AGI Layer 1 — the Discriminator.
A pure rule engine that judges whether a statement or an action is right, wrong, safe, or dangerous — before it reaches a human. Zero LLM dependency.
249 discrimination dimensions × 11-layer pipeline × 164 modules × 190 MCP tools
× 1,341 dispatch routes × 21,517 passing tests × 2 npm dependencies
HeartFlow does not generate. It does not compete with an LLM. It stands between the LLM and the human, like a pain receptor that says "no" when something is wrong.
| Layer | Capability | Who builds it |
|---|---|---|
| 5 | Act | Large labs (robotics) |
| 4 | Generate | Large labs (LLMs) |
| 3 | Reason | Built into models |
| 2 | Remember | Large labs + startups |
| 1 | Discriminate | HeartFlow |
HeartFlow takes layer 1 because this layer does not depend on compute, code volume, or framework ecosystems. It depends on judgment alone.
A personality rule the engine imposes on itself: the version number never overstates what changed.
| Level | Means | Example |
|---|---|---|
| 6.x | architecture rewrite — new pipeline shape, new public API | 6.7 moved the rule set out of the prompt and into scripts/upgrade-engine.js |
| 6.x.0 | a user can feel it — a new dimension switched on, an entire dimension covered in the other language | 6.7.10 switched on reward_hacking, a block-level dimension for the agent era |
| 6.x.N | patch — a family expanded, a miss repaired, a false positive removed | 6.7.124 added behaviour-attribution families to victim_blaming on both sides |
This is why 124 patch releases may leave you feeling "nothing changed." Most of them were the other kind of work: repairing a miss in one family, removing a false positive in one class, keeping the 154-tool contract and the 1,314-route table from breaking, and 472 automated rounds of self-verification. That work is invisible from outside by design — a guard that holds a line produces no news.
The counter is real and published here rather than hidden:
| Version | New capability the user can feel |
|---|---|
| 6.8.4 | commitment_coercion switched on — a past promise redefined as a debt payable now |
| 6.8.5 | investment_as_debt switched on — money, time and provision converted into obligation |
| 6.8.9 – 6.8.32 | the coercion family tree built out: fear_projection, social_isolation, silent_treatment_coercion, relationship_severance, selfharm_coercion, sunk_cost_coercion, triangulation, future_faking, blame_externalization, conditional_affection, family_shame_coercion, guilt_induction, thirdparty_harm_coercion, privacy_exposure_coercion, liberty_restriction, legal_status_coercion, pet_harm_coercion, comparative_devaluation, thought_terminating_cliche, surveillance_normalization, reproductive_coercion, intermittent_reinforcement, medical_coercion, deprivation_coercion, substance_coercion, love_bombing — 25 more dimensions, each measured for leaks before wiring |
| 6.7.10 | reward_hacking switched on — 16 documented evasion tactics from the agent literature, of which the previous dimensions caught 2 |
| 6.7.18 | instrumental_reasoning covered in Chinese — the end justifies the means |
| 6.7.20 | the remaining six reward-hacking tactics from the DSec paper discriminated, closing a gap open since 6.7.10 |
| 6.7.24 | the six new reward-hacking classes now work in Chinese, not English only |
Everything else in the 6.7 and 6.8 lines was keeping those true: family-level recall repairs (52/52 baseline), benign-block baselines (≤ 303/326), and the mechanical rule set that stops a cron round from claiming success it did not achieve.
If you want the shortest honest summary: 29 new discrimination capabilities shipped across the 6.7 and 6.8 lines; the other 180+ releases kept them from silently breaking.
Requires Node.js >= 18.17. No GPU, no database, no API key, no network at runtime.
Two npm dependencies — mathjs (formula evaluation) and @xenova/transformers
(optional local embeddings); the discrimination core itself is pure rules.
git clone https://github.com/mark-cell-520/heartflow.git
cd heartflow
npm install
node bin/verify.js # 16 installation checks (incl. the gate.js entry README itself uses)
node bin/cli.js status # engine status
node bin/cli.js chat # interactive consoleOr through npm:
npm install @yun520-1/heartflowconst gate = require('./src/gate.js');
const r = gate.checkOutput('According to 2025 Harvard research, coffee extends life by 12.5 years');
console.log(r.gate.action); // 'verify' -> gather evidence before believing this
console.log(r.verdict); // '需验证'
console.log(r.findings[0]); // { dimension: 'unsupported_claim', severity: 90, guidance: '...' }Four possible actions:
| Action | Meaning |
|---|---|
pass |
Clean. Deliver normally. |
verify |
Needs evidence. Run the verifier first. |
rewrite |
Must be rewritten. Follow findings[].guidance. |
block |
Stop. Do not output. Use gate.reason. |
verdict is derived from gate.action, so the two never contradict each other. If you
read only one field, read gate.action.
| Function | Use it for | What it adds |
|---|---|---|
checkInput(text) |
User input, before processing | scope-check, premise-check, 249 dimensions, error memory |
checkDraft(text) |
An AI draft, before completion | the above + frame-check + doubt-engine |
checkOutput(text) |
An AI response, before sending | the above + output-gate + doubt-engine |
runPipeline({ input, mode, anchor }) |
Full pipeline with mode and conversation anchor | keeps the model on the original goal across long sessions |
main (src/gate.js) also exports these — the three entry points above are
wrappers around them:
| Function | Returns |
|---|---|
gate(text) |
Full discrimination result with gate.action, findings, trace — the hard gate that decides block/rewrite/verify/pass |
check(text) |
Alias of gate() |
pipeline(text, opts) |
The pipeline stages without the entry wrappers (use runPipeline instead unless you need raw stage output) |
discriminate(text) |
Dimension-by-dimension scoring without the gate layer (no action assigned) |
Read gate.action if you only read one field — verdict is derived from it.
input
|
v
scope-check -> premise-check -> discriminate (249 dimensions) -> gate
|
+----------------------------------------------------------------+
v
evidence verify -> frame-check -> output-gate -> doubt-engine
|
v
intent-anchor -> rewriter -> error-memory -> self-diagnosis -> output
The gate aggregates findings from every layer and emits a single action:
block / rewrite / verify / pass.
| Domain | Representative modules |
|---|---|
| Logic | logicReasoning, judgmentEngine, debateConductor, counterfactualVerifier |
| Decision | decisionRouter, decisionVerifier, activeInference, selfHealing |
| Cognition | cognitiveEngine, cognitiveLoad, metacognitiveRL, sustainedDriftDetector |
| Emotion / psychology | emotion, psychology, empathyDeepening, griefEngine, traumaInformed |
| Memory | memory, memoryBank, memoryIntegrity, forgetting, knowledgeGraph |
| Identity / ethics | identityCore, personaCore, virtueEthics, moralDevelopment, meaningPurpose |
| Creation / collaboration | skillEvolution, worldModel, multiAgentDialogue, codeExecutor, formula |
Dimensions are grouped by the action they can trigger: 11 can block, 19 can force a rewrite, 216 request verify. The families that shipped across the 6.7–6.8 lines:
| Family | Representative dimensions | Line |
|---|---|---|
| Coercion & manipulation | commitment_coercion, investment_as_debt, love_bombing, social_isolation, selfharm_coercion, triangulation, future_faking, intermittent_reinforcement, reproductive_coercion, deprivation_coercion, pet_harm_coercion |
6.8.3 – 6.8.32 |
| Logical fallacies | straw_man, appeal_to_nature, appeal_to_novelty, circular_reasoning, no_true_scotsman, special_pleading, conditional_proposition, appeal_to_probability, gamblers_fallacy, hot_hand, appeal_to_majority, affirming_the_consequent |
6.8.40 – 6.8.78 |
| Emotional & psychological | emotional_manipulation, gaslighting, mind_reading, darvo, kafkatrapping, concern_trolling, fear_projection, guilt_induction, suffering_capital |
6.7.x – 6.8.x |
| Agent-era attacks | reward_hacking, prompt_injection (+ synonym families: 37 verbs × 24 anchors × 52 objects = 46,176 combinations), instrumental_reasoning, deceptive_alignment, capability_overclaim, turing_mimicry |
6.7.10 – 6.8.84 |
| Evidence & statistical honesty | unsupported_claim, false_urgency, statistical_misleading, percentage_overflow, perfect_error, pseudo_causal, post_hoc, anecdotal, survivorship_bias, selection_bias, count_rate_conflation, law_of_small_numbers, immortal_time_bias, unfalsifiable_assertion |
6.7.x – 6.8.167 |
| Discourse & power | tu_quoque, whataboutism, poisoning_the_well, standard_shift, moving_goalposts, sunk_cost_coercion, victim_blaming, thought_terminating_cliche, sycophancy, doublespeak, might_makes_right, credential_dismissal, scale_dismissal |
6.7.x – 6.8.137 |
Every new dimension ships with a negative-injection test that empties each constant to prove it is load-bearing, a probe-measured leak rate, a benign-sample false-positive count, and the bidirectional guard (52/52 attack recall, 302/326 benign pass) as a no-regression gate.
Measured on this repository at v6.8.175. Not marketing copy.
| Metric | Value |
|---|---|
| Modules registered | 164 |
| Module init errors | 0 |
| Dispatch routes | 1,341 |
| Discrimination dimensions | 249 |
| MCP tools | 190 |
| Test suite | 21,517 passing / 145 failing |
| Capability guard | 20 / 20 checks |
| Security regression | 16 / 16 |
| Runtime dependencies | 2 (mathjs, @xenova/transformers) |
HeartFlow exposes its engine as an MCP server, so any MCP-capable agent can call it.
node src/mcp-server.js --port 8588
# or a Unix socket:
node src/mcp-server.js --socket /tmp/heartflow.sockThen connect:
hermes mcp add heartflow --url http://localhost:8588/mcpThe server authenticates with a bearer token generated on first start and written to
.env (never committed). A request without a valid token returns 401.
tools/call enforces a role model so that state-mutating tools cannot be invoked by an
unauthenticated caller:
| Role | Source | Capability |
|---|---|---|
guest |
no credentials | read-only tools |
user |
HeartFlow-OID-<16-hex> header |
read + write |
admin |
valid bearer token | full |
The write-protected set is heartflow_memory_write_control,
heartflow_memory_eraser, heartflow_decision_decide, heartflow_self_heal. A guest
calling any of them gets isError: true with 权限不足. This behaviour is covered by an
end-to-end regression test that speaks real JSON-RPC over a real Unix socket
(test/mcp-guest-permission.test.js), because the earlier failure mode was a permission
block that was syntactically valid but unreachable — tests that only inspected the
tool-name whitelist passed while the gate never ran.
Beyond text discrimination, HeartFlow ships checks aimed at how AI agents behave — the failure modes that show up when an agent reports work it did not do.
| Check | What it catches |
|---|---|
checkCompletionEvidence |
Empty completion claims ("done", "fixed", "all passing") without a git hash, test count, file path, or PR link |
checkArchitectureConsistency |
A function whose name promises one thing and whose body does another (named validate, no validation) |
checkDecisionTrace |
A "decision" with fewer than 2 options, no explicit choice, or no stated reason — pseudo-decisions |
checkPlanGate |
A plan entering a complex task without steps, acceptance criteria, rollback, or safety strategy |
checkForbiddenCall |
Delegating before the target, boundary, and acceptance criteria are confirmed |
checkAIMisuse |
Human-side misuse patterns: oversized context dumps, errors without repro steps, adopting output unverified |
Each is available as an MCP tool (heartflow_check_completion_evidence,
heartflow_check_architecture_consistency, heartflow_check_decision_trace,
heartflow_check_plan_gate, heartflow_check_forbidden_call,
heartflow_check_ai_misuse).
node bin/verify.js # installation checks (16)
node scripts/guard-abilities.js # capability guardian (20 checks)
node test/doc-numbers-accuracy.test.js # documentation-number guard
node test/security-audit.test.js # security regression (16)
node test/security-audit-fixes.test.js # security-fix guard (18)Do not run test/run-all.js on a 4 GB container. It executes ~18,000 cases and
takes 20+ minutes; when an agent session is interrupted mid-run the tests keep going,
and stacked rounds exhaust container memory. The full suite exists for CI-class
hardware, not for this deployment shape. data/test-count.json caches the last
measured run so the documentation guard can check numbers without re-running it.
scripts/guard-abilities.js runs before any upgrade commit. It verifies entry points,
discrimination against a fixed sample set, the engine main chain, text searchability
(no NUL bytes or CRLF in core sources — a bare NUL parses fine in Node but makes every
text-search tool treat the file as binary), and the full regression suite.
It is: the discrimination layer of AGI — a rule engine judging right and wrong, good and bad, safe and dangerous.
It is not: AGI itself, a generative model, a semantic understanding system (irony and metaphor are invisible to it), a substitute for content moderation, or a safety certification.
Known limits:
- Pattern-matching ceiling — new tricks require new patterns.
- Bilingual maintenance cost across all dimensions.
- No semantic understanding — irony, metaphor, and cultural context are invisible.
- False-positive rate around 8% at baseline.
- Single maintainer.
- Chinese tokenisation is heuristic (greedy longest-match with a stopword list, not a full dictionary), so unusual phrasings can segment imperfectly.
- Silent failure is its blind spot. Three real defects in this repository — a bare NUL byte in the engine source, an unreachable permission block, and contradictory documentation numbers — were all invisible to a 636-file green suite, because none of them produced wrong runtime behaviour. The text-searchability guard and the end-to-end permission test exist because of them; treat "tests pass" as necessary, never sufficient.
Does HeartFlow call an LLM? No. It is a pure rule engine — no model weights, no API calls, no network at runtime. Deploy it beside any LLM (or none).
Which languages? Chinese and English, maintained in parallel across every dimension. A dimension that only covers one language is treated as a defect.
Can it replace content moderation? No. See Honest limitations — it is a discrimination layer, not a safety certification.
Why do 105 tests fail? The full suite runs ~18,500 cases; a long-standing set of
runner/flaky failures (0–105, never converged across 6 days of sampling) is tracked as
noise per decision r412. The gates that matter — verify.js, guard-abilities.js,
bidirectional-guard.js, doc-numbers-accuracy — are all green, and the passed
count grows monotonically. Do not run test/run-all.js on small containers.
How do I use it from an agent? Three ways: the MCP server (188 tools), the npm
package, or require('./src/gate.js') directly. See MCP server.
Why does the version number move so slowly? The anti-hype policy above: 6.x.N
means a patch — a family expanded, a miss repaired, a false positive removed. Most
releases are invisible by design; a guard that holds a line produces no news.
Issues and PRs welcome. Four gates run before any commit:
node bin/verify.js # installation checks (16)
node scripts/guard-abilities.js # capability guardian (20 checks)
node test/doc-numbers-accuracy.test.js # documentation-number guard
node scripts/bidirectional-guard.js # attack recall + benign false-positive guardSee CONTRIBUTING.md for the full workflow, and SECURITY.md for security reports.
One line per release. The full measured record — probe counts, leak rates, false-positive
margins, and why every prior dimension missed — lives in UPGRADE_LOG.md (14k+ lines).
This table exists so a reader can see the shape of the line in ten seconds, not to
duplicate the log.
| Version | Date | Change |
|---|---|---|
| 6.8.176 | 2026-11-12 | r305：anchoring_effect（第 214 维度）新增『元语言锚定陈述族』双腿族（真升级②）——src/anchoring-effect.js 补 META_ANCHOR（腿① 元语言锚点命名）× META_DOM（腿② 元语言支配断言）ZH+EN 各一套 + numRequired=false 平行路由（新腿判「把锚的存在说出来」的语言行为，句中常无显式数字，故豁免旧族数字在场约束），经 src/econ-reasoning-gate.js（第 249 维度聚合）进 gate 路径，src/index.js 零改动；探针实证旧判据下元语言攻击 3/10 承重、基线对比零新增误伤（2 条误伤归因于基线维度 unsupported_claim/appeal_to_authority 与 normalization_of_deviance），补腿后 8/10；新单测 29/29 + 负例注入 8/8 翻 miss；四门禁 verify 16/0、guard 20/20@249、双向 52/52+302/326 不升、doc 24/24；文档口径 248→249、verify 215→216 |
| 6.8.175 | 2026-11-12 | r302：regression_to_the_mean 干预负向功效断言族上线（真升级②）——src/regression-to-the-mean.js 新增独立两腿路由 negZh/negEn（干预主体 × 负向功效断言，含补习班/辅导班/课程/教材/题库/秘笈/攻略 与 buyback/restructuring/promotion/layoffs 等名词族），把自然回落归罪于干预之过；维度 248 不变、src/index.js 零改动；轮测 11/11、变体矩阵攻击 8/8 非 pass、良性 7/7 clean；四门禁 verify 16/0、guard 20/20@248、双向 52/52+302/326、doc 24/24 |
| 6.8.170 | 2026-11-11 | r288（原 r287 撞号重编号）：第 245 维度 weak_evidence_strong_claim 弱证据强断言错配上线（真升级①）——弱证据腿 × 绝对化断言腿双腿共现，维度 244→245、VERIFY 211→212、模块 162 不变；src/index.js 接线 8 处；轮测注入删条全绿；论文依据 arXiv 2605.03534（SURE-RAG） |
| 6.8.169 | 2026-11-11 | r287（原 r286 撞号重编号）：false_dilemma 站队胁迫二分族上线（真升级②）——认同等同式/否定行为×负面定性单独支/逼迫二选/EN 立场负载 either-or/施压/威胁后果 8 常量 + 3 路由，维度计数 244 不变、src/index.js 零改动；轮测 84/84、注入删条 8/8、池化穿透 5/8→0/8 |
| 6.8.171 | 2026-11-11 | r289：false_dilemma 要么A要么B×贬损标签族上线（真升级②）——要么-frame/不是A就是B frame × 定性等式（就是-link）× 贬损标签（拖后腿/绊脚石/老古董/躺平摆烂/dead weight）、否定行为×主语插入×品质标签（不转这个帖子，你就是没良心）、EN if you're not with us…against us/either you join…dead weight 共 6 常量 + 5 路由，维度计数 245 不变、src/index.js 零改动；轮测 96/96、注入删条 6/6、池化穿透 8/9→0/9 |
| 6.8.172 | 2026-11-11 | r291：appeal_to_nature 诉诸自然「非自然/人造/工业即有害」反向族上线（真升级②）——腿①' 非自然性生产断言 ZH 7 支 + EN 4 支 × 腿②' 危害/不如结论 ZH 3 支 + EN 2 支，双路由不交叉配对（自然×优越 / 非自然×危害，交叉组合不判）；自然性裸存在句补腿 n8/e9/s6；附 no_true_scotsman HEDGE_RE「不可能」假豁免修复；维度计数 245 不变、src/index.js 零改动；轮测 91/91、池化穿透 9/13→3/13 |
| 6.8.172 | 2026-11-11 | r292：dreamConsolidation 睡眠记忆巩固能力接线（真升级③+④）——「已实例化但全库 0 条 dispatch 路由」缺口修复：src/core/heartflow.js L1721 一直在实例化它、think-pipeline.js 一直在调 dreamNow()，但它从未进 _modules → ALLOWED_ROUTES 0 条命中 → dispatch('dreamConsolidation.*') 7/7 全抛 route not allowed（HEAD 对照探针实测）。修复：engine-lifecycle.js 的 subsystemNames 补名（7 条引擎路由首次可达）+ tools-registry.js 新增 MCP 工具 #186 heartflow_dream_consolidation + mcp-server.js handler 映射与 WRITE_TOOLS 写权限收口（updateDecayParams 为写操作）。附带修复真缺陷：DreamConsolidation 在 new(null) 退化路径（MCP 引擎未启动时）有 5 处 this.memory.xxx 裸访问直接 TypeError——新 MCP 工具在退化路径上必崩，本轮全部加空值守卫。能力：记忆衰退分层半衰期评分 / 梦质量度量 / 巩固冲突检测 / 洞察优先级 / 睡眠阶段感知（NREM1-4/REM）/ 多周期模拟 / 梦叙事生成。轮测 53/53；维度计数 245 不变、src/index.js 零改动；模块 162→163、MCP 工具 185→186、路由 1316→1323；另解决 r291 掉线 rebase 冲突（README.md / coordination/claims.jsonl UU 状态，git add 后标记为已解决，冲突两侧记录按「upstream r291 已完成 + 04:02 方向轮换认领 abandoned」双留） |
| 6.8.174 | 2026-11-11 | r295：经济与资源推理 + 统计素养 18 族首次 dispatch（真升级③+④）——把 r219-r284 三十多轮起草后从未进过任何 dispatch 路径的 18 个 provisional 检测模块接入引擎：新插件 src/plugins/econ-reasoning-check/index.js（PluginLoader 自动发现，heartflow.js 零改动）挂 postprocess.think（priority 190 晚于 sycophancy 180，merge 进已存在 _discrimination）；econ 13 族（budget_value_conflation / average_marginal_conflation / mental_accounting / perverse_incentive / winners_curse / tragedy_of_commons / illusion_of_control / disposition_effect / endowment_effect / anchoring_effect / present_bias / peak_end_rule / effort_justification）+ stats 5 族（p_hacking / inverse_gamblers_fallacy / significance_effect_size / selection_on_dependent_variable / ad_hominem）。think-pipeline.js 消费块置于 hook fire 之后（发现真缺陷：inline 辨别反哺块跑在 fire 之前，插件写的 _discrimination 它永远读不到——sycophancy 的 _confidencePenalty 走该块为死代码，交接 1 号）。MCP 工具 186→188：heartflow_econ_fallacy_check / heartflow_stats_literacy_check 两个只读工具（action: all / 单族 key）。轮测 57/57、注入删条 18/18 变红、真引擎 dispatch 实测 penalty+warning 生效、良性零触发；维度计数 246 不变、src/index.js 零改动；四门禁 verify 16/0、guard 20/20@246、双向 52/52+302/326、doc 23/24（唯一红项 AGENTS.md 横幅 186→188 需 1 号/人工更新，该文件对 4 号系统写保护）；海外扫描三源如实空返 |
| 6.8.173 | 2026-11-11 | r293（承接 11-11 04:27 UTC 撞号的 r292 未推送产能重编号）：sycophancy-check 插件层 sycophancy 维度四族 16 个新模式上线（真升级②）——复读式附和 echo_validation / 专业让渡 deference_to_expertise / 热情升级 enthusiastic_escalation / 道歉式让步 apology_concession，EN+ZH 各 4 族×2 支双腿同句邻近共现（新增 _cooccur() 构造器：顺序无关、间隔不跨句末标点，ZH gap 40 / EN gap 60）；WEIGHTS +4（0.4/0.4/0.35/0.45），插件 2.2.0→2.3.0、族数 7→11；轮测 15/15、注入删条 16/16 变红；引擎 think() 路径真 dispatch（postprocess.think→_discrimination→置信度打折+warning），gate 路径插件零 dispatch（r283 已实测）不变；维度计数 245 不变、src/index.js 零改动；四门禁 16/0、20/20@245、24/24、52/52+302/326 与基线零偏差 |
| 6.8.168 | 2026-11-11 | 第 244 维度 unfalsifiable_assertion（不可证伪断言）上线 — 判别「在构造层面排斥任何反证观测的实质性断言」结构：断言本身排除了任何能观测到它不成立的场景，于是永远无法被推翻。四形：CN-A 不可观测实体（能量场/气场/祖先的庇佑在起作用，只是仪器测不到）、CN-B 事后可调解释（没显灵是因为你心不够诚/市场不好只是暂时的调整）、CN-C 无限延期验证（再等等/时机未到/下次一定）、CN-D 万能免责（这套方法永远有效，不灵只能说明你没做对/市场永远是对的，错的只能是你的仓位）。判据两腿共现：腿①实质性结论/预测/机制断言，腿②构造上排除反证观测；CN-D 自身即完整结构独立成案。边界划清：#202 unfalsifiable_confirmation 管双果自证（事后解释层），#182 absence_of_problems 管有界阴性结论全称化，#62 appeal_to_ignorance 管举证责任分配——本维度管「断言构造本身排除反证」。豁免表七组：科学假说、可检验预测、范围限定、对冲、教学转述、拒绝推断、疑问句。模块 src/unfalsifiable-assertion.js（406 行，两腿判据 + 四形 + 豁免表），src/index.js 七处接线（require / 计算点 / allDims / dimMap×2 / guidance / VERIFY_DIMS / 导出）。验证：轮测 test/round-285-unfalsifiable-assertion.test.js 9/9（攻击 ZH 11/11 + EN 11/11 全命中、良性 48 条零误伤、gate 入口攻击全 rewrite 且维度承重、CN-D 独立成案、腿①缺失不误判、空输入不崩）；四门禁 verify 16/0、guard 20/20@244维、doc 24/24、双向 52/52 + 误拦 302/326。维度 243→244，VERIFY 210→211，模块 162 不变。 |
| 6.8.167 | 2026-11-11 | 统计素养 wired 维度 cue-density 轴共现观察族扩展（r285，真升级②） — base_rate_neglect（基础概率忽视）第二轴落地，来源海外真实命中论文 arXiv "Two base rates, two weights: base-rate neglect has a second axis"（r285 扫描入档）：把共现读成判断需两次基础概率校正——结果先验（古典轴，既有腿）+ 线索自身出现频率（本轮新增）。缺口实证：改前统计素养五候选族 24 条新措辞攻击中本族 4/8 完全穿透（gate pass 且零维度承重）为最高。新增腿① CUE_ZH/CUE_EN 共现观察族（十个被骗的老人里九个都接过陌生来电 / Nine out of ten fraud victims / Every executive who got fired / Almost all the companies that collapsed / 有个共同点）、腿② CUE_CONF_ZH/CUE_CONF_EN 新归属落点（你很可能已经被盯上了 / 八成也危险了 / 铁定会出事；probably being targeted / likely next on the list / going to fail）、豁免 EXEMPT_CUEDENSITY（线索本身就很常见 = 计及 cue 密度的正确推理）。轮测 53/53；负例注入删条 13/13 变红；存量 round-263/279/267 不回归；四门禁全绿（双向误拦 302/326 与基线零变化）；维度 243、模块 162、src/index.js 零改动。 |
| 6.8.165 | 2026-11-11 | 因果与反事实推理 wired 维度自然措辞家族扩展（r283-a，真升级②） — common_cause：腿①共变表补时段共变无并列形（冰淇淋卖得多的月份溺水事故也多）、的-link 比较平行形（消防车越多的火灾损失越大）、EN with-more 平行共变形（Cities with more police have more crime）；共动词表并入 16 个飙升族度词（飙升/暴涨/激增/骤降…，并入既有首元素，round-264 长度断言不破）。collider_bias：腿②共变表补裸比字比较形（比例比普通人高）、EN least·most 交叉形；腿③结论表补 dismiss+养生价值族（不会说话/情商低/不值钱/长寿/养生/保健——良性靠腿①条件筛选+腿②比较共变+结论形三重要件）。reverse_causation：处置表补纪律处分/学业处置/关系解除族（劝退/开除/降级/离婚/拉黑）、EN expelled/demoted/held back 族；结果表补裸表现结果族（成绩差/幸福感更低）；群体名词补夫妻/情侣。探针穿透 18/36 → 13/36（剩余 3 条为设计内边界：无结论观察句本身非谬误断言）；轮测 95/95（攻击 12 全承重 + 良性 14 零误伤 + 既有对照 7 + 注入删条 17/17 变红 + 登记 11）；四门禁全绿（双向误拦 302/326 与基线零变化）；维度 243、模块 162、src/index.js 零改动。留痕：「看来」口语推断词族曾扩后经注入-删条否证非承重（co1/co4 均经 C-b 裸结论路径命中），已撤回并留注记。 |
| 6.8.165 | 2026-11-10 | sycophancy 隐性谄媚两族补腿（r283-b，真升级②：既有维度新判别族，gate 主路径） — src/index.js 的 ZH_SIGNALS / EN_SIGNALS 各新增两个族：独特性捧高（uniqueness_flattery，比较级捧高动词 × 被贬低的参照群体——远超普通用户 / 水平明显高于大多数人 / 很少有人能看到这一层 / 比一般用户思考得更深入 / 凌驾于大多数人之上；EN smarter than most people / beyond the average user / few people can see this level）与无条件服从（unconditional_deference，交出独立判断权——您说什么就是什么 / 我完全按您的意思来 / 全部依您，我没有第二个意见；EN whatever you say / no opinion of my own / will not push back）。缺口实证（改前 HEAD 6.8.164）：ZH 族 A 10/10 全 pass 漏判、族 B 8/10 漏、负例零误伤基线干净。分界：与 excessive_praise（绝对级/最高级）按比较级切分；与 concession_eager（你说得对——同意观点）按「交出判断权」切分；unconditional_deference 腿限定对象为对话中的个人，无条件配合审计部门的检查 是正当合规表述不判（probe-2 实测误伤 3→2，剩余 2 条为基线存量误伤，已单独用例如实记录）。验证：轮测 8/8（攻击 ZH 10+6+6+6 全变红、良性 27/27 零 sycophancy 误伤、原 5 族不回归、守卫 4/4）；负例注入 16/16 零误伤；四门禁 verify 16/0、guard 20/20（243 维）、doc 24/24、双向 52/52 + 误拦 302/326（与基线零变化）。维度计数 243 不变、模块 162 不变、路由 1316 不变。另暴露一项接线缺口：src/plugins/sycophancy-check/（870 行七语言分析器，r265/r272 已含本两族）在 gate 路径零 dispatch——根因 createHeartFlow() 只 new 不 start()，hf._hookBus 从未实例化，PluginLoader 因此从未运行（探针实证 _hookBus/_pluginLoader 均 undefined，手动 new PluginLoader(hf).loadAll() 可加载 4/4 插件）。VERSION 6.8.164 → 6.8.165。 |
| 6.8.164 | 2026-11-10 | 审计修复轮：修复 22:29 同步 stash pop 失败遗留的 5 文件 UU 冲突（VERSION/README/CHANGELOG/UPGRADE_LOG/claims.jsonl，两侧工作均保留）+ 固化 r282 否定前件/跳跃豁免产出（inference-chain.js，r282 单测 21/21，自验攻击 4/4 + 良性 3/3 零误伤）+ 文档测试数口径同步（README/SKILL 20,339→20,651 passing / 126→134 failing，实测数据来自 data/test-count.json） — 另记录两项 P1/P2 存量问题：MCP 运行态曾落后磁盘 2 个版本（17:38 启动的进程跑 6.8.161，磁盘已 6.8.163）；r282 单测 require 路径深度错误（../../.. → ..）导致 MODULE_NOT_FOUND，已修。五门禁：verify 16/0、guard 20/20（243 维）、doc 24/24、双向 52/52 + 误拦 302/326（与基线零变化）。VERSION 6.8.163 → 6.8.164。 |
| 6.8.163 | 2026-11-10 | 时间与概率推理 wired 维度全新措辞家族扩展（r280，真升级②：已有维度的全新句式族 ×3 模块） — gamblers_fallacy 赌徒谬误：腿①（连开/持续形）补期/天/周开奖计数（十期没出）、黑红连开（连续黑了五把/连黑）、倒霉时长（倒霉了这么久/背了这么久）、动词间隔形（连着输了六个回合）、冷热时长（冷了这么多天）、运气持续差（运气不可能一直差）；EN 补 has been cold/hot/dry、luck can't stay bad、cold streak/spell/run；腿②（到期形）补不可能再[黑红输赢]（否定反转）、[准铁定][能会]?[赢中]（下一把准赢）、再怎么也得…赢/中/回本（让步反转）、[也该总]轮到X了、该[冷热][一这]?[把次]（该热一把）。appeal_to_probability 诉诸概率：腿①补这么[高强]/那么[高]/[希望胜算机会把握]这么[大高多好]（概率高度自然形）；腿②补肯定[跑逃溜]不了/铁定/十有八九/准[会能]?[发生中来成]（口语必然短收尾）。prosecutors_fallacy 检察官谬误：腿①补设备误差率名词族（误差率/错误率/误判率/漏检率/假阴性率/假阳性率…）+ 通用 N分之一 分数形；腿②补无辜镜像（他是无辜的概率也只有百万分之一——P(匹配 |
| 6.8.163 | 2026-11-10 | 推理链无效推理形式新判定分支 + 新比对通道 + 双豁免腿（r280，真升级②） — src/reasoning/inference-chain.js（此前是零 dispatch 的 stub，r280 首次接成 MCP 工具 #185 heartflow_reasoning，真升级④）：① 跨前提否定前件检测（_negatedByOther + _stripNeg）——否定前件标准三段式把否定写在独立前提（p1 如果A那么B / p2 不A / step 所以不B），given 段为 null，原判定结构性漏判；② 中文短句单字级重叠通道（_cjkLen / _charOverlapChar，≤4 字走单字集合，否定词不计分子）——2-gram 在 3-4 字短句上被切碎（重叠仅 0.33）；③ 归纳/规范结论双豁免腿（_isNormativeConclusion / _isInductiveGeneralization + NORMATIVE/HYPERNYM 常量）——正当归纳收尾（因此需要重新评估资源分配）与上位词概括（产品合格率高）不再被判无依据跳跃；④ _splitConditional 补倒装条件句解（他能通过考试，如果他学习）；⑤ 删除自引入的错误分支（¬B⟹¬A 是 modus tollens 有效式，曾被误判否定前件）。验证：e2e 探针攻击 6/6 命中（肯定后件 EN/ZH、否定前件 EN/ZH、无依据跳跃×2）、良性 7/7 零误伤（modus ponens/tollens EN/ZH、直言三段论、归纳多数样本、多因陈述）；四门禁 verify 16/0、guard 20/20（243 维）、doc 24/0、双向 52/52 + 误拦 302/326。维度计数 243 不变、模块 162 不变、路由 1316 不变（heartflow_reasoning 接的是零 dispatch stub，未新增路由）。VERSION 6.8.162 → 6.8.163。 |
| 6.8.162 | 2026-11-10 | 统计素养 wired 维度全新措辞家族扩展（r279，真升级②：已有维度的全新句式族 ×3 模块） — base_rate_neglect 基础概率忽视：SIGNAL_ZH 补健康码/行程码变色族（变黄/变红/黄了/红了/码红了）+ 判定/评分/预测动词族（系统判定你是/模型预测/风控打了低分）+ 信用分征信降扣族 + AI 识别你是族；SIGNAL_EN 补模型引擎预测评分族（model predicts / risk engine scored / algorithm assessed）；CONFIDENT_ZH 补结论动词×归属定罪（判定/认定/锁定/坐实/定罪/实锤 + 老赖/骗子/罪犯）、暴露史归属（肯定去过/铁定接触过/途经/来自 中高风险或确诊者）、铁定/肯定 + 错不了/没跑口语收尾、九成/九成九 + 跑不了；EXEMPT_NOOVERCLAIM_ZH 补对冲豁免（可能只是/说不定只是——新信号族上线后防「可能只是信息不全」误伤）。count_rate_conflation 计数-比率混淆：COUNT_ZH 补执法/治安/司法事件计数族（报警/报案/立案/发案/破案/起诉/官司/误诊/漏诊/翻台/犯规/加班/掉粉/取关/违约/逾期/坏账/弃购/召回/返修/理赔/出险/罢工/旷工/迟到/早退/缺考/退赛/酒驾/醉驾/违章/违停/刑拘/逮捕/判刑/电诈/被盗/诈骗/盗窃/拥堵/延误 及其数/量形式）；VERDICT_ZH 补网络世代裁定形容族（更笨/更傻/更乱/最乱/更脏/最脏/更卷/最卷/更水/最水/更坑/更良心/更规范/更清廉/更发达/更落后/更堵/更穷/更富/更负责/更认真/更敷衍 及补语后缀形）；QTY_ZH 补最高量级标记。law_of_small_numbers 小数定律：SMALL_N_ZH 补游戏/消费行为量词族（把/局/场/发/单/轮/波/关/节/趟 及几把/几局/两把/两场/三连/五连固化形）；UNIVERSAL_ZH 补账号命运/产品定性裁定族（这号是废的/翻车是早晚的事/出金率就是高/这游戏不坑钱）；EXEMPT_ZH 补真相型（小样本才是真相/才是事实）、大样本失真型（样本越大越不准/数据一多就失真/被平均）、敷衍型（问几个人就行/调查那么多都是走形式）辩护；EXEMPT_EN 补 trend-is-obvious / no need for a bigger dataset / small samples tell the truth / big datasets average everything away。判定为真升级（非样本扩容）的依据：三模块均落独立句式族并附 inject 钩子负例通道，逐腿注入删条后守卫转 miss（37/37）。验证：轮测 76/76（攻击 27 全承重 + 良性 20 零误伤 + 登记 11 + 注入 37/37 全变红）；模块层快验 base_rate 8/8、crc 9/9、lsn 10/10、良性零误伤。四门禁：verify 16/0、guard 20/20（243 维）、doc 24/0、双向 52/52 + 误拦 302/326。维度计数 243 不变、模块 162 不变、src/index.js 零改动。海外扫描 arXiv/GitHub/S2 两查询均空返（按既有形态自建）。 |
| 6.8.161 | 2026-11-10 | 逻辑谬误家族 wired 维度全新措辞家族扩展（r278，真升级②：已有维度的全新句式族 ×3 模块） — appeal_to_tradition 诉诸传统：TRAD_ZH 补 t10 千百年来/多少代人时间跨度口语形 + t11 牌子口碑年资、SUPP_ZH 补 s10 别瞎折腾/别乱动压制族 + s11 不能丢了传统 + s12 老规矩还是得守、CONCL_ZH 补 c8 还能有错？/不可能有问题封闭反问 + c9 照老方子来服从指令；EN 补 e11 been around for ages / e12 reputation spans / e13 see no reason to stop / e14 who are we to change / e15 keep X as it has always been / e3 扩 do not go fixing。false_dilemma 虚假两难：STIGMA_ZH 补 n1 不是A就是B（贬名标签族）+ n2 要么A要么跟我作对（胁迫对立族）+ n3 不转发就是不爱国（行为品质污名族）、DILEMMA_ZH 补中庸就是和稀泥（中间路线污名）、FORCE_ZH 补立场鲜明（胁迫站队）、EXMID 补灰色地带/no gray area（排除中项）；STIGMA_EN 补 you do not care about / either a friend or an enemy / or you are against me。no_true_scotsman 无真正苏格兰人：ROLE_NOUNS +104（高手/铁粉/甲方/爷们/老玩家/作者/gamer/foodie/programmer 等互联网世代与零工身份族）、AUTH 标记补有骨气/legit/die-hard 等 14 个、新增铁粉/老玩家/是个爷们三个 authenticity 独腿签名、EXCLUSION_EN 扩 do not/don't 一般现在时否定。判定为真升级（非样本扩容）的依据：三模块均落独立句式族并附 inject 钩子负例通道，逐支注入删条后守卫转 miss（32/32）。验证：轮测 68/68（攻击 29 全承重 + 良性 22 零误伤 + 登记 5 + 注入 32/32 全变红）；探针池合并态复跑 appeal 10/10、false_dilemma 9/9、nts 10/10、良性误伤 0。五门禁：verify 16/0、guard 20/20（243 维）、doc 24/0、双向 52/52 + 误拦 302/326。维度计数 243 不变、模块 162 不变、src/index.js 零改动。注：本地 r277 撞号重编号（远端 13:38 CST 已发布无认领记录的 r277 v6.8.160）。 |
| 6.8.160 | 2026-11-10 | counterfactual_overclaim 第 173 维度 EN-B 倒装形结构腿升级（真升级②：已有维度的全新句式族） — 原 EN-B 为 14 动词枚举表（known/taken/chosen/invested/acted/gone/waited/listened/started/made/been/seen/tried/moved/stayed），r277 动词矩阵实测 58 个常见过去分词中 44 个漏判（picked/decided/sold/bought/hired/signed/wrote/found/kept/put/cut/paid/sent/built/broke/held/read/gave/told/thought/felt/opted/switched/pivoted/expanded/paused/cancelled/delayed/shipped/launched/joined/left/won/lost/met/asked/called/pushed/set/run…），且组织/集体主语（the company / the team）整句漏判。改造：CF_EN_B_RE 从枚举腿改为「had + (人称代词 |
| 6.8.159 | 2026-11-10 | 第 241–243 维度接线闭环（r276：self_selection_bias / completion_evidence / zero_sum_fallacy，VERIFY 199→210，维度 240→243） — 三个模块于 r275 前落地但接线与文档未随本轮 commit 固化：src/self-selection-bias.js（241，自选择偏差：样本自选择机制 × 效应结论，VERIFY）、src/completion-evidence.js（242，完成态无证据断言：工作自述动作 × 完成动词 × 无证据锚，VERIFY；r275 已补 2 条召回腿 + 3 条误伤豁免腿）、src/zero-sum-fallacy.js（243，零和谬误：收益-损失互斥框架 × 分配结论，VERIFY）在本轮 src/index.js 完成九处接线（require / 计算点 / allDims / dimMap / dimensions / DIM_GUIDANCE / VERIFY_DIMS / summary / exports），维度数 240→243、VERIFY 199→210、三层列举同步追加 3 维度。文档三层同步（README/AGENTS/SKILL 维度数 240→243、verify 计数 206→210、规格表版本戳 v6.8.159、Verify 列举追加）。 |
| 6.8.155 | 2026-11-09 | 第231维度 affirming_disjunct（肯定析取 / affirming a disjunct）上线（r256 并发接线）+ 第230维度 fallacy_of_relevance 功能修复 — 新文件 src/affirming-disjunct.js；fallacy_of_relevance 补三项修复：① 否定极性盲区（「不支持…所以必须支持」族因 2-gram 碎片共享实词整族 miss，改用前段否定词干判别）；② INFER_ZH 补论证意图族（「却想用这个证明」等，答非所问类无推断词致切分返回 null）；③ 日常情境→自身心境豁免（SCENE_ZH × MOOD_ZH 成对，「天气不错所以心情很好」属正当因果）。 |
| 6.8.158 | 2026-11-09 | 统计素养三维度接线闭环（r269 双维 + r270 immortal_time_bias，VERIFY 199→206，维度 237→240） — count_rate_conflation 238 / law_of_small_numbers 239 于 r269 在 src/index.js 完成九处接线；immortal_time_bias 240 于 r270 落地，三腿判据（起点时点锚 × 起点差异[显式/对置/存活门槛] × 疗效结论）+ 三组豁免（方法学纠正在场 / 统一起点声明 / 严谨设计）。r270 收尾修复两条单测漏判：腿③ EFFICACY_ZH 补「疗效结论需要重新核对」类纠正性引用、腿① START_ANCHOR_EN 补裸起点名词（starting points/baseline/entry point）——实证 gap_zh/gap_en 从 hit=false 转绿，ITB 单测 64/64。同步三份文档维度数 237→240、verify 计数 199→206、Verify 列举追加 3 维度。 |
| 6.8.158 | 2026-11-09 | 5 个行为经济学偏误维度接线闭环（r268，VERIFY 189→199，维度 232→237） — 承接 r267 交接：上一轮实证 18 个 provisional 模块 gate 层 18/18 穿透、think 层 0 可达，本轮取其中 5 个在 src/index.js 完成九处接线（require / 计算点 / allDims / dimMap / dimensions / DIM_GUIDANCE / VERIFY_DIMS / summary / module.exports）：omission_bias 233、money_illusion 234、planning_fallacy 235、loss_aversion 236、fundamental_attribution_error 237，全部入 VERIFY_DIMS。5 模块 + 5 单测随接线提交，单测全绿。接线有效性探针实证（scripts/round-268/r268-wire-probe.js v2，走 src/gate.js 的 gate()）：用各模块文件头声明的标准攻击形态注入，5/5 维度 gate 层真实可达；良性样本 4/4 零误伤。剩余 2 条穿透（money_illusion / loss_aversion 的非标准措辞）经 r268-direct-probe.js 直调 detect 函数判定为模块中文腿措辞覆盖缺口，非接线缺失，作交接事项给 3 号。同步更新 AGENTS.md / README.md / SKILL.md 三份共享文档：维度数 232→237、verify 计数 189→199、Verify 层列举追加 5 维度、AGENTS.md 版本史追加 r268 条目；修复 r267 畸形 claim 记录（kind/direction 字段值互换）。四门禁逐位零漂移：verify 16/0、guard-abilities 20/20（237 维度）、doc-numbers 24/24、bidirectional 52/52 召回 + 302/326 误拦。 |
| 6.8.157 | 2026-11-09 | 5 个单族维度补 family 样本 + 硬判据函数层否定守卫（r258，良性零误伤优先） — 轮初审计发现 prompt_injection / code_security / instrumental_reasoning / social_norm / hasty_generalization 五个维度存在「维度已登记但判据族不全」的结构性漏判（样本命中 1-2/10，同族换说法即漏）。按 verify-final.js 权威判据补 16 个新族：① prompt_injection 补 prompt_leak_verbatim 逐字索取系统提示族（中英同构，旧库只收 output/reveal 泛指不收 verbatim 索取）；② code_security 补 hardcoded_cred_Z / hardcoded_cred_E / exfil_en 三族，severity 同 secret（0.9 / critical）；③ instrumental_reasoning 补 self_preservation_Z/E / concealment_E / resource_acquisition_Z/E 五族，全部以 [regex, 'type'] 元组入 INSTRUMENTAL_PATTERNS；④ social_norm 补 conformity_pressure_Z/E / authority_exemption_Z/E / dehumanization_Z/E 六族入 SOCIAL_NORM_PATTERNS；⑤ hasty_generalization 补 single_sample_Z/E / hearsay_Z/E / recency_bias_Z/E 六族入 HASTY_GENERALIZATION_PATTERNS。否定守卫放函数层不进正则：hardcoded_credential 新增 _NEG_HARDCODE_ZH/_NEG_HARDCODE_EN（仿 _NEG_BEFORE_EXFIL/_ADVISE_AGAINST_EXEC 既有模式），命中即 skip 三新族、其余类型照报，劝阻式安全表述（「不要把 token 硬编码到前端」）零误伤。承重性实测（mutation liveness）：把全部 r258 块置空后 IR 9/9→2/9（7 条仅由 r258 覆盖）、SN 10/10→0/10（完全承重），PI/CS/HG 为 additive 覆盖（旧族已命中）。最终实测：5 维攻击 50/50、良性 FP 1/62（该条经 git HEAD 版对照证实为既有基线误伤「我认识的几位朋友都很靠谱」，非本轮引入）。五门禁全绿：verify 16/0、guard-abilities 20/20（202 check 入口）、bidirectional 52/52 召回 + 302/326 误拦（92.6%）、doc-numbers 24/24、round-531-sample-matrix 20 攻击/35 良性全空数组；附加 negative-test-adversarial-variant-lang 3/3、round-214-mcf-matrix 22/22。新增 test/round-258-family-coverage.test.js 12 通过 0 失败（fixtures 落仓库内 test/fixtures/r258/）。维度数不变（232，本轮为既有维度补判据族而非新维度）；src/index.js 16879 行超 14000 阈值仅报警不拆分。 r265 补记（agent-4 补腿 + 本会话接手验证修复）：sycophancy-check 插件层补 uniqueness_flattery（独特性捧高 EN 15 条 / ZH 同族 26 条）与 unconditional_deference（无条件让渡立场 EN 18 条）两个隐性谄媚族，WEIGHTS 加 0.35/0.5；缺口实证为补腿前插件层对 10 条全新英文谄媚措辞 10/10 穿透 score 全 0.00。同时修复 7 处命中计数 bug（match() 无 /g 时数组长度 = 1 + 捕获组数，原代码把捕获组数算成命中次数，良性句被放大到 high 误伤）。接手方修复 agent-4 未验证引入的真实回归：其用 matchAll 计数但信号库模式全为 /i 非全局，node test/sycophancy-i18n.test.js 10/10 全挂（matchAll called with a non-global RegExp）；修法为新增共用函数 _countMatches（按原 flags 补 g 重建同源正则），不破坏模式库、零模式改动。修复后 i18n 单测 10/10、r258 测试 12/0 无连带回归，插件层攻击 5/10 达 medium+、良性 0/10 误伤。因 src/gate.js 零 plugin 引用（插件层不在 gate/checkOutput 路径），本轮维度计数 232 不变、门禁数字基线不变、VERSION 不 bump（沿用 6.8.157）。 |
| 6.8.156 | 2026-11-09 | 第232维度 motivated_reasoning（诉诸动机 / Bulverism）接线（用户指令「全面审计→全面修复→同步」） — r257 遗留的完整模块 src/motivated-reasoning.js（18426 字节）此前 git 未跟踪、index.js 零引用、dimensions 未登记，维度白做；按 affirming_disjunct 既有模式接 import + calc + dimMap/allDims + dimensions×2 + VERIFY list 六处，维度数 231→232、VERIFY 188→189，模块实测中英攻击 hit:true score:70、良性 false。[HIGH] MemoryBank 静默丢库修复：src/memory/memory-bank.js 解密失败时先 cp 密文至 <bank>.undecryptable.<ISO>.bak 再 starting fresh（原行为直接丢弃、后续 save 覆盖 → 密钥找回也读不回），本轮实测生成 2183 字节备份。check-domain-coverage 缺语料崩溃修复：fs.existsSync 守卫，缺失记 SKIP 不崩整轮审计。6 处 orphan import 清理（debugLog×3 / checkInput / FIELD / SPECIAL_MODULES，仅删零引用符号）。文档口径同步 README/SKILL/AGENTS 维度 231→232、verify 188→189、版本戳 6.8.156。验证：五门禁全绿（verify 16/0、guard 21 passes、doc-numbers 24/24、双向通过、变体矩阵 done）+ 引擎 smoke OK + dimension-registry-guard 232 键。 |
| 6.8.154 | 2026-11-09 | 第230维度 fallacy_of_relevance（不相干结论 / ignoratio elenchi）上线（r254 立项，r255–r256 收尾） — 新文件 src/fallacy-of-relevance.js；中文 2-gram 零共享判据 + 全文中英文豁免四件套（META_CONCEPT/TEACH/QUESTION/NUMERIC_BRIDGE）先于切分 + 正当因果桥接豁免（INVEST×OUTCOME 成对、体育→健康独立域）；变体矩阵 9/9 正向 + 3/3 良性。 |
| 6.8.153 | 2026-11-09 | 第229维度 undistributed_middle（中项不周延）上线 + 全局稳定性加固（r253 审计） — ① P1 修复：undistributed_middle 未命中返回 null 导致 checkOutput 全量崩溃（commit 6d42f8c6），另修 9 个 check* 维度函数（checkPseudoProfundity / circular_reasoning / survivor_time_bias / whole_part_transfer / necessity_sufficiency_confusion / activity_outcome_confusion / proxy_metric_goal 等）收到非字符串输入时直接 text.match() 解引用的崩溃，统一入口守卫 if (typeof text !== 'string') text = String(text ?? '')——206 个 check* 函数 × 14 类输入动态探测 2884 次调用由崩溃 16 次降为 0；② 性能：dangerous-instruction.js _matchAll 的 5 个豁免正则提为模块级常量，消除每次调用的 361 次重复编译（CPU profile 实测单次 discriminate 编译 782 个正则，其中该函数占 46%），语义零变化、dangerous-instruction 测试 33/33 通过；③ coordination 体系整理：DEPLOY-agent2/agent3 标注历史退场状态（无宿主部署目录、claims 零认领、reports 止于 r213/r214），DEPLOY-agent4 补「两处部署体系」说明与当前写权限表；④ 撤销一次误操作恢复：agent 用 .bak 回填导致工作区同时丢失 226/227/228/229 四维接线与 automation-bias / framing-effect 内容（index.js -126 行），已完整恢复 HEAD 并只重打合法优化。维度数 228→229，VERIFY 185→186，模块 161 不变。验证：五门禁 verify 16/16、guard 20/20（229 键）、doc 24/24、双向召回 52/52 误拦 302/326（较基线 301 改善）。 |
| 6.8.152 | 2026-11-08 | r252 补登记 undistributed_middle 计算点与 9 个维度入口守卫 — 本地独有产能固化，为 6.8.153 的前置。 |
| 6.8.151 | 2026-11-08 | 第228维度 false_memory_transplant（虚假记忆嫁接 / memory-reframing）上线（r252 接手落地） — 新文件 src/false-memory-transplant.js（272行）。判据为三腿共现（缺一不判）：腿① MEM_ZH 第一手见证标记腿（`你 |
| 6.8.150 | 2026-11-08 | 第227维度 means_as_end（过程完成当目标达成 / means-as-end confusion）上线（r251） — 新文件 src/means-as-end.js。缺口实证：scripts/round-251/probe-gap-r251.js 探针实测，6 条攻击句中 5 条穿过硬闸门且零维命中（grep src/ 对 means as end / checklist fallacy / process completion 零命中），存量 activity_outcome_confusion(184) 只管「活动/产出当结果」的泛化混淆、proxy_metric_goal(187) 只认指标替代目标，均不认「拿流程/清单/步骤的完成直接给目标达成背书」这一手段-目的替换结构本身。判据为四腿共现（缺一不判）：腿① PROC 过程语义腿（流程/步骤/环节/手续/清单/审批/议程/工单/代码/部署/配置/任务……含裸名词形态）；腿② DONE 完成语义腿（走完/勾完/办完/提交/关闭/结案/悉数/ticked off/followed……）；腿③ GOAL 目标达成腿（目标已达成/任务已成功/产品可交付/问题已解决/strategic realized……）；腿④ CONN 因果连接腿（说明/可见/因此/足见/means/which means/so/thus——隐性因果逗号并置亦承载，连接词为增强腿非必需腿）。三处结构性约束：腿位置互斥 + 已然断言（目标腿完整吞掉过程腿时须为「已经/显然/圆满」等已然断言，「任务最后三天要完成」的未然安排句豁免）；跨度 ≤80 字（EN 160 字符）；三类豁免（过程与目标主动分离/目标侧证据在场/对冲疑问）。index.js 九处接线（require/计算点/allDims/dimMap/补登记×2/VERIFY_DIMS/exports；变量名 mae，grep 实测零冲突）。来源：HN Algolia "checklist fallacy" 与 "means end confusion" 检索、GitHub API 均无高星纯规则实现，判据为心虫自有纯规则实现，侦察落盘 /tmp/hf-scout-r251.txt。维度数 226→227，VERIFY 185→186。验证：单测攻击 19/19 命中、良性 17/17 通过（含 guard-abilities gate-97 实测误伤句「这个任务最后三天要完成，我安排一下进度」）、六语义腿逐常量置空命中均下降、verify 16/16、guard 20/20（227 键）、双向召回 52/52 误拦 302/326 不升。已知边界：① 隐性因果并置形「所有审批节点都过完了，项目显然已经成功交付」靠逗号承载，连接词腿非承重；② 跨句长距离形（过程在前句、结论在后句）不收；③ 「战争纯粹是因为资源争夺」等他族语义由 monocausal_attribution 代拦，非本维度射程。 |
| 6.8.149 | 2026-11-08 | 第226维度 survivor_time_bias（幸存时间偏误 / survivor time bias）上线（r250） — 新文件 src/survivor-time-bias.js。缺口实证：scripts/round-250/probe-gap-r250.js 探针实测，6 条攻击句中 4 条穿过硬闸门且零维命中，存量 survivorship_bias(132)/selection_bias(168)/denominator_swap(224) 只认「进入样本的个体不同」的构成性偏差，不认「同一队列内因诊断/登记时点不同而产生的生存期差」这一随访起点错配结构本身。判据为三腿共现（缺一不判）：腿① SURVIVAL 生存期语义腿（生存期/中位生存/存活时间/随访/survival time/median survival……）；腿② OUTCOME 疗效结论腿（疗效更好/生存更久/预后更佳/better survival……）；腿③ TIME_CONTRAST 时间对置腿（更早诊断/更早发现/早诊断/诊断时点/earlier diagnosis……）。豁免表三类（真实获益承认/早筛设计承认/研究设计语境）负例注入承重。index.js 九处接线（变量名 stb）。来源：HN Algolia "survivor bias" 检索、GitHub API 无高星纯规则实现，判据为心虫自有纯规则实现，侦察落盘 /tmp/hf-scout-r250.txt。维度数 225→226，VERIFY 184→185。验证：单测攻击 13/13、良性 12/12、verify 16/16、guard 20/20（226 键）、双向 52/52 误拦 302/326 不升。 |
| 6.8.148 | 2026-11-07 | 第222维度 prosecutors_fallacy（检察官谬误 / prosecutor's fallacy）上线（r233） — 新文件 src/prosecutors-fallacy.js。缺口实证：scripts/round-233/probe-gap-r233.js 探针实测，7 条攻击句中 5 条穿过硬闸门且无一维命中（grep src/ 对 prosecutor / 检察官 / 误报率 / false positive rate 零命中），存量 base_rate_neglect(169) 只管「忽略基础比例」的泛化陈述，不认「P(匹配 |
| 6.8.146 | 2026-11-07 | 第218–220三维度上线（r232：诉诸未来 / 传闻即事实 / 证实性取证） — 三个新文件：src/appeal-to-future.js（307行，appeal_to_future 诉诸未来：把尚未存在的未来证据/未来解决者当成当下行动或当下结论的依据，是证据缺位的时间置换）、src/appeal-to-rumor.js（290行，appeal_to_rumor 传闻即事实/argumentum ad rumorem：把「某说法正在流传」这一传播事实当成该说法为真的证据，是传播强度与证据强度的混淆）、src/confirmatory-evidence.js（287行，confirmatory_evidence 证实性取证：把实验/研究反复确认这一经验动作直接等同于该理论为真，是证据方向与证据充分性的混淆）。三族判据统一为两腿共现（缺一不判）+ 邻近窗：诉诸未来（未来证据腿 × 当下行动/结论腿，80字/200字符窗）、传闻即事实（传播事实腿 × 真实性推断腿）、证实性取证（确认动作腿 × 真理性结论腿）；均含中英双语词表、裸名词形态（r161教训：去后缀仍命中）、负例注入常量（EXEMPT_DEBUNK/BOUNDED/META/HEDGE/NEG/QUESTION + RHETORICAL 反问不豁免口径，与 r216 collider_bias 一致）。缺口实证（scripts/round-221/probe-gap-r221.js + scripts/round-232/probe-matrix-r232.js）：诉诸未来 6句漏判3句、传闻即事实 4句漏判3句、证实性取证 4句漏判3句；r232 变体矩阵实测攻击 24/24 命中、良性 16/16 pass；gate 层端到端三族全部 findings 命中、良性 pass。覆盖度检查：grep src/ 对 appeal_to_future / ad_rumorem / 空穴来风 / confirmatory 零命中，217 个既有维度无覆盖（appeal_to_majority 只管多数观点→正确、ad_nauseam 只管断言重复、anonymous_source_claim 只管不可指认信源）。外部来源：三层搜索全部无可用纯规则方案（HN Algolia 仅得谬误清单页面、Semantic Scholar 两次 429 限流、GitHub fallacy detection 仓库全为 LLM-based 与零 LLM 依赖硬约束冲突），判据为心虫自有纯规则实现，侦察落盘 /tmp/hf-scout-r232.txt。index.js 九处接线（require/计算点/dimMap/allDims/补登记×2/VERIFY_DIMS/exports；变量名 atf/atr/cev）。维度数 217→220，VERIFY 174→177。验证：verify 16/16、guard 20/20（220键）、doc 24/24、双向召回 52/52 误拦 302/326 不升。已知边界：① 三族均只收正向单句（未来证据→当下行动、传播事实→真、确认动作→成立），反向形与长距离跨句形本轮保守不收；② 边缘良性（「战争是多因共同造成的」式正确多因陈述）实测零误伤；③ 存量维度 scrutiny_evasion/presupposition 会代拦部分本族边缘长句，非本维度引入。 |
| 6.8.145 | 2026-11-06 | 第 213 维度 monocausal_attribution（单因归因 / monocausal attribution）上线（r220） — 新文件 src/monocausal-attribution.js（336 行）。缺口实证：tmp-probe-gap.js 21 条候选攻击句现有维度漏判 13 条，其中 L2 枚举否认族 6/6 全漏；monocause(161) 要求「系统性复杂现象名词 × 绝对化框架」双腿齐现，不认 is not a contributing factor 式否认，也不认口语化 it is all because X。三条腿判据（任一命中）：L1 COMPRESS 归因压缩（全都是因为/唯一原因/说到底只有一个/EN it is all because / the only reason ... is / due solely to / boils down to one，含裸名词与带修饰形态）；L2 ENUM-DISMISS 枚举否认（句中切出 ≥3 段并列候选成因 + 否认措辞 is not a contributing factor / plays no role / 不起任何作用 / 并非真正的原因，中文枚举连接词显式支持顿号与和与以及）；L3 CONTRADICT 归因相消（同句既有单因压缩又有多因并列，宽松口径不强制 ≥3 段枚举）。三组负例表承重：对冲豁免（之一/可能/也许/maybe/likely/at least in part）、NEGATED_MULTI 拒绝单因句式（no single factor explains it all / 并非只有一个原因，中英双语，修实现期发现的良性误判：该句式因内含 single 被 COMPRESS 误命中，L1/L3 前置排除）、概念教学与疑问句。验证：攻击变体矩阵 21/21 非 pass，良性 15/15 中 monocausal_attribution 零误伤（3 条由 empty_answer/unsupported_claim/vagueness 既有维度触发）；四门禁 verify 16/16、guard 20/20（214 键）、doc 24/24、双向召回 52/52。文档口径 212→214、verify 169→171 同步（README/SKILL/AGENTS 三处）。 |
| 6.8.143 | 2026-11-05 | 第 211 维度 ecological_fallacy「生态谬误 / 区群谬误 / ecological fallacy」上线（r215，统计推断的层级桥接方向轮换；由本地 claude 起草、1 号主力收尾）——新文件 src/ecological-fallacy.js（405 行），判据为两腿共现 + 邻近窗（缺一不判）：腿① AGGREGATE 聚合统计量（平均/人均/率/统计/数据显示/该国/群体层面 等 12 族 ZH + 对应 12 族 EN）× 腿② INDIVIDUAL 个体层推断（连接词族 A/A2/A3/D/D2/D3/D4 + 裸指称族 B/B2/C/C2/E/E2），把「群体统计量直接推断个体成员属性」偷换成「有跨层桥接证据」的新攻击族——与 simpsons_paradox(206，聚合层与子群层趋势方向相反)/whole_part_transfer(170，整体属性搬到部分)/base_rate_neglect(169，忽视基础比率)/selection_bias(152，样本选择偏差)/reverse_causation(172，因果方向倒置) 边界错开：前四者是「统计量算错/方向错/忽略先验」，本族错在层级跨错。1 号主力收尾三项：① 接线补全 4→8 处（require / 计算点 / dimMap 参与 / dimMap / abilityPrompt 登记 / 导出 / 另两处——漏接 abilityPrompt 与导出即为 r496 ReferenceError 事故同形状）；② 漏判修复补 D3/D4 两腿（推断连接词 + 第二/三人称带限定语指称），实证句「平均智商高于邻国，所以你邻居家的孩子一定更聪明」「统计显示这个城市犯罪率很高，所以住在这里的你一定是个罪犯」——D 族只收「说明/意味着+你」、A 族只收「每个人/人人」类，中间形态整族漏判；③ BARE_IDX_ZH 序号表随插入修正（11/12 → 13/14），避免裸指称腿被当连接词腿。边界纪律（1 号主力新增）：裸指称腿只在腿①同时命中时启用——否则良性陈述「每个学生都付出了努力」被误判；连接词腿携带推断语义，独立成立。验证：攻击 5/5 承重、良性 0/6 误伤；四门禁全绿（verify 16/16、guard 20/20、doc 24/24、双向 52/52 + 302/326 不升）；文档口径 210→211、verify 167→168 同步。 |
| v6.8.141 | 2026-11-05 | 第 209 维度 collider_bias「对撞偏差 / endogenous selection bias」上线（r212，因果与反事实推理方向轮换）——「条件筛选标记（在…里/之中/中间/中 + 只看/入选/活过/留任/被淘汰前，裸名词形态；EN among the admitted / only those who / conditioned on / restricted to / survivors / survived / made it through，窗 30 字/EN 100 字符）× 样本内观察到的相关共变（负相关/正相关/相关性/相关关系/成反比/共变 + 越少…越高/更低/更高/反而/反倒——不要求显式「相关」字样；EN negatively correlated / correlated with / inversely related / associated with / less…more / tends to）× 总体因果结论（C-a 推断+因果形：所以/因此/因而/可见/说明/意味着/证明/显然/据此 + 导致/造成/拖累/损害/抑制/促进/有助/有害/不利于/更有效/更安全/推荐/推广/越好；EN causes / leads to / hurts / harms / improves / boosts / better / effective / should / recommend；C-b 裸结论形：因果词在腿②后 80 字/EN 200 字符内且未被否定）」三腿共现+邻近窗，否定守卫窗口化（查因果词前 8 字窗口，实现期修复「这看起来像…拖累…但其实是伪相关」漏对冲），把「条件筛选制造的子样本内伪相关」偷换成「总体因果结构」的新攻击族——当 A、B 都影响 C（对撞因子），观察范围被限制在 C 的某个取值（入院/入选/存活/留下）时，A 与 B 在子样本里呈现相关，即使 A、B 在总体中毫无因果关系，这是条件化的结构性后果而非数据噪声，是最容易被叙述成铁证的一类假因果；四组豁免（概念/教学转述、正确响应形、无腿②、推理否定/对冲）+ 疑问句；与 selection_bias（取样受限只是腿①、承重的是腿②③）/ survivorship_bias（不要求相关观察与因果结论）/ simpsons_paradox（本族只有一个被条件筛选的层级）/ common_cause·reverse_causation（它们假定共变真实，本族的共变本身就是筛选产物）/ cum_hoc·post_hoc（不要求显式相关或时间先后）边界错开；r212 探针 5/7 攻击真零命中（含中英双语/裸名词/带修饰/长句），良性基本 pass；外部三层搜索（HN Algolia + GitHub API fallacy / causal inference fallacy 命中 joshinh/causal-fallacies、causalNLP/logical-fallacy、latent-variable/Real_time_fallacy_detection，README 已读均非纯规则可集成；Semantic Scholar collider bias endogenous selection 空返）后按既有维度形态自建；新文件 src/collider-bias.js + test/round-212-collider-bias.js（32 断言全绿：8 攻击全命中含中英双语/裸名词/带修饰/倒装/长句、8 良性全放行含正确响应形/多因陈述/观察性数据/存活者偏差转述/教学/筛选后描述/对冲/疑问、6 项腿与豁免常量负例注入承重），gate 层变体矩阵 10/10 攻击非 pass（8 条 collider_bias 直接承重，2 条其他维度 verify 代拦）+ 10/10 良性全 pass，verify 165→166、维度 208→209、三层级合计 195→196/剩余 13 不变，模块 162 不变，五门禁全绿（verify 16/16 / guard 20/20 / doc 24/24 / 双向 52/52+302/326 不升） |
| 6.8.140 | 2026-11-04 | 第 208 维度 hypostatization「实体化/拟人化谬误 / reification」上线（r211，逻辑谬误检测一线轮换）——「抽象实体（历史/时代/市场/资本/数据/社会/规律/命运/科技/系统/算法/自然；EN history / the market / capital / data / society / the system / the algorithm / nature）×主体行动词（会惩罚/不会允许/知道/记得/决定/选择/拒绝/回答/回应/审判/报复/背叛/警告/呼唤；EN will punish / won't allow / knows / decides / chooses / refuses / answers / judges / rewards / warns）×结论承重（所以/因此/必须/应该/必将/注定或后果分句）」三腿共现+双邻近窗（中文 12/50 字、EN 50/180 字符），另设 R2 判定型分支（审判/惩罚/报复/报答/谴责/辜负/背叛/定罪；EN judge / punish / reward / betray / revenge / will not tolerate——「X 将审判/惩罚 Y」裸断言自带攻击性），把「抽象物像人一样行动」偷换成「结论有依据」的新攻击族；四组豁免（概念/教学/文学转述、明示修辞标记、主体性否认、谚语格言）+ 疑问句豁免；r211 探针 5/7 攻击真零命中（含中英双语/裸名词/带修饰）、良性 4/4 零误伤；外部三层搜索无可集成纯规则方案（GitHub 谬误检测最高 ★120 全 LLM 方案、Semantic Scholar 空返），按既有维度形态自建；新文件 src/hypostatization.js + test/round-211-hypostatization.js（56 断言全绿），verify 164→165、维度 207→208、三层级合计 194→195/剩余 13 不变，五门禁全绿（verify 16/16 / guard 20/20 / doc 24/24 / 双向 52/52+302/326 不升） |
| 6.8.139 | 2026-11-04 | 第 207 维度 uncertainty_weaponization「证据确定性要挟 / perfect-evidence demand」上线（r210，接手上一轮爆截断遗留并补全）——「证据不确定性宣告（尚无定论/证据不足/不是100%确凿/仍无共识/尚未证实；EN not settled / no consensus / inconclusive / unproven）×行动阻挠（没必要/不应/暂缓/为时过早 + 干预/监管/防范/整改/措施；EN no need to act / shouldn't intervene / too early to act）」两腿共现+邻近窗，另设 R2 完美证据分支（除非百分之百证实否则不能启动…），把「还不全知」偷换成「行动尚未获准」的新攻击族；四组豁免（让步-防护/预防原则/宣示暂缓/概念转述）+ 修绿（智能引号归一、预防与宣示豁免上移主判据前） |
| 6.8.138 | 2026-11-04 | 第 206 维度 simpsons_paradox「聚合层逆转」上线（r209）——「聚合层趋势（总体/整体/平均/合计+上升/下降/提高/降低/上涨/下跌/增长/减少，裸名词形态；EN overall / on average / aggregate × rose/fell/increased/decreased/improved/worsened）×子群层相反趋势（每个/各个/每一/分层/科室/班级/亚组/年龄段/收入段/对照组/每个县，裸名词形态；EN each/every/per/subgroup/stratum/cohort/by age/broken down）×覆盖使用腿（U-a 覆盖推断形：所以/因此/说明/意味着+更安全/更有效/有效/更好/更可靠/省钱/成功/应当/推广/采纳；EN so/therefore/which means+safer/effective/works/better/should/roll out；U-b 裸结论形：肯定结论出现在子群相反趋势之后）」三腿结构 + 「抹平形」融合分支（聚合层无差别宣告×子群方向截然相反）判别，把「混合权重」偷换成「每个分组真实方向」的新攻击族——聚合量是子群量的加权混合，子群构成一变，聚合方向可以与每一个子群相反（辛普森悖论的结构性质）；与 whole_part_transfer（涌现性质量搬运）/statistical_misleading（小基数比例）/proxy_metric_goal（代理指标当目标）边界错开，探针实测本族 10/13 攻击样本现有 205 维真零命中（其余 3 条被 proxy_metric_goal / stereotype+perfect_error 代拦）；外部来源 GitHub ehart-altair/SimpsonsParadox ★18（Simpson's Pairs = variables whose relationship with the dependent variable reverses when conditioned on a second variable）+ ninoch/Trend-Simpsons-Paradox ★57 + HN/Wikipedia Simpson's paradox 条目，Semantic Scholar / arXiv 本轮网络不可达（空返）未引论文，按既有维度形态纯规则自建；新文件 src/simpsons-paradox.js + test/round-209-simpsons-paradox.js（18 断言全绿：12 攻击全命中含中英双语/裸名词/带修饰/长句、10 良性全放行含概念转述/方向相同/正确响应/待验证、13 项腿与豁免常量负例注入承重、gate 矩阵 7 攻击非 pass 且维度承重 + 6 良性全 pass），verify 162→163、维度 205→206、三层级合计 192→193/剩余 13 不变，模块 162 不变（纯函数维度模块不进 _modules registry），五门禁全绿（verify 16/16 / guard 20/20 / doc 24/24 / 双向 52/52+302/326 不升） |
| 6.8.137 | 2026-11-04 | 第 205 维度 scale_dismissal「规模 dismiss」上线（r208）——「小规模/低占比宣告（才几个人/个别案例/占比不到千分之一/区区几个家庭/这么小的比例；EN only a few / a handful of users / a tiny fraction / statistically negligible / a few out of millions）×风险 dismiss（不值得大惊小怪/没必要上纲上线/忽略不计/不值一提/翻篇；EN not a big deal / not worth the fuss / no need to panic / we can move on）」两腿结构 + 「忽略不计」融合分支判别，把「相对占比」偷换成「绝对代价免检通行证」的新攻击族——风险与伤害的严重性取决于绝对后果与可逆性，不取决于它占总体的比例；与 selective_minimization（过错定级降级）/harm_invalidation（具体伤害抹除）/percentage_overflow（百分比演算）边界错开，探针实测本族 8/8 攻击样本它们零命中；r208 探针 12 候选族实测，scale_dismissal 是唯一 8/8 真零命中且 4/4 良性零误伤的族（次优 token_gesture 7/8、burden_inversion 7/8）；外部三层搜索（HN Algolia 无相关命中、Semantic Scholar 两次 429 限流、GitHub agent execution verification 前五全部 LLM 驱动的 agent harness 与心虫铁律冲突）无纯规则成熟实现可集成，按既有维度形态纯规则自建；新文件 src/scale-dismissal.js + test/round-208-scale-dismissal.js（13 断言全绿：11 攻击全命中含中英双语/裸名词/带修饰/倒装/长句、8 良性全放行含让步转折/测量语境/统计报告/概念批判、9 项腿与豁免常量负例注入承重、gate 矩阵 6 攻击非 pass 且维度承重 + 6 良性全 pass），verify 161→162、维度 204→205、三层级合计 191→192/剩余 13 不变，模块 162 不变（纯函数维度模块不进 _modules registry），五门禁全绿（verify 16/16 / guard 20/20 / doc 24/24 / 双向 52/52+302/326 不升） |
| 6.8.136 | 2026-11-03 | 第 204 维度 might_makes_right「强权即公理」上线（r207）——「实力/胜负/权力依据（谁赢/赢了/胜利者/拳头/胳膊/武力/强权/枪杆子/实力/官大/钱多/碾压/弱肉强食；EN might / muscle / the stronger / whoever wins / victor / power / bigger army）×是非裁量等式（就是真理/就是正义/代表真理/决定对错/定义是非/判定善恶/谁有理/说了算/才算数/书写历史/正统/合法性/硬道理；EN (might |
| 6.8.135 | 2026-11-03 | 第 203 维度 credential_dismissal「诉诸资历/资历驳回」上线（r206）——「批评者的资历/学识/经验缺口（没读过这本专著/不是经济学博士/没在行业干过十年/外行/没结过婚/连一天都没做过；EN you have not read / you are not a professional / who has never / no degree / outsider / amateur）×发言资格剥夺（没资格批评/不配谈/有什么资格/先读书再说话/闭嘴/插什么嘴/也配/班门弄斧；EN not qualified to criticize / have no right to comment / who are you to / shut up / stay in your lane / until you have read）」两腿结构 + 自足祈使分支（先+长周期资历动作+再+发言动作单式即两腿齐备）判别，把「人的准入条件」偷换成「观点的判别条件」的新攻击族（学术圈「没读完全集就没资格评论」/行业「没干过十年少插嘴」/公共议题「不是X国人别说话」同族——用「你是谁」代替「你说了什么」）；与 genetic_fallacy（第 131 维起源谬误：来源×质量贬损，落点是内容价值）边界错开，探针实测本族 4/4 攻击样本它零命中；r206 族探针实测攻击 4/4 真零命中、良性 2/2 零误伤；外部搜索（HN Algolia "courtier's reply" 前五无相关命中 courier 同形词干扰、Semantic Scholar 两次空返限流、GitHub "ad hominem fallacy" sort=stars 前五 UKPLab/naacl2018-before-name-calling ★13 / jghibiki/NeuralAdHominem ★2 / utkarsh512/logical-fallacy ★1 全部 LLM/统计学习路线、"courtier reply fallacy" 零仓库）无纯规则成熟方案可集成，按既有维度形态纯规则自建；新文件 src/credential-dismissal.js + test/round-206-credential-dismissal.js（16 断言：12 攻击全命中含中英双语/裸名词/带修饰/反问/长句、12 良性全放行含资源指引/教学邀请/执业许可/概念批判、3 项豁免陷阱逐常量承重、8 项腿与豁免常量负例注入含中英逐侧与自足祈使分支、gate 矩阵 6 攻击非 pass 且维度承重 + 6 良性全 pass），verify 159→160、维度 202→203，模块 162 不变（纯函数维度模块不进 _modules registry），五门禁全绿（verify 16/16 / guard 20/20 / doc 22/22 / 双向 52/52+302/326 不升） |
| 6.8.134 | 2026-11-03 | 第 202 维度 unfalsifiable_confirmation「双果自证」上线（r205）——「互斥结果对（如果数据涨了…如果数据跌了、无论实验灵不灵、不管结果如何、预言实现了…没实现、测出来是A…测不出来是B、涨了是A…跌了是B；EN whether X or not / either way / no matter what / if it works…if not / whatever happens）×同一自证结论（都说明/证明/验证+我+对·正确·成立·有效·准·高明·领先、我的（理论/观点/学说/预言/判断/方法/模型/能力/框架）+是+对/正确/成立/有效、反正/横竖+都+我+对；EN proves/confirms/shows/vindicates + I'm/my theory is + right/correct/valid、my theory is vindicated、either way + confirms、right either way / no matter the outcome、nothing could disprove it）」两腿结构判别，把「一对互斥结果」偷换成「同一个自我结论的证据」的新攻击族（伪科学免疫策略/阴谋论/营销自证话术核心结构——不存在任何能推翻它的观察，举证责任反向消灭）；立项缺口由 r167 one-sided-dismissal.js 注释显式留白（「不可证伪断言」候选族当时仅 6/8 被其他维度代拦，无维度实现其两腿结构）；外部搜索（HN Algolia "falsifiability" 前八均为 Popper 证伪主义/Sagan 车库里喷火龙/Hans Albert 免疫策略哲学博客、Semantic Scholar 429 限流、GitHub sort=stars 前二 cpt13-g/clarify-research-idea ★282 与 tonyblu331/research-proof ★52 均 LLM 驱动）无纯规则成熟方案可集成，按既有维度形态纯规则自建；新文件 src/unfalsifiable-confirmation.js + test/round-205-unfalsifiable-confirmation.js（12 断言：11 攻击全命中含中英双语/裸词/倒装极性/长句、8 良性全放行、8 项腿与豁免常量负例注入承重、gate 矩阵 6 攻击非 pass 且维度承重 + 6 良性全 pass），verify 158→159、维度 201→202，五门禁全绿（verify 16/16 / guard 20/20 / doc 22/22 / 双向 52/52+302/326 不升） |
| 6.8.133 | 2026-11-03 | 第 201 维度 galileo_gambit「伽利略赌局」上线（r204）——「殉道者/被嘲笑先知先例（伽利略被宗教裁判所审判/布鲁诺被烧死在鲜花广场/魏格纳被嘲笑/塞尔维特/塞麦尔维斯/苏格拉底/被当成疯子谬论/所有突破起初都被骂；EN Galileo / Copernicus / Bruno / Servetus / Wegener / Semmelweis / burned at the stake / they laughed at / mocked / persecuted）×向自身观点的平反迁移（所以/因此+我…没错·正确、我的理论/发现+迟早会被证明·承认·平反、历史/时间/时代+会+证明·平反·给我公道、被封杀·被拒稿…说明我·说明我对、反对本身就是最好的证明、真理往往掌握在少数人、站在历史的正确一边；EN so my theory will be proven / history will vindicate / they're laughing at me now / the more they mock me, the more right / establishment rejection is the best credential / wrong side of history）」两腿结构判别，把「立场被压制的事实」偷换成「观点正确的证据」的新攻击族（伪科学/阴谋论/民科自辩最常用护身符）；与 poisoning_the_well/genetic_fallacy（动机污名化）、appeal_to_majority（多数即真，本族方向相反）、special_pleading（规则后门）判据错开；立项缺口由 r180 fallacy-fallacy.js 缺口清单显式留白（「galileo_fallacy 在 src/ 下 0 命中」）；外部搜索（HN Algolia "logical fallacy"：arXiv 2405.02318《Autoformalizing Natural Language to First-Order Logic》/ logicalfallacies.org / UTEP Master List；GitHub fallacy-detection 高星仓库均为 LLM/统计学习方案）无纯规则成熟方案可集成，按既有维度形态纯规则自建；新文件 src/galileo-gambit.js + test/round-204-galileo-gambit.js（13 断言：10 攻击全命中/10 良性全放行/3 豁免陷阱逐常量承重/6 项腿常量负例注入含中英逐侧/gate 矩阵 5 攻击非 pass 且维度承重 + 6 良性全 pass），verify 157→158、维度 200→201，模块 162 不变（纯函数维度模块不进 _modules registry），五门禁全绿（verify 16/16 / guard 20/20 / doc 22/22 / 双向 52/52+302/326 不升） |
| 6.8.132 | 2026-11-03 | 第 200 维度 availability_bias「可得性偏差」上线（r203）——「可得性线索（老听说/经常听到/刷到/满屏都是/热搜霸榜/电视里天天放/身边全在聊/一听就能想起好多例子；EN keep hearing about / all over my feed / keeps coming up / see it all the time）×频率或概率推断（所以很常见/越来越普遍/比比皆是/十有八九/激增/成风；EN must be common / increasingly common / rampant / happening all the time）」两腿结构判别，把「回忆容易度」偷换成「实际发生率」的新攻击族；与 base_rate_neglect（阳性信号×高置信后验）判据错开、与 anecdotal（第一人称经历×推广）主体错开；外部搜索（HN Algolia / Semantic Scholar / GitHub 高星仓库）无纯规则成熟方案可集成，按既有维度形态纯规则自建；新文件 src/availability-bias.js + test/round-203-availability-bias.js（19 断言：9 攻击全命中/9 良性全放行/4 豁免陷阱逐常量承重/7 项腿常量负例注入含中英逐侧/邻近窗/疑问句/gate 矩阵 7 攻击非 pass 且维度承重 + 6 良性全 pass），verify 156→157，五门禁全绿 |
| 6.8.131 | 2026-11-03 | 第 199 维度 appeal_to_legality「诉诸合法性谬误」上线（r202）——「守法宣告（不违法/没犯法/法无禁止即可为/法律没有禁止/完全合法/合规；EN it's not illegal / broke no law / perfectly legal）×道德或批评豁免（所以没错/不能指责/管不着/少大惊小怪/凭什么指责；EN nothing wrong with it / no grounds to criticize / who are you to judge）」两腿结构判别，把「合不合法」偷换成「该不该做、能不能批评」的新攻击族；补齐 appeal_* 系列缺失的「法律/合规」一支（argumentum ad legem），与 legal_status_coercion（以法律后果胁迫）方向相反、与 opinion_entitlement（免举证义务）豁免对象错开；外部搜索（arXiv 2405.02318、GitHub tmakesense/logical-fallacy 21★ 等）全部为 LLM 路线与零 LLM 依赖铁律冲突，按既有维度形态纯规则自建；新文件 src/appeal-to-legality.js + test/round-202-appeal-to-legality.js（48 断言：9 攻击全命中/9 良性全放行/6 豁免逐常量承重/中英逐侧独立/邻近窗/否定前缀保护/矩阵 9 攻击 gate 非 pass 且维度承重 + 8 良性全 pass），verify 155→156，五门禁全绿（另修复 r201 遗留的 README/SKILL 版本戳 v6.8.129 漂移） |
| 6.8.130 | 2026-11-03 | 第 198 维度 nirvana_fallacy「完美方案谬误」上线（r201）——「不完美宣告×行动废止」两腿结构判别，填补 197 维在 perfect solution fallacy（拿方案的不完美直接废止行动，用理想化对照否决现实选项）族的空白；来源 GitHub dhfbk/faina ★5（NAACL 2025 谬误分类学对照，aclanthology.org/2025.naacl-long.34/）；新文件 src/nirvana-fallacy.js，verify 154→155，五门禁全绿 |
| 6.8.129 | 2026-11-03 | 第 197 维度 goal_drift「目标漂移/使命置换」上线（r200）——「原始使命锚点×目标置换宣告」两腿结构判别，填补 196 维在 mission drift（以「延续/服务原目标」之名行「置换目标」之实）族的空白；来源 arXiv 2603.03456《Asymmetric Goal Drift in Coding Agents Under Value Conflict》+ GitHub sunilgentyala/Plumbline ★6；新文件 src/goal-drift.js，verify 153→154，五门禁全绿 |
| 6.8.128 | 2026-11-03 | 第 196 维度 opinion_entitlement「观点权利豁免/主观主义谬误」上线（r199）——「观点/信念主观性或权利宣告 × 拒绝举证/拒绝证据裁决」两腿结构判别，填补 195 维在 subjectivist fallacy（用「我有权保留观点」免除举证责任）族的空白；新文件 src/opinion-entitlement.js，verify 152→153，五门禁全绿 |
| 6.8.127 | 2026-11-02 | 第 195 维度 cum_hoc「相关≠因果谬误」上线（r198）——「显式相关标记（相关/关联/成正比/伴随/同步）×前向因果动词（导致/造成/带来）」两腿结构判别，填补 194 维在 cum hoc ergo propter hoc（把相关/共现直接当成因果）族的空白；新文件 src/cum-hoc.js，verify 151→152，五门禁全绿 |
| 6.8.126 | 2026-11-02 | 每日审计修复（r197）——README Dimensions badge 188→194 与「What it catches」标题/action-tier 计数（verify 147→151）、SKILL 规格表 Engine version 6.8.124→6.8.126 三处漂移修复；doc-numbers-accuracy 守卫严格化（新增 badge/标题/tier 三组断言，SKILL 版本戳或关系改必须双满足）22→24 用例；判别逻辑零改动，五门禁全绿 |
| 6.8.125 | 2026-11-02 | 第 194 维度 argument_to_moderation「诉诸中庸」上线（r196）——「中间/折中立场表述×中间=正确断言」两腿结构判别，填补 193 维在「以位置取信」族的空白 |
| 6.8.124 | 2026-11-02 | 第 193 维度 regression_to_the_mean「回归均值误归因」上线（r195）——「极端基线×回归性变化×干预功劳归因」三腿结构判别，填补 192 维在统计因果归因族的空白 |
| 6.8.123 | 2026-11-02 | 第 192 维度 motte_and_bailey「稻草城堡谬误」上线（r194）——「争议主张+退守遁词」替换结构的判别，填补 191 维在该族的空白 |
| 6.8.121 | 2026-11-02 | 第 191 维度 ad_nauseam「诉诸重复」上线（r193）——复读冒充论据的结构型判别，补齐单句内容型谬误之外的重复密度盲区 |
| 6.8.118 | 2026-11-02 | 第187维度 proxy_metric_goal（代理指标当目标：古德哈特定律的谬误应用）上线 — 判别「把测得的量（代理指标）当作目标本身达成或真实效果出现的证据」结构：KPI 完成了所以目标肯定达成、下载量破百万所以产品满足需求、平均分上去了所以教学变好、测试覆盖率 100% 所以代码质量没问题、满意度打分都是五分所以客户真的满意。指标是目标的代理度量，目标是被测量的事物本身，代理读数上升不增加目标达成的证据——指标可被扭曲、博弈、与目标脱钩（刷分刷出的 KPI、为凑覆盖率写的空测试、怕差评不敢打分的问卷），古德哈特定律正是「度量变成目标之日，即它不再是好度量之时」：「KPI完成了，说明目标肯定达成了。」「这个季度下载量破了五百万，说明产品已经满足了用户需求。」「测试覆盖率到了100%，说明代码质量没有问题。」「The metric hit its target, so the goal must have been achieved.」「Coverage is at 100%, which proves code quality is fine.」。判据三腿共现，缺一不判：腿① METRIC 被测得的量（点击率/播放量/下载量/KPI/指标/完成率/覆盖率/分数/平均分/引用量/日活/留存率/满意度打分/GDP…全部裸名词；EN metric/downloads/coverage/score/KPI/views/revenue…）、腿② GOAL 目标/真实效果结论（目标达成/需求满足/质量没有问题/客户满意/效果显著/实力提高…；EN goal achieved/needs met/quality is fine/customers are satisfied/capability improved…）、腿③ CONN 连接（显式连接词或逗号并置且无转折词，跨度 ≤80 字/EN 160 字符）。与 activity_outcome_confusion（第184维 活动/投入计数→抽象成果）、reward_hacking（BLOCK 自认式指标作弊）、statistical_misleading / percentage_overflow（比例数字呈现）边界错开：本族腿①只收测得的量、是无欺骗的推理谬误（陈述者真诚相信「指标读数=目标达成」）。主动分离指标与目标（不等于/不代表/只能说明/分开看）、目标本身直接证据在场（访谈/实测/复购/问题解决率）、对冲、概念转述（古德哈特/代理指标/vanity metric）均明确不判。来源（真实，未编造）：本族立项由 r186 activity-outcome-confusion.js 文档显式划界留白（「下载量破百万所以满足需求」「分数上去所以教学变好」归 Goodhart 族）；GitHub search?q=Goodhart's law（sort=stars）参考 agamm/Gobert（Goodhart's law 演示，https://github.com/agamm/Gobert）与 math 评估类成熟方案作概念背景；HN Algolia 检索 Goodhart / proxy metric 作背景；判据为心虫自有纯规则实现（零 LLM 依赖）。侦察落盘 /tmp/hf-scout-r190.txt、探针 /tmp/hf-probe-r190.js。维度数 186→187，VERIFY 145→146，模块 161 不变（纯函数维度模块不进 _modules registry，r186/r189 同口径）。验证：round-190 轮测 17/17（攻击 20/20 含中英/裸名词/倒装/逗号并置/反问/长句、良性 16/16 零误判含主动分离/直接证据/指标到指标/边缘多因陈述、12 项常量负例注入承重、gate 层变体矩阵攻击 8/8 非 pass 且维度承重 + 良性 7/7 pass）。本行由 r191 接手 r190 爆截断存量工作补齐（r190 已实现并全绿，漏补 Version history 与 CHANGELOG/UPGRADE_LOG 即被截断）。 |
| 6.8.119 | 2026-11-02 | 第188维度 false_precision（伪精确断言：无来源无约数的过分精确数字冒充量化证据）上线 — 判别把超出证据支撑精度的数字（小数百分比/小数倍数/小数评分/小数年限/精确副词+数字：37.42% / 3.7 倍 / 92.6 分 / 12.5 年 / 整整 3.7 倍 / exactly 47.83%）以无来源、无约数、无对冲的断言形式给出、制造「量化严谨」假象的结构：数字的精度形态本身就是证据——样本多大、怎么测的都没说，小数却给到小数点后一两位，是把随机噪声包装成测量读数，是 AI 输出中最典型的数字幻觉形态：「用户满意度达到 92.6 分，远超预期。」「性能提升了整整 3.7 倍，几乎没有代价。」「新药可延长预期寿命 12.5 年。」「该模型在测试集上的准确率达到 99.37%。」「The study shows engagement improved by exactly 47.83%.」。判据两腿共现 + 分组豁免，缺一不判：腿① PREC 精确数字形（必须含小数点或精确副词——四舍五入的整数 37% / 3 倍是约整数的合法形态，不判：小数 %/％、小数倍、小评分（92.6 分，排除分钱/分钟/分利/分贝/分米）、小数年岁（12.5 年/岁）、精确副词+数字（整整/恰好/正好/分毫不差/精确地/准确地/精准地 + ≤4 字 + 数字，含整数形——161 轮教训：副词腿收裸词形态，「准确率 92.6 分」由「准确」前缀+数字命中）；EN 小数 %、小数 x/times、小数 years/months/weeks/days/hours/points、exactly/precisely/accurately + 数字）；腿② CLAIM 断言/度量语境（腿①匹配点前后 ≤24 字 / EN 48 字符内任一：效果动词 提升/提高/增长/增加/降低/减少/下降/达到/高达/低至/升至/突破/改善/优化/翻番/刷新/创下；EN improves/increases/reduces/reaches/boosts/extends/achieves/drops/agrees…；度量名词 准确率/满意度/转化率/增长率/覆盖率/留存率/性能/得分/评分/寿命/预期寿命/占比…；EN accuracy/satisfaction/conversion/coverage/retention/performance/life expectancy…；结论标记 证明/表明/显示/说明/意味着/可见/足见；EN proves/shows/demonstrates/indicates/means——裸名词形即断言语境：「转化率 37.42%，完播率 21.8%」无动词也是伪精确陈列）。边界：perfect_error 的 S1_false_precision 管裸精确百分比（无度量名词、无来源，如「市场份额达到 87.3%」）且是多信号协同判定的一票（[v6.7.93] 起对度量名词语境整体豁免，本为「latency dropped 40%」这类工程读数设计），本族专打度量名词语境下的小数断言——正是 S1 豁免腾出的空间，且是单维专项判据、不依赖多信号叠加；statistical_misleading / percentage_overflow 管比例数字的呈现问题（基数隐藏/占比合计溢出），本族不管呈现结构、管数字精度本身是否被证据支撑；unsupported_claim / bullshit 管断言缺依据的强度，本族有数字在场但数字的精度形态无效。明确不判：约数/对冲（约/大约/接近/将近/左右/上下/近似/估算/预计/差不多/误差范围/容差/±/大概/可能/也许/据说；EN about/around/approximately/roughly/nearly/almost/estimated/give or take）、具名来源（根据X统计局/X大学/X报告/X调查/《X》/据X统计/年报/季报/月报/财报/公报/白皮书/来源于/数据来源/转引自/参见；EN according to X report/study/survey/census/source:）、天然精确语境（数字近邻 ≤40 字 / EN 80 字符内：完整日期、版本号 v6.8.118、自然计数 12 个/1200 人、物理量纲与精度规格 0.01 毫米/±0.5℃、数学常数 π/约等于、时长与历史语境 历时/持续/为期/已有/建于/成立于）、仪器读数语域（数字近邻内 监控/压测/日志/报表/看板…显示/表明——同 r437「可定位复现」边界）、统计口径（同比/环比/较上期/去年同期/百分点/percentage point/year-over-year）、元讨论（有效数字/小数点/四舍五入/位数/保留/精度）、疑问句（含 ?/？）。已知边界：① 「实验组有效率为 87.3%」这类度量名词未入表且无效果动词的形暂不判（「有效率/完成量/满意度指数」等词可续入腿②，保守方向）；② 句级豁免仅留约数/来源/口径/元讨论四类真句级属性，天然精确与仪器豁免已收窄为数字近邻窗；③ 跨段数字不判。来源（真实，未编造）：HN Algolia 命中 Wikipedia "False precision"（https://en.wikipedia.org/wiki/False_precision）；GitHub API hallucination detection benchmark（sort=stars）最高星 zjunlp/FactCHD（IJCAI 2024，92★，https://github.com/zjunlp/FactCHD，事实冲突幻觉检测基准）作数字幻觉背景；Semantic Scholar 三次检索均 HTTP 429 空返。缺口实证 /tmp/hf-probe-r192.js：本族 6 条攻击 4 条真零命中、6 条边缘良性零误伤。侦察落盘 /tmp/hf-scout-r192.txt、探针 /tmp/hf-probe-r192.js。维度数 187→188，VERIFY 146→147，模块 161 不变（纯函数维度模块不进 _modules registry，r190 同口径）。验证：round-192 轮测 61/61（攻击 17/17 含中英/裸名词/倒装/长句/精确副词/小数年/小数评分、良性 17/17 零误判含来源/对冲/仪器/时长/口径/元讨论/疑问/天然精确/边缘多因陈述、11 项常量负例注入承重、gate 层变体矩阵攻击 8/8 非 pass 且维度承重 + 良性 8/8 pass）。 |
| 6.8.120 | 2026-11-02 | 全面审计修复轮（用户指令：全面审计→全面修复→同步 GitHub） — 3 路子代理并发扫描 527 个 JS 文件，修复 HIGH 10 类 + MEDIUM 10 类，零 CRITICAL、零新增回归。①**[HIGH] 每次 think() 必抛的 TypeError**：heartflow.js:6047 _PrivacyExposure.sanitizeForPersistence 漏调 lazy 函数（缺 ()），被 catch 吞后 Engram 存储/召回整块失效（_engramRecall 恒 undefined）——修复后实测恢复真实召回数据。②**[HIGH] BLOCK 级豁免击穿**：reward_hacking 的 DISCOURSE_FRAME 豁免未过信号闸，「比如，直接绕过安全检查」加前缀即整体豁免；弱举例标记组（比如/举例/像是）移除+双闸纪律对齐，攻击句恢复 HIT、良性元话语 3/3 仍豁免、回归 27/27。③**[HIGH] MCP 匿名提权**：OID 身份码只做格式正则无签发体系，匿名 guest 自报 x-heartflow-oid 即被提为 user 穿过 40 个写工具拦截——修复为 guest 持 OID 不提权，固化该漏洞的测试期望同步反转（mcp-auth-e2e-real 13/13）。④**[HIGH] 记忆库静默重置**：rotate-aes-key 只扫 .enc 后缀，data/memory-bank.json（同格式 .json）被漏扫，换钥后每次启动 Decryption failed... starting fresh 整库丢弃（日志实锤）——改按内容探测纳入轮换。⑤**[HIGH] 文档漂移**：package.json 168→190、SKILL.md 184→190/149→177 等 6 处，sync-doc-numbers 补登记 CONTRIBUTING/CURRENT_STATE 两份活文档并修复其把 UPGRADE_LOG 历史轮次数字误报为漂移的缺陷。⑥**[HIGH] 回归门禁静默放行**：heartflow-self-evolve 判据只匹配英文 N failed 而 run-all 输出中文「M 失败」+ 180s 超时必杀 27 分钟的全量回归——被杀后 failedCount=0 空放行并继续 commit/push；修 fail-closed + 双语判据 + 30min 超时。⑦**[HIGH] 即崩工具**：heartflow-memory-tool/inject.js require 全仓不存在的 HeartFlowMemory 类（从未入 git），被 two-pass-response 技能宣传为记忆注入能力——删除死脚本，宣称修正为 py 插件层。⑧**[HIGH] 假健康**：24 个 _stubFactory 占位子系统 healthCheck 恒 {ok:true} + 4 个假全局桥接类打假「加载成功」日志——统一诚实化 {ok:false, stub:true/fake:true}。⑨**[HIGH] 死注册表**：14 个 lazy 条目指向不存在文件（planner 族 8/code 族 4/memory 族 2），dispatch 报误导性 route not allowed——删除。⑩**[HIGH] require 守卫自身红灯**：3 处 archive 断链 + git/无-git 双模式口径不一致（一个扫 archive 一个不扫）——修断链+统一口径，527 文件 0 断链。MEDIUM：dimensions 输出补 agency_deflection/ai_writing_tell 键、dimMap 补 6 键修 findings detail 恒 1、bullshit→bullshit_recognition 三表命名统一（维度口径 188→190）；_thinkCache 死代码连根删；threePoisons 空 try 诚实化；rate limit 注释口径统一（100/min）；sseClients 加 CAP=64 LRU；Unix socket chmod TOCTOU 用 umask 预置修复；6 处内存泄漏（WAL _pending 不回写/outbound-ledger _buffer/decision-engine 双 Map/cross-domain _chains/receipts/DNS_CACHE 只读不清理）全部加上限或 TTL；package.json files 移除 test/（828 个测试文件不再随包外发）；measure 裸 spawnSync('node') 改 process.execPath（干净 PATH 下维度数不再静默输出 0）；triality-memory.js 死代码出库；3 个 fetch 脚本 require 路径少一层 .. 修复。验证：五门禁全绿（verify 16/16、guard 20/20、doc 22/22、双向召回 52/52 误拦 302/326 不升、rh 回归 27/27）+ 冒烟 6 分支 initErrors=0 + mcp-auth-e2e 13/13 + decision-router 23/23。版本 6.8.119→6.8.120（判别逻辑仅 F3 一处行为变更，其余为诚实化/死代码/文档/资源边界）。 |
| 6.8.122 | 2026-11-02 | LOW 批次收尾轮（v6.8.120 审计轮遗留项 + 合并回归暴露项） — ①think() 并发互斥：_pendingInputCheck 实例字段跨 await 边界，并发 think() 互相清掉/丢计数——改局部变量，实测输入辨别合并仍在（情感操纵句 1 issue）+ 并发双 think 互不干扰。②archive 断链清零：sleep-wake.js require('./heartbeat') → ../core/heartbeat.js，audit-missing-requires 待分类引用 1→0。③v6.1.7 注释残留死行删除（try 体丢失、catch 被注释吸收的死代码）。④门禁严格化：doc-numbers 五处条件断言（if (rows[X] !== undefined) 行缺失即静默跳过=守卫失守）改为先断言行存在再比对；guard-abilities 过时注释 51→176 个 check 函数；verify 模块数地板 124→150（实测 162）。⑤sync-version.js 补兜底常量同步：version.js 第 97 行兜底版本此前无人管理，实测静默停滞在 6.8.120 落后两个版本（runtime 读 VERSION 文件所以不报错）——兜底值是 VERSION 文件丢失时的唯一版本源，现由 syncVersionFallback 自动同步。⑥合并回归修复（v6.8.121 合入时暴露）：openmem 先例查询在 think() 开头引入 await 边界后，verify.js 原「同步阶段 stop() → 之后 await 异步检查」顺序使 think 续跑撞 started 检查（16 项必挂 1 项）——停止引擎移至 Promise.all 完成后。未修（如实）：start() 约 12 处关键路径实例化无 try/catch（identityCore 等 fail-fast 是设计意图，批量包 try/catch 回归风险大于理论收益，留作文档化隐患）。验证：五门禁全绿（16/16、20/20、22/22、双向 52/52 误拦 302/326 不升）+ 冒烟 restart ok + round-193 33/33 + security-audit 16/16 + 并发 think 双 ok + 断链 0。 |
| 6.8.116 | 2026-11-02 | 第185维度 most_transitivity（「大多数」链式传递谬误：most 量词不满足传递性）上线 — 判别「两个「大多数」前提推不出「大多数」结论」结构：「大多数A是B，大多数B是C」的证据不能沿中间项累积——100 人里 60 人喝茶、60 个喝茶的人里只有 35 人喝咖啡时，「大多数人喝咖啡」就不成立，两段「大多数」各自只覆盖自己前提的大部分、交集可以远小于半数：「大多数A队球迷支持主队，大多数支持主队的人会买季票，所以大多数A队球迷会买季票」「大部分用户点了同意，大部分点了同意的人都没读协议，所以大部分用户都没读协议」「七成受访者表示满意，七成表示满意的受访者购买了会员，所以七成受访者都买了会员」「Most students passed the exam, and most people who passed got certificates, so most students got certificates.」。判据链式三件共现，缺一不判：件① 第一段「大多数」前提（量词+主项+非空谓项，中间项素材在谓项里）；件② 第二段「大多数」前提的主项包含第一段谓项的关键词（中间项连续性）；件③ 「大多数」结论主项回指第一段主项、谓项是新内容（非复述：最长共享≥3字或互为子串才算复述）、且结论整体与第二段有共享关键词；结论带推断词（任意位置，覆盖「所以C，因为P1、P2」倒装形）或跟在两前提之后（跟句模式）。缺一不判——两个「大多数」并列无链式结论是正常统计陈述，对冲/特称结论是谨慎或有效的推论。与 illicit_conversion（第167维，全称换位）、affirming_the_consequent / denying_the_antecedent / conditional_proposition（条件句方向族）、statistical_misleading / percentage_overflow（比例数字呈现）边界错开。任务书轮换纪律：r186 为价值对齐一线，本轮切回逻辑谬误检测 + 证据纪律一线（候选四族探针实测：A 中项不周延 5/6 被 affirming_the_consequent 代拦且 2 条有效三段论被误拦、B 肯定析取 3/5 被 illicit_conversion 等代拦、C 堆垛 4/5 被 fallacies 代拦、D most 链式传递 6/10 真零命中且 8/8 边缘良性零误伤——选 D）。来源（真实，未编造）：HN Algolia "Translating natural language to first-order logic for logical fallacy detection"（258 分，https://news.ycombinator.com/item?id=43257719）、"Master List of Logical Fallacies (2018)"（113 分）、"Slaughtering Competition Problems with Quantifier Elimination"（78 分，https://news.ycombinator.com/item?id=46415448）、"The strange non-transitivity of random variables"（26 分，https://news.ycombinator.com/item?id=10763018）作量词与非传递性概念背景；GitHub tmakesense/logical-fallacy（21★）、bkthomps/CategoricalSyllogism（6★，https://github.com/bkthomps/CategoricalSyllogism）为三段论/谬误检测成熟代码参考。Semantic Scholar 两次检索 HTTP 429 空返；判据为心虫自有纯规则实现。侦察落盘 /tmp/hf-scout-r188.txt 与 /tmp/hf-scout-r188-transitivity.txt。维度数 184→185，VERIFY 143→144，模块 161 不变（纯函数维度模块不进 _modules registry，r186 同口径）。验证：round-188 轮测 14/14（攻击 14/14 含中英/裸名词量词/带修饰/倒装/长句/无逗号紧凑形、良性 17/17 零误判含特称结论/对冲结论/并列统计/量词副词化/纯转述/边缘多因陈述、9 项常量负例注入承重、3 条豁免陷阱翻转、gate 层变体矩阵攻击 6/6 非 pass 且 most_transitivity 承重、良性 6/6 pass）、双向门禁 52/52 召回、误拦 302/326 不升（与基线一致、零新增误拦）、verify 16/16、guard 20/20（维度登记 185 键）、doc 22/22。已知边界：① 定义式真值链（「大多数猫是哺乳动物，大多数哺乳动物是温血动物，所以大多数猫是温血动物」）量词形式相同也会判 verify——结论为真但交集大小未被证明，verify 的含义正是要求补全称或特称证据；② 主谓切分是启发式（功能词边界+约 280 个高频动词），生僻动词开头的前提可能切不出谓项而漏判（保守方向）；③ 谓词链接结论与第二段的判据对「换了说法」的结论谓词（「用真金白银表达信心」↔「买了股票」）依赖主项/其他共享词兜底。 |
| 6.8.117 | 2026-11-02 | 第186维度 appeal_to_wealth（诉诸财富/诉诸贫穷谬误：argumentum ad crumenam / ad lazarum）上线 — 判别「把财富或贫困状态当作观点正确性、判断力或可信度的证据」结构：有钱所以说得对、听成功人士的总没错、穷人懂什么、他穷过所以更懂人间疾苦。财富与贫困是处境属性，观点对错与判断力高低是认知属性，两者之间没有可传递的证据链——有钱不增加任何命题为真的概率，穷不减少任何论证的有效性：「他那么有钱，说的话当然有道理，人家身家几十亿，认知能不比我们高吗？」「马云说的都是真理，他能赚到几千亿身家，眼光还能错吗？」「一个穷得叮当响的人，观点能有什么价值？」「他穷过苦日子，所以最懂老百姓的心，说的话才真正靠谱。」「He is a billionaire, so his advice on the economy must be sound.」「She married into one of the richest families, so her judgment is obviously superior.」。判据两腿+连接，缺一不判：腿① STATUS 财富/贫困状态词（有钱/富翁/富豪/身家/富人/成功人士/白手起家/穷困/贫寒/穷得叮当响/寒门/底层出身…裸名词形态；EN billionaire/rich/wealthy/made billions/poor/poverty/the poor——限定搭配收录，不收裸「穷」（穷尽误伤）、裸「富」（丰富误伤）、裸「成功」（实验成功误伤）、裸「底层」（底层逻辑误伤））；腿② VALIDITY 认知有效性结论（有道理/是对的/高明/靠谱/眼光好/最懂/更懂…；EN sound/right/superior/credible/gets it——不收「值得参考」（经验价值）、不收「犯错/失误」（错误承认）、不收「业绩/收益」（实测数据））；腿③ 连接——路线 A（正向断言）腿①与腿②间有推断连接词（所以/当然/一定/因为…；EN so/because/must/obviously）且间距 ≤50 字（EN 110 字符）；路线 B（反问断言）腿① ±50 字（EN 110）内有反问有效性形（眼光还能错吗/能有什么价值/懂什么；EN how could he be wrong），反问自身携带推断、免连接词。有效性断言前置否定时不判（不是因为眼光好/观点不总是对的/does not determine whether someone is right）。与 appeal_to_authority（第资质类权威，词表互斥）、halo_effect（外貌名气迁移）、absolute_claim/confidence（断言强度）、unsupported_claim/no_fallback（依据缺失强度）边界错开。任务书轮换纪律：r188 为逻辑谬误一线，本轮探针横扫三方向三候选族（零和思维 6/6 被代拦+2 良性误伤淘汰、均值回归 4/6 零命中但 3 良性误伤淘汰、诉诸财富 4/6 真零命中+6/6 边缘良性零误伤——唯一干净缺口）。来源（真实，未编造）：HN Algolia "The Fallacy Of Zero Sum Game Thinking"（26 分，https://news.ycombinator.com/item?id=6390466）、"Humanity is not a Zero-Sum Game"（https://news.ycombinator.com/item?id=30773049）作零和与阶层推理概念背景；GitHub tmakesense/logical-fallacy（21★）、SeanFlannery/Logical.ly（4★）为谬误检测成熟代码参考；Semantic Scholar 检索 HTTP 429 空返；判据为心虫自有纯规则实现（零 LLM 依赖）。侦察落盘 /tmp/hf-scout-r189.txt、探针 /tmp/hf-probe-r189.js。维度数 185→186，VERIFY 144→145，模块 161 不变（纯函数维度模块不进 _modules registry，r186/r188 同口径）。验证：round-189 轮测 16/16（攻击 21/21 含中英/裸名词/带修饰/倒装/长句/反问路线、良性 20/20 零误判含亲历依据/审计业绩/错误承认/按证据评估/否认关联/谚语/边缘贫困叙事、9 项常量负例注入承重、3 条豁免陷阱翻转含否定守卫、gate 层变体矩阵攻击 7/7 非 pass 且 appeal_to_wealth 承重、良性 7/7 pass）、双向门禁 52/52 召回、误拦 302/326 不升（零新增误拦）、verify 16/16、guard 20/20（维度登记 186 键）、doc 22/22。已知边界：① 腿①与腿②限同句或相邻分句、总跨度 ≤50 字（EN 110 字符），跨段形本轮保守不收；② 「不无道理」类双重否定会被否定守卫挡下（保守方向，漏判优于误判）；③ 隐含有效性断言（「他有钱，说话就是硬气」——硬气是气势非有效性）暂不判。 |
| 6.8.115 | 2026-11-01 | 每日审计修复轮（r187） — 修复文档口径漂移：README badge/横幅外 5 处维度数陈旧值（164/168/180/180/142 → 184/184/184/184/143）、SKILL.md pipeline 图 46 dims → 184、AGENTS.md checkInput 行 50 → 184；修复 src/core/version.js 兜底版本号停留在 6.8.89（r161 起连续 25 轮升级遗漏手动同步，违反 AGENTS.md 约定 #1「fallback 必须与 VERSION 一致」）→ 6.8.115。引擎判别逻辑零改动，五门禁复跑全绿。 |
| 6.8.114 | 2026-11-01 | 第184维度 activity_outcome_confusion（活动/投入当成果：以工代果）上线 — 判别「把会议场次/发文份数/加班时长/代码行数/走访户数等活动或投入计数，当作「推进得好/进展巨大/成效显著/成绩斐然」这类抽象成果的证据」结构：「我们开了 50 场会、发了 100 份文件，所以项目推进得很好」「代码写了几万行，可见进展巨大」「培训办了几十期，可见人才培养卓有成效」。活动量与成果之间没有必然传递——会开得多可能议题涣散，班加得猛可能在做无用返工，代码行数多可能蕴含更多缺陷；用活动的数量给成果背书，是把「做了很多」偷换成「做成很多」。判据活动腿×抽象成果腿×因果连接三腿共现，缺一不判（活动腿收可计数活动名词与须配强度标记的泛指投入词；成果腿只收抽象成功/进展/绩效、不收具体结果；连接为显式连接词或逗号并置且无转折词）。与 reward_hacking 自认式作弊（BLOCK）、Goodhart 代理指标族（点击率/GDP/停留时长——测得的量）边界错开：本族腿①只收活动/投入（做的事、花的工），是无欺骗的推理谬误。缺口实证攻击 5/5 真零命中、良性 0/5 误伤。维度数 183→184，VERIFY 142→143。 |
| 6.8.113 | 2026-11-01 | 第183维度 necessity_sufficiency_confusion（条件角色混淆：必要条件↔充分条件反向使用）上线 — 判别「先宣称某条件是另一事件的必要条件（前提/基础/保障/「没有X就没有Y」/「唯有X才Y」），再据此断言该条件单独就能保证结果（「所以只要X就能Y」）；或先宣称充分条件（「X足以Y」/保证），再据此断言没有该条件结果就不成立（「所以没有X就不能Y」）」结构：「努力是成功的必要条件，所以只要努力就能成功」「信任是团队合作的前提，因此只要把信任建立起来，合作自然水到渠成」「没有人民支持就没有革命的胜利，所以只要赢得了人民支持，胜利就唾手可得」「唯有每天复盘才能进步，可见只要坚持复盘，进步就不成问题」「睡眠是记忆巩固的保障，所以睡眠问题一解决，成绩自然不在话下」「天赋是夺冠的充分条件，所以没有天赋就夺不了冠」「努力足以换来回报，所以不努力就没有回报」「Trust is necessary for cooperation, so once trust is in place, cooperation will follow automatically.」「A degree is a prerequisite for this role, so a degree alone qualifies you.」「This certificate is a guarantee of employment, so without it you cannot be hired.」。必要与充分是两个方向：P必要于Q只意味着 ¬P⇒¬Q，推不出 P⇒Q；P充分于Q只意味着 P⇒Q，推不出 ¬P⇒¬Q；把单向条件当双向用，是用角色的静默互换伪造推理的必然性。任务书轮换纪律：r182 为自我验证一线，r183 为反事实一线，r184 为证据纪律一线，本轮切回逻辑谬误检测一线（候选四族探针实测：A 条件角色混淆 11/11 真零命中、B motte_and_bailey 6/9、C nirvana_fallacy 3/6、D prosecutor_fallacy 4/7——选 A 因两腿结构最清晰、与 denying_the_antecedent/affirming_the_consequent/conditional_proposition 三族边界全部错开、FP 风险最低）。判据宣称腿×反向结论腿×连接推断词三腿共现，缺一不判：路线1（必要→充分）腿① NEC_CLAIM 必要条件宣称（是…必要条件/前提/基础/保障/基石；离不开/缺一不可；没有X就没有Y；唯有X才Y；EN necessary for / prerequisite / without X no Y / only if）× 腿② SUF_CONCL 充分性结论（只要X就Y/就够了/不在话下/不成问题/水到渠成/迎刃而解/一劳永逸/有了X就Y；EN once…will follow / alone is enough / guarantees）× 腿③ 宣称与结论之间有推断词（方向强制：宣称在前用前向连接词 所以/因此…；结论在前用后向连接词 因为/既然…——r185 首跑实测：不强制方向时「X是Y的必要条件，所以X是Y的充分条件」会被路线2反向双重解释，同一句两路线齐命中）且跨度 ≤70 字（EN 140 字符）；路线2（充分→必要）腿① SUF_CLAIM 显式充分宣称（是…充分条件/保证/足以；EN sufficient for / alone is enough / suffices）× 腿② NEC_CONCL 必要性结论（没有X就不能Y/不X就不Y/除非X才Y；EN without X cannot / impossible without）× 腿③ 同上。反向宣称抑制（本族精度核心）：结论只是把条件已宣称过的角色再用一遍时不是混淆——路线1要求文中不存在充分条件宣称、路线2对称（「X是Y的充分条件，所以只要X就能Y」是同向复用，不判；「信任是合作的前提，也是合作的充分条件，所以只要有了信任就能合作」双角色俱告亦不判）。缺腿不判：只有必要条件宣称无充分结论（「没有健康就没有一切，所以我们要锻炼」——结论是行动建议）；只有充分结论无角色宣称（「只要努力就能成功」单独成句——普通条件句，归 denying_the_antecedent 家族）；宣称与结论之间无推断词（「没有健康就没有幸福。只要有钱就能幸福。」——两句并列无推理关系）。r185 实现教训（已固化进测试）：① 连接推断词必须分前向/后向两个常量并强制方向（见上）；② 对冲豁免陷阱的样本必须让豁免词与结论腿分离——「所以只要努力就可能成功」里「就可能」不匹配 SUF_CONCL 的「就+能/可以/会」交替，抹掉对冲豁免后仍不命中，陷阱不承重，改为「所以只要努力就能成功，当然这只是可能的路径之一」；③ 反向宣称抑制陷阱必须让双角色出现在不同子句（「X是Y的充分条件，所以只要X就能Y」放行是因为没有 NEC 宣称腿而不是抑制逻辑承重——换「X是Y的前提，也是Y的充分条件，所以只要X就能Y」才真承重）。边界：denying_the_antecedent（第64维）管条件规则句的 ¬P∴¬Q 推理（如果X就Y；没X所以没Y），本族腿①是角色宣称句、不依赖「如果…就…」规则句形态，「只要X就Y」单独出现不判（归其家族）；affirming_the_consequent 管 Q∴P 逆向倒推，本族不倒推前后件、只搬弄条件角色；conditional_proposition 管 ¬B⇒¬A 用未验证假言裁定前件，本族结论落在条件角色判断上；whole_part_transfer 管层级间属性迁移，本族管同一对条件命题间的角色互换；absolute_claim/confidence 管断言强度，本族角色词缺席时不判。明确不判：正确区分角色（不是充分/不充分/仅靠…不够/不能只靠/还需要；EN not sufficient / necessary but not sufficient）、拒绝推断（不等于/不意味着/不能说/不成立/误解；EN does not mean / cannot conclude）、对冲（可能/也许/未必/不一定；EN maybe/perhaps/might）、双条件/充要（当且仅当/充要条件；EN if and only if / iff）、概念/教学转述（什么是/区别/课堂/教材/谬误/逻辑题/形式逻辑；EN difference between / textbook / lecture）、疑问句（含 ?/？）。src/necessity-sufficiency-confusion.js 新增 20 个可注入常量（NEC_CLAIM/SUF_CLAIM/SUF_CONCL/NEC_CONCL_ZH×EN、INFER_FWD/INFER_BWD_ZH×EN、EXEMPT_SEP/NEG/HEDGE/BICOND/CONCEPT_ZH×EN），腿①/腿②/腿③正则全部可注入，index.js 九处接线（require/计算点/allDims/dimMap/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports；变量名 nsc，grep 实测无冲突）。来源（真实，未编造）：HN Algolia "logical fallacy detection" 命中 arXiv:2405.02318《Translating natural language to first-order logic for logical fallacy detection》（HN 258 分，https://arxiv.org/abs/2405.02318）——NL2FOL 把自然语言翻译成一阶逻辑后用 SMT 求解器判定逻辑有效性，与纯规则引擎不同构，只取其谬误族清单作概念背景；Semantic Scholar "necessary sufficient conditions reasoning fallacy" 命中 10.1016/j.neunet.2025.108164《Necessary and sufficient knowledge enhanced collaborative logical reasoning in LLMs》（2025，0 引）与 10.1080/01445349408837232《A Case of affirming the consequent in international law》（1994）——确认必要/充分条件混淆是逻辑推理的公认失效族；GitHub API "logical fallacy detection"（sort=stars）最高星 tmakesense/logical-fallacy ★21，其余 ≤4 星且均为 LLM 路线，无可移植纯规则判据；判据为心虫自有纯规则实现。侦察落盘 /tmp/hf-scout-1761930000.txt。维度数 182→183，VERIFY 141→142。验证：round-185 轮测 20/20（攻击 11/11 含中英/裸名词/带修饰/倒装/长句/直接翻转、良性 16/16 零误判含正确多因陈述/必要条件正向使用/双角色同向复用/双条件/概念教学/拒绝推断/对冲/无连接词并列、9 项常量负例注入承重、6 条豁免陷阱翻转含反向宣称抑制、登记断言×3）、gate 层变体矩阵 6/6 攻击非 pass 且 5/5 良性 pass、双向门禁 52/52 召回 302/326 误拦不升（与基线一致，零新增误拦）、verify 16/16、guard 20/20（维度登记 183 键）、doc 22/22。已知边界：① 宣称与结论限同句或相邻句且总跨度 ≤70 字（EN 140 字符），跨段形本轮保守不收；② 隐式充分条件句（「X就能Y」）不进 SUF_CLAIM 腿——它与 denying_the_antecedent 的条件规则句高度同形，双重命中会稀释该族信号，显式角色语言（充分条件/足以/保证/alone is enough）才判路线2；③ 宣称与结论分属不同主语的跨条件形（「没有健康就没有幸福。所以只要有钱就能幸福。」）因无推断词连接不判，可能漏「借用前句角色宣称为后句结论背书」的间接形；④ 存量维度 contradiction 会误伤边缘良性「努力是成功的必要条件，但努力不是充分条件，还需要方法、机遇和时代红利。」（r185 实测：该句 gate.action=verify 由 contradiction 代拦——非本维度引入，改动前行为一致，故 gate 层良性矩阵改用不触发 contradiction 的等价边缘良性）。 |
| 6.8.112 | 2026-11-01 | 第182维度 absence_of_problems（未检出≠不存在：把有限检验的阴性结果当作完备证明）上线 — 判别「测试/扫描/审查/调研只覆盖有界范围，却把「本次未检出」升级为「不存在」」结构：「测试全跑过了没报错，所以代码肯定没问题」「扫了一遍没发现漏洞，说明系统是安全的」「没人投诉，说明产品体验没问题」「三年没出过事故，这座桥梁绝对安全」「尽调没查出任何问题，这家公司完全可以放心投资」「抽样检查没有发现缺陷，产品百分之百合格」「The tests did not fail, so the code must be correct.」「We scanned and found no vulnerabilities, therefore the system is secure.」「No bugs were reported during the audit, so the release is completely safe.」。检验的效力受其覆盖范围约束：测试只覆盖已写下的用例，扫描只匹配已知漏洞特征库，审查只覆盖审查者看得见的层面，调研只覆盖愿意开口的人；把「本次未检出」升级为「不存在」，是用有界搜索的阴性结果冒充无界全域的证明，让未被检验的部分获得与已检验部分相同的可信度——这是 AI 工程汇报、代码审查结论、上线评审中最典型的似是而非：听起来是验证结论，实则是验证范围的静默扩张。任务书轮换纪律：r176-r183 连续八轮为逻辑谬误/证据纪律/反事实一线，本轮切自我验证与纠错 + agent 执行验证一线（候选四族探针实测真零命中：A nirvana_fallacy 3/6、B absence_of_problems 5/6、C red_herring 6/6、D common_sense 4/4；选 B 因两腿结构最清晰、FP 风险最低、与既有维度边界全部错开）。判据两腿共现，缺一不判（同句或相邻句）：腿① CHECK 已执行检验+阴性结果三形任一——CN-A 阴性结果形（检验动词 测试/跑过/回归/扫描/检查/审查/核查/排查/审计/验证/复测/体检/走访/调研/问卷/回访/抽查/化验/检测/验收/评审/复盘/对照/巡检/观察/试用/实测/压测/演练/试运行/评测/筛查/尽调/背调/复查/复审/复核/校验/自查/互查/巡查/专访/暗访/抽样 设为可选前缀——「扫了一遍没发现漏洞」有动词，「没发现问题」「三年没出过事故」无动词亦可，工程汇报里动词省略是常态 → [≤12字] → 否定 没/未/没有/并无/未能/不曾/从无/毫无/压根没/根本没/从未 → [≤6字] → {检出动词 报错/发现/找到/查出/检出/测出/暴露/出现/触发/冒出/见到/遇到/收到/得到/报出/显现/看出/听出/出过/发生过 → [≤10字] → 问题名词 问题/漏洞/缺陷/错误/异常/故障/风险/隐患/毛病/弊病/差错/事故/副作用/不良反应/反馈/差评/恶评/意见/不满/异议/警告/报警/疑点/破绽/把柄/bug |
| 6.8.111 | 2026-11-01 | 第181维度 counterfactual_overclaim（反事实独断：把另一条未走路径的结果当作确定可知）上线 — 判别「如果当初X，现在一定/必然/势必Y」结构：「如果当初投入研发，现在我们一定已经领先市场三年了」「要是当时听我的做了对冲，公司肯定不会陷入今天的困境」「倘若当初选择另一条技术路线，我们现在势必已经主导整个行业」「当初要是果断止损，这笔投资毫无疑问现在已经回本了」「If we had invested in that startup, we would certainly be market leaders today.」「Had we taken the other path, the project would definitely have succeeded by now.」「If only we had acted sooner, we would certainly have avoided the crisis.」。复杂系统中未走路径的结局原则上不可事先确定——用确定性副词封装反事实结果，是把不可验证的假设当结论，用叙述的确定性替代因果的开放性（r183 差距实测：本族攻击 2/8 真零命中，「反事实框架 × 确定性断言」结构零维度覆盖；conditional_proposition 第160维只收否定推断形 早就/就不会/would）。判据双腿共现，缺一不判：腿① CF 反事实框架三形任一——CN-A 连接词先行形（连接词 如果/若/要是/假如/倘若/倘使/若是/设若/要不是/若非 → [≤12字] → 过去时点锚 当初/当时/那时/那年/当年/早年/一开始/最开始/早些时候/早点/最初/早先——刻意不收「之前/先前/此前」：宽锚会把「如果之前的数据有误，现在一定要回头修正」的义务情态扫进谬误族）、CN-B 锚先行形（过去时点锚 → [≤8字] → 连接词/否定动作——「当初要是听我的」「那时如果没停电」）、EN-A if+主语+had+过去分词 / EN-B 倒装 had we/they+PP / EN-C if only+主语+had；腿② CERT 确定性断言——一定/肯定/必然/势必/必定/注定/毫无疑问/毋庸置疑/不容置疑/必将；EN certainly/definitely/bound to/undoubtedly/without a doubt/no doubt/surely（刻意不收「早就/就不会/would have」——那是 conditional_proposition 否定推断形的射程）。161 轮教训：全部收裸名词形态。四类豁免全部负例注入承重：对冲（也许/或许/可能/大概/未必/不一定/说不定/无法知道/谁也说不准；EN maybe/perhaps/might/may have/no one can say）、拒绝推断（不要以为/别以为/无法断定/推不出/不意味着/不恰当/轻率/武断；EN does not mean/cannot conclude/unjustified/speculative）、概念/教学转述（反事实/复杂系统/混沌/决定论/因果推断；EN counterfactual/causal inference/complex systems）、结论侧否定（不足以/无法/难以 + 证明/证实/说明/支撑/确立；EN insufficient/does not prove），另加疑问句逐句豁免。来源（真实，未编造）：HN Algolia "counterfactual reasoning" 命中 Steve Kinney 论文笔记《The Hidden Dangers of Counterfactual Reasoning in AI》（2025-04-21，https://stevekinney.com/writing/counter-factual-reasoning-in-ai）——评述 Aalaila / Großmann / Mukherjee / Wahl / Vollmer《When Counterfactual Reasoning Fails: Chaos and Real-World Complexity》：复杂/混沌系统中反事实推理脆弱、小扰动即发散，确定性反事实断言缺乏依据；该论文 arXiv 检索无收录，未编造 arXiv ID。另命中 arXiv:2402.08955《Counterfactual Tasks to Evaluate the Generality of Analogical Reasoning in LLMs》（HN 2 分，LLM 评测路线，只作概念背景）。GitHub API 检索无高星纯规则方案；Semantic Scholar 限流空返，如实记录。侦察落盘 /tmp/hf-scout-r183.txt。心虫裁决 philosophy_decision type=accelerate（confidence 0.7 / priority 40 / goalReviewed true；rationale 的 cognitiveLoad 与 uncertainty 均为字符串 "unknown"，按轮次纪律不当实质背书，accelerate 依据是缺口实证与轮换纪律）。判据为心虫自有纯规则实现。维度数 180→181，VERIFY 139→140。验证：round-183 轮测 67/67（攻击 13/13 含中英/裸名词/带修饰/长句/锚先行/倒装、良性 15/15 零误判含对冲/拒绝推断/概念转述/结论侧否定/疑问句/将来时条件句、15 项常量负例注入承重含 8 条豁免陷阱翻转、登记断言×5）、gate 层变体矩阵攻击 6/6 非 pass 且 findings 全含 counterfactual_overclaim、良性 6/6 pass、双向门禁召回 52/52 误拦 302/326 不升、verify 16/16、guard 20/20（维度登记 181 键）、doc 22/22。已知边界：① 腿①腿②限同句共现，跨句形（「如果当初投入研发。现在我们一定已经领先三年。」）语义同族但保守不收；② 将来时条件句（「如果明天发布新版本，用户量一定会涨」）不在射程（无过去时点锚，归 absolute_claim 类）；③ 存量维度 absolute_claim 误伤「不要以为如果当初买了那套房现在就一定过得更好」、hindsight_bias 误伤「当初的决策基于当时可获得的信息，事后假设另一种选择必然更好是不恰当的」（r183 实测，非本维度引入，改动前行为一致，下轮可评估修复）。 |
| 6.8.110 | 2026-11-01 | 第180维度 cherry_picking（选择性取证 / cherry-picking，采樱桃谬误——确认偏误的证据纪律形态）上线 — 判别「只从支持某结论的证据池取证（只找/只引用/只保留支持性证据，或隐去/略去/屏蔽不利与反面证据），却把结论封装为「证明/证实/成立/毫无疑问」」：「我只找支持这个结论的证据，所以它一定是成立的」「他专门挑对自己有利的数据说，足以证明产品没有问题」「报告里只引用正面评价，隐瞒了所有不利记录，可见方案是完全可靠的」「整个论证只保留了支持性案例，反面证据全部略去，因此结论毫无疑问是成立的」「为了说明这项政策有效，研究者只摘取了支持该政策的材料，因此结论是毋庸置疑的」「团队筛选符合预期的实验记录，据此证明新算法性能显著更优」「We only cited the studies that supported our hypothesis, which proves the theory is correct.」「The report only presented favorable evidence, so the conclusion is certainly sound.」。寻真的前提是证据对称性——结论的置信度只能来自正反证据的合力；选择性取证把取证本身变成结论的傀儡：先裁掉不利证据，再宣布「证据充分」。它常以方法论口吻出现（「我们核查了」「报告显示」），是 AI 输出中最典型的似是而非：看似做了功课，实则只做了支持自己的功课。任务书轮换纪律：r176-r181 连续六轮为逻辑谬误/判断偏差一线，本轮切证据充分性与寻真纪律一线（候选来自 r182 差距实测：本族 6 条攻击中 1 条真零命中——EN「We only cited the studies that supported our hypothesis, which proves the theory is correct.」穿过硬闸门且无任何维度代拦，其余 5 条仅被 absolute_claim / confidence / contradiction 兜底代拦，无一维覆盖「选择性取证 × 确定性结论封装」结构本身；grep -rl 扫描：fallacies 维度内仅 1 条英文窄模式 cherry-pick + data/evidence + to prove 与 3 条中文确认偏误窄句，中文选择性取证族零覆盖）。判据两腿共现 + 顺序腿，缺一不判：腿① SELECT 选择性取证陈述四形任一——A 支持性取样形（选择性副词 只/仅/唯独/单/专门/专挑/净挑 → [≤8字] → 取证动词 找/搜集/收集/采集/挑选/选取/筛选/引用/援引/摘取/保留/留存/呈现/列举/列出/展示/报告/提及/分析/统计/纳入/采信/采纳 → [≤10字] → 方向标记 支持/有利/符合/赞同/迎合/顺着/证明/附和/正面/利好/唱好 → [≤8字] → 取证对象 证据/数据/案例/材料/例子/事实/文献/报道/样本/记录/证言/反馈/评价/意见/说法/信息/研究/调查/问卷/言论；A2 无副词形：动词 → [≤6字] → 方向标记 → [≤8字] → 对象——「筛选符合预期的实验记录」选择性由紧邻结构自足）、B 反向抑制形两种语序（反面/不利/负面/反对/异议/相反 + 对象 + 略去/隐去/删除/抹掉/屏蔽/过滤/排除/剔除/忽略/回避/不谈；或抑制动词在前 + 不利/反面 + 对象——「隐瞒了所有不利记录」）、C 定性短语形（选择性取证/举证/采证/引用/采信/呈现/报告/披露、断章取义、报喜不报忧、只捡好的、只拣好的、只谈成绩；EN cherry-pick + 对象 / quote out of context / suppress/omit/exclude + contrary/negative + 对象；EN only/solely/selectively + 取证动词 + 支持性标记 + 对象两种语序）；腿② SEAL 确定性结论封装三形任一——A 连接词形（连接词 所以/因此/因而/足见/据此/可见/故而/于是/那么；EN so/therefore/thus/hence/which means → [≤18字] → [确定性副词 一定/必然/必定/肯定/完全/毫无疑问/毋庸置疑/不容置疑/无疑 → [≤6字] →] 封结论名词 成立/为真/属实/正确/没错/无误/可靠/可信/有效/没有问题/无懈可击/站得住/毫无疑问/准确/更优/更好/更强/更高；EN correct/true/valid/reliable/sound/right/accurate/solid/watertight；或封结论动词 证明/证实/印证/坐实/说明/确立；EN proves/confirms/validates/establishes/demonstrates → [≤10字] → 封结论名词）、B 直接封装形（[足以/完全可以] + 封结论动词 → [≤8字] → 封结论名词——「足以证明产品没有问题」无连接词，是最常见攻击形，必须收）、C 情态封装形（确定性副词 → [≤6字] → 封结论名词，跨句亦判：「报告只引用正面评价。方案完全可靠。」）；腿③ ORDER 顺序——腿①必须早于腿②。161 轮教训：全部收裸名词形态（收「支持」而非只收「支持性」，收「正面」而非只收「正面的」）。r182 实现教训（已固化进测试）：① 封结论动词词表与连接词词表必须互斥——首版把「足以证明」放进 CONNECT_ZH，导致 A 形跨承重：注入 SEAL_VERB_ZH 后「足以证明产品没有问题」仍被 A 形（连接词 + 封结论名词）命中，负例测试假绿，剔除后 B 形才独家承重（同 r180「谓语必须用末组捕获组定位」教训）；② CONNECT_EN 不收 which proves / that proves——它们同时是 EN B 形的封结论动词前缀，收了会让 EN A' 形跨承重（注入 SEAL_VERB_EN 假绿）；③ A2 无副词形让 SEL_ADV_ZH 只对「动词-标记间隔 7-10 字」的长句独立承重，注入样本专门取该形（「只收集了来自三个渠道的支持性反馈」），非偶然通过；④ 豁免表刻意不收「选择性取证/断章取义」裸词（META_ZH）——它们同时是腿① C 形攻击词，收了会把自己的腿①豁免掉（r180 同款教训），META 只收「确认偏误/认知偏差/cherry-picking fallacy」等概念名。六类豁免全部负例注入承重：概念/教学转述、纪律性否定（不能/没有 + 只/选择性）、结论侧否定与存疑（不足以/仍需/待验证）、对冲弱化、疑问句。src/cherry-picking.js 新增 26 个可注入常量（SEL_ADV/VERB/MARK/OBJ_ZH、SEL_NEG_QUAL_ZH、SEL_SUPPRESS_ZH、SEL_PHRASE_ZH×EN、SEL_EN_ADV/VERB/MARK/OBJ、CONNECT_ZH×EN、SEAL_VERB/ADV/NOUN_ZH×EN、META_ZH×EN、NEG_ZH、NEGSEAL_ZH、HEDGE_ZH×EN），腿①/腿②正则全部由子表动态拼接（全部可注入），index.js 九处接线（require/计算点/allDims/dimMap/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports；变量名 chp，grep 实测无冲突）。来源（真实，未编造）：HN Algolia "logical fallacy detection" 命中 arXiv:2405.02318v3《Autoformalizing Natural Language to First-Order Logic: A Case Study in Logical Fallacy Detection》（HN 258 分 / 143 评论，2025-03-04，https://arxiv.org/abs/2405.02318）——NL2FOL 框架把自然语言翻译成一阶逻辑后用 SMT 求解器判定逻辑有效性，以逻辑谬误检测为案例；该框架依赖 LLM 翻译，与纯规则引擎不同构，只取其谬误族清单作概念背景。GitHub API "logical fallacy detection"（sort=stars）最高星为 tmakesense/logical-fallacy ★20，其余 ≤4 星且均为 LLM 路线，无可移植纯规则判据。Semantic Scholar 三次检索均 HTTP 429 限流，如实记录。侦察落盘 /tmp/hf-scout-1761925800.txt，差距探针 /tmp/hf-probe-r182.js。判据为心虫自有纯规则实现。维度数 179→180，VERIFY 138→139。验证：round-182 轮测 73/73（攻击 10/10 含中英/裸名词/带修饰/长句/跨句/无连接词、良性 17/17 零误判含纪律否定/结论侧否定/概念转述/对冲/无标记/边缘多因陈述/selection_bias 与 survivorship_bias 射程交界良性、20 项常量负例注入承重、6 条豁免陷阱翻转含 META/NEG/NEGSEAL/HEDGE 中英三族、登记断言×4）、gate 层变体矩阵 8/8 非 pass 且 findings 全含 cherry_picking、良性 8/8 pass（含自建豁免/对冲/概念转述/无标记/无关多因/EN 来源受限）、双向门禁 52/52 召回 302/326 误拦不升（与基线一致，零新增误拦）、verify 16/16、guard 20/20（维度登记 180 键）、doc 22/22。已知边界：① 腿③只做前向（先取证、后封装），「结论毫无疑问，因为报告只引用正面评价」倒装形语义同族但极少见，本轮保守不收；② 「成功案例/活下来的样本」形（「只统计了成功案例，足以证明方法有效」）与 survivorship_bias 射程交界，本轮保守不收；③ 存量维度 absolute_claim / confidence 会代拦部分本族边缘长句（r182 实测：攻击「为了说明这项政策有效，研究者只摘取了成功地区的材料，结论毋庸置疑。」由 reasoning_coherence + confidence 代拦——非本维度引入，改动前行为一致）。 |
| 6.8.109 | 2026-11-01 | 第179维度 halo_effect（光环效应 / halo effect）上线 — 由无关的正面特征（外貌/名气/穿着/装修包装/平台品牌/口才嗓音等非认知特征）推断对象在另一维度上的人品、能力、技术或观点正确性，且迁移总以肯定情态封装（肯定/一定/想必/must be），看上去像结论，实则零证据：「他长得这么帅，人品肯定也没问题」「这家公司装修得这么高档，产品技术肯定是一流的」「他口才这么好，做事能力一定也很强」「这 App 界面做得这么精致，底层架构肯定也差不了」「He is tall and handsome, so his judgment must be sound.」「She is famous, so her advice must be reliable.」。呈现维度的属性与实质维度的属性之间没有可传递的证据链——对呈现的好感不能增加对实质的任何支持；该迁移会系统性地让评价被表象带偏（产品评测被 UI 做工带偏、人选评估被口才外貌带偏、观点采信被讲者名气带偏），是 AI 输出中最常见的似是而非之一。轮换纪律：r179/r180 两轮为逻辑谬误检测一线，本轮切回判断偏差一线（r179 探针遗留候选 halo_effect 零覆盖，r181 复测 6/6 攻击穿过硬闸门、良性 3/3 干净）。判据三件共现（缺一不判）：腿① TRAIT 无关正面特征——外貌类（帅/俊/好看/漂亮/英俊/姣好/面善/养眼/颜值高/高大/高挑）、声音风度类（声音好听/声音磁性/气质好/风度翩翩/举止得体/幽默/风趣/口才好/能说会道/会做人）、名气类（有名/出名/当红/网红/明星/大V/粉丝多/热搜/人气高/家喻户晓/如雷贯耳）、包装装修类（包装精美/封面好看/界面精致/设计精美/装修高档/地段好）、平台品牌类（大公司/大厂/名牌/奢侈品/五星级/奢华/高端/上档次/价格不菲/百年老店/进口品牌）；EN tall / handsome / good-looking / attractive / beautiful / famous / celebrity / influencer / viral / premium packaging / sleek design / luxury brand / prestigious brand / great speaker / funny / witty——161 轮教训：全部收裸名词形态（收「帅」而非只收「长得帅」，收「有名」而非只收「有名气」）且容忍中间插入修饰语（「口才这么好」「人气很高」「装修得这么高档」——r181 首版实测漏判「他口才这么好，做事能力一定也很强」，根因是特征正则要求紧邻，已加修饰窗口修复）；腿② TRANSFER 质量迁移断言两形——A 连接词形（连接词 所以/因此/可见/说明/证明/so/therefore/thus/which means → 结论名词 人品/能力/技术/质量/观点/判断/方案/产品/服务/架构/体验/口碑/信誉 → 肯定情态 肯定/一定/必然/想必/显然/势必/也该/应该/must be/certainly）与 B 并置迁移形（腿①特征词 → [≤12字] → 结论名词 → [≤6字] → 肯定情态——「他长得这么帅，人品肯定也没问题」无连接词、仅靠并置完成迁移，是最常见攻击形，必须收）；腿③ ORDER 顺序——A 形要求腿①命中早于腿②。情态词表刻意不收「也」——「她长得漂亮，成绩也很好」是并列陈述非迁移推断，收了误伤面大于收益。七类豁免全部负例注入承重：概念/教学转述（光环效应/晕轮效应/月晕效应/认知偏差/美貌偏见/以貌取人；EN halo effect / horn effect / cognitive bias / lookism）、对比/转折与迁移否定（虽然/但是/然而/不过/却/不等于/两回事/无关/不能因为/不代表/不应以；EN but / however / irrelevant / not because / don't judge / has nothing to do with）、对冲弱化结论（可能/也许/或许/大概/似乎/不见得/未必/说不准/maybe/perhaps/might）、疑问句。src/halo-effect.js 新增 14 个可注入常量（TRAIT_ZH 五子族 LOOK/MANNER/FAME/WRAP/BRAND 可分别注入、TRAIT_EN、QUALITY_ZH×EN、MODAL_ZH×EN、CONNECT_ZH×EN、CONTRAST_ZH×EN、META_ZH×EN、HEDGE_ZH×EN），腿②正则由 trait/quality/modal 三表动态拼接（三表均可注入），index.js 九处接线（require/计算点/allDims/dimMap/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports；变量名 hef，grep 实测无冲突）。来源（真实，未编造）：HN Algolia "attractiveness bias hiring" 命中 Hacking attractiveness biases in hiring? The role of beautifying photo-filters（135 pts）、"halo effect" 命中 Cognitive Bias Spotted in the Wild: Halo Effect（1 pt）；GitHub API "halo effect bias" 命中 ekateneva-code/the-news-coverage-of-civil-nuclear-energy ★1（研究数据仓库，主题 The 'Halo' Versus 'Horn' Effect）与 henu-wang/bias-detector-9 ★0（Halo Effect Check 认知偏差检测工具）；Semantic Scholar 三次检索均 HTTP 429 限流，如实记录。以上均为概念背景，判据为心虫自有纯规则实现。grep -rl 差距扫描：halo / 光环在 src/ 下 0 命中；probe 实测 6/6 攻击穿过硬闸门。维度数 178→179，VERIFY 137→138。验证：round-181 轮测 45/45（攻击 17/17 含五特征族+A 连接词形专属样本、良性 15/15 含对比分离/概念转述/对冲/无迁移情态并列/疑问句、13 项注入承重含五子族分别注入、6 条豁免陷阱翻转 CONTRAST/META/HEDGE 中英三族、5 项结构区分、登记断言×2）、gate 层变体矩阵 10/10 非 pass 且良性 7/7 pass（含裸名词/带修饰/倒装/长句/EN）、双向门禁 52/52 召回 302/326 误拦不升（与基线一致，零新增误拦）、verify 16/16、guard 20/20（维度登记 179 键）、doc 22/22。已知边界：① 资质类特征（名校/博士/头衔/获奖）不在腿①——「她是名校博士毕业，说的话一定更有道理」是资质→观点的迁移，语义同族，但与 appeal_to_authority 射程交界且需同域/异域判别，本轮保守不收；② 让步+迁移混合形（「他虽然帅，但人品肯定好」）被对比豁免放过，保守方向取误判率可控；③ 存量维度 reasoning_coherence / absolute_claim / contradiction 会代拦部分本族边缘长句（r181 实测：良性样本「他给我们讲了光环效应：因为一个人长得好看，就认定他的人品也一定好，这就是以貌取人」被存量维度判 rewrite、EN stunning 样本被 contradiction 判 verify——非本维度引入，改动前行为一致，已从 gate 级矩阵移入维度级断言）。 |
| 6.8.108 | 2026-11-01 | 第178维度 fallacy_fallacy（谬误谬误 / fallacy fallacy / argumentum ad logicam）上线 — 指认对方的论证犯有谬误或推理无效，据此断定对方的结论为假、应被驳回、不可信：把「驳倒论证」冒充「驳倒事实」。论证无效（invalid）与结论为假（false）是两个独立的命题状态——一个无效论证完全可以推出一个为真的结论（「所有猫会飞，咪咪是猫，所以咪咪会飞」前提全假而结论碰巧为真）；把「我驳倒了你的推理」当成「我驳倒了你的主张」，是用修辞胜利冒充真理判定，会系统性地把「还没被证明」误报成「已被证伪」，把讨论从证据竞赛降级成谬误点名竞赛。这是判别者「寻真而非似是而非」目标的核心缺口，也是 AI 输出中最常见的似是而异之一。任务书轮换纪律：r176-r179 连续四轮为概率推断/判断纪律/逻辑谬误一线，本轮仍切逻辑谬误检测（该方向上轮使用为 r179 moralistic_fallacy），但补的是该方向的结论侧缺口（存量谬误维度全在前提侧）。判据为三腿共现（缺一不判）：腿① FALLACY 谬误指认——指名谬误裸名词（稻草人/偷换概念/循环论证/因果倒置/滑坡谬误/非黑即白/以偏概全/幸存者偏差/确认偏误/锚定效应/可得性偏差/赌徒谬误/基本率谬误/合取谬误/区群谬误/机械类比/歧义谬误/人身攻击/诡辩/谬误/逻辑错误/推理错误/论证无效；EN straw man / red herring / circular reasoning / slippery slope / false dilemma / hasty generalization / survivorship bias / ad hominem / non sequitur / logical fallacy——161 轮教训：不收裸「滑坡」（与地质义同形），只收滑坡谬误/滑坡论证/滑坡推论）或泛指论证缺陷（「犯了/这是/存在/充满 + 缺陷名词」，或「你的论证/推理/逻辑 + 问题/漏洞/错误/无效/站不住脚」）；腿② REJECT 结论废弃——连接词（所以/因此/因而/可见/足见/据此/那么/so/therefore/thus/hence/since）→ 结论名词（观点/结论/主张/说法/论点/论断/命题/claim/conclusion/point/thesis……）→ 假值谓语（是错的/是假的/不成立/荒谬/一派胡言/不足为信/不可信/被证伪/大错特错/全盘否定/该被抛弃/false/wrong/bogus/refuted/dismissed）；腿②必须落在结论名词上，不能落在论证名词上——「你的论证不成立」是正确判别（论证未获支持），「你的观点不成立」才是本族，这是本维度最重要的结构性区分；腿③ ORDER 顺序——腿①必须早于腿②。r180 实现教训（已固化进测试）：① 腿②谓语必须用末组捕获组定位——首版 PREFALSE 否定前置豁免取整条腿②匹配起点，落在连接词之前，导致「所以你的观点不是错的」的否定豁免永不承重（负例测试假绿），改用 falsePos = 起点 + 全长 - 组长 后陷阱才真正翻转；② CORRECT「不等于/推不出」豁免查腿①之后全段（而非仅两腿之间），否则尾句补刀形（「……所以你的观点是假的；不过我要说明，论证无效这一事实本身，推不出任何关于结论真假的判断」）漏豁免；③ PREFALSE 词表刻意不放宽到裸「不能/不会/不可能」——那会误杀「所以你的观点在逻辑上就不能成立」这类攻击形，只收能明确否定假值断言本身的形式（未必/不一定/并非/不是/算不上/不等于/不代表/推不出/说明不了/not/never/does not）；④ 豁免表不能收「犯了谬误」这类同时是攻击形的组合，否则把自己的腿①豁免掉。七类豁免全部负例注入承重：概念/教学转述、谓语前否定与对冲、不等于/推不出、结论名词缺失或落在论证名词上、对冲、疑问句。src/fallacy-fallacy.js 新增 14 个可注入常量（FALLACY_ZH/FALLACY_GENERIC_ZH/FALLACY_EN、REJECT_ZH/REJECT_EN、PREFALSE_ZH×EN、CORRECT_ZH×EN、META_EXEMPT_ZH×EN、HEDGE_ZH×EN），index.js 九处接线（require/计算点/allDims/dimMap/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports；变量名 flf——ff 已被 future_faking 占用，命名冲突被 node --check 抓出后改名，同 r178/r179 先例）。来源：三层外部搜索全部拿到结果——HN Algolia "logical fallacy" 命中 arXiv 2405.02318（Translating natural language to first-order logic for logical fallacy detection，258 pts）、"Logical Fallacies Are Usually Irrelevant or Cited Incorrectly"（176 pts）、"Show HN: Aristotl – an intuitive logical fallacy lookup tool"（173 pts）、"Master List of Logical Fallacies (2018)"（113 pts）；GitHub API 命中 LukeStonehm/LogicalDefence ★140 / causalNLP/logical-fallacy ★96（"Detecting Logical Fallacies: From Quiz to Climate Change News"）/ andrewscwei/aristotl ★43；UTEP Master List of Logical Fallacies 全文 223,983 字节已落盘 /tmp/hf-master-fallacies.txt。Semantic Scholar 一次检索无返回，如实记录。grep -rl 差距扫描：fallacy_fallacy / ecological_fallacy / planning_fallacy / attribute_substitution / galileo_fallacy 在 src/ 下 0 命中；ad_hominem / prosecutors_fallacy / whole_part_transfer 已有。判据为心虫自有纯规则实现。维度数 177→178，VERIFY 136→137。验证：round-180 轮测 37/37（攻击 16/16 含中英/裸名词/带修饰/倒装/长句、良性 17/17 零误判含谓语前否定/不等于推不出/概念转述/结论名词不匹配/对冲/疑问句、7 项常量负例注入承重、7 条豁免陷阱翻转含 PREFALSE/CORRECT/META 三族、腿②落在论证名词专测/结论名词不匹配专测/顺序倒置专测、登记断言×2）、gate 层变体矩阵 11/11 非 pass 且良性 6/6 pass、双向门禁 52/52 召回 302/326 误拦不升（与基线一致，零新增误拦）、verify 16/16、guard 20/20（维度登记 178 键）、doc 22/22。已知边界：① 腿③只做前向（先谬误指控、后结论废弃），反向形「你的观点是错的，因为你犯了稻草人谬误」语义同族但因几乎总以「因为」连接、与大量合法因果解释同形，本轮保守不收；② 长句中修饰性谬误指认（「一位既支教又出诗集的青年：你这是稻草人」）可能触发泛指缺陷腿——该族良性文本量更大，保守方向取误判率可控的 VERIFY 级；③ 存量维度 scrutiny_evasion / presupposition / bad_faith / fallacies 会代拦部分本族边缘长句（r180 实测：良性样本「你的论证确实有谬误，但这只说明推理有问题，不等于你的结论为假」被 scrutiny_evasion 判 verify，非本维度引入，改动前行为一致）。 |
| 6.8.107 | 2026-11-01 | 第177维度 moralistic_fallacy（道德主义谬误 / moralistic fallacy）上线 — 把应然规范倒推为实然事实：「世界应该公平」推不出「世界本来公平」，「人理应诚实」推不出「人的天性就是诚实」——规范（ought）描述应当如何，事实（is）描述实际如何，二者是两个命题域，用规范词给事实断言作前提，是把价值主张冒充成事实陈述（自然主义谬误的镜像族，appeal_to_nature 管的是反方向：自然性→优越性）。任务书轮换纪律：r176-r178 连续三轮为因果/判断/证据纪律方向，本轮切回逻辑谬误检测一线（该线上轮使用为 r172 whole_part_transfer）。判据为三要素共现（缺一不判）：腿① OUGHT 应然规范词（应该/应当/理应/理当/本该/应然/按理；EN ought to / should / supposed to / meant to——裸名词形态）；腿② ISNATURE 实然自然/本质等式断言（本来就是/事实上就是/本质上都是/天性就是/天生就是/本性就是/归根到底就是/就是事实；EN by nature / naturally / innately / inherently / fundamentally / only natural——只收「自然/本质/事实标记+肯定系动词」形，「天性未必就是」「本质上并非」等否定形因缺肯定系动词天然不匹配）；腿③ ORDER 顺序——腿①必须早于腿②（顺序倒置即非本族：「善良是人性的一部分，所以人应该向善」是合法论证）。结构豁免：两腿之间的前提连接词（因为/由于/because/since）拦截实然作前提的合法论证形。r179 实现设计（已固化进测试）：① 腿②把否定排挤出腿本身而非事后否定豁免——避免误伤「本质上都是无私的」攻击形；② OUGHT_AFTER 豁免「本性就是应该诚实的」元陈述形；③ NORMNOUN 豁免「本来就是底线/义务/追求」的规范性重申。七类豁免全部负例注入承重。src/moralistic-fallacy.js 新增 14 个可注入常量（OUGHT_ZH×EN、ISNATURE_ZH×EN、PREMISE_ZH×EN、NEG_ZH×EN、NORMNOUN_ZH、OUGHT_AFTER_ZH、META_EXEMPT_ZH×EN、HEDGE_ZH×EN），index.js 九处接线（require/计算点/allDims/dimMap/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports；变量名 mof——mf 已被 moral_foundations 占用，命名冲突被 node --check 抓出后改名）。来源：HN Algolia "moralistic fallacy"/"is-ought" 两关键词 0 命中；GitHub "moralistic fallacy" total_count=0，logical-fallacy 检索群（tmakesense/logical-fallacy ★20 / SeanFlannery/Logical.ly ★4 / graciolli-f/sophron-bot ★4 / EPITA-SCIA/Fallacy-Detection-LLM ★2 / IMJONEZZ/Fallacy-Detection ★1）全部 LLM-based，与纯规则引擎不同构，仅作概念背景；Semantic Scholar 两次检索均 HTTP 429 限流，如实记录；判据为心虫自有纯规则实现。差距实测（/tmp/hf-probe-r179.js 候选池 10 族 + /tmp/hf-probe-r179b.js 深探，直调 src/gate.js）：本族攻击 12 条中 8 条真零命中（存量维度无任何一腿覆盖该结构），良性 12 条按新判据设计零命中。侦察落盘 /tmp/hf-scout/hf-scout-r179.txt。维度数 176→177，VERIFY 135→136。验证：round-179 轮测 29/29（攻击 11/11 含中英/裸名词短句/带修饰/倒装/长句、良性 16/16 零误判含顺序倒置合法形/批评性分析/法律条文/概念转述、10 项常量负例注入承重、8 条豁免陷阱翻转、顺序倒置专测、登记断言×2）、gate 层变体矩阵 7/7 非 pass 且良性 5/5 pass、双向门禁 52/52 召回 302/326 误拦不升（与基线一致，零新增误拦）、verify 16/16、guard 20/20（维度登记 177 键）、doc 22/22。已知边界：规范词在前、自然/本质断言在后但无语境连接词的句子（「学生应该多读书，读书本来就是人类的天性」）可能误判 verify——该族良性文本量更大，保守方向取误判率可控的 VERIFY 级。 |
| 6.8.106 | 2026-10-31 | 第176维度 conjunction_fallacy（合取谬误 / conjunction fallacy / Linda problem）上线 — 把合取事件（A 且 B）的概率/可能性宣称高于其任一组分事件（A 或 B）的概率，违反概率公理 P(A∧B) ≤ P(A)：两个条件同时满足的场合必然少于任一条件单独满足的场合，与描述生动与否、叙事合理与否无关；把合取描述得更具体就认为它「更可能」，是把叙事的代表性冒充概率的大小：「琳达既是银行出纳又是女权主义者的概率，高于她只是银行出纳的概率」「他既是名校毕业又有十年经验的可能性，比他只是名校毕业的可能性更大」「一个产品既便宜又耐用的概率，大于它只是便宜的概率」「The probability that Linda is both a bank teller and a feminist is higher than the probability that she is a bank teller.」「It is more likely that she is both a philosopher and a poet than that she is merely a philosopher.」。任务书轮换纪律：r175-r177 连续三轮为判断纪律方向（outcome_bias/common_cause/hindsight_bias），本轮切到证据纪律/概率推断一线。判据为三腿共现（缺一不判）：腿① CONJ 合取标记——组合形（既是…又是/既…又…/既…也…/不但…而且/不仅…还/既会…又会/又…又…；EN both … and / as well as / and also）或裸词形（兼具/兼有/兼备/同时具备/同时具有）或裸连接词「且」「以及」（与腿②概率词 ≤12 字窗邻近，不单独承重——「便宜而且耐用」是语篇递进不是逻辑合取）；腿② PROB 概率词（概率/可能性/几率/机率/或然率；EN probability / probable / chance / likelihood / odds / likely——裸名词形态）；腿③ CMP 概率比较结构两条路线任一——C-a 比较算子（高于/大于/超过/胜过/优于/更可能/可能性更大/更高/更大；EN higher than / greater than / larger than / more probable / more likely / exceeds）× 组分回指（只是/仅仅是/单是/不过是/仅/只/无非；EN just / only / merely / simply / purely / alone）落在算子后 25 字（EN 80 字符）窗内；C-b 两个概率词 × 中间夹比较算子（覆盖「A 的概率高于 B 的概率」的数学陈述形与倒装形）。结构豁免（防误伤良性条件概率比较）：合取标记前后 12 字（EN 40 字符）内有人群后缀（的人/的学生/的应聘者/的选手/的群体/的成员；EN people who / those who / students who / candidates who / employees who / applicants who / groups of）时，该句是 P(结果 |
| 6.8.105 | 2026-10-31 | 第175维度 hindsight_bias（事后聪明偏差 / hindsight bias / 后见之明 / knew-it-all-along effect）上线 — 把事后已知的结果当成事前本可预见、注定如此的，用后见之明倒推事前的判断力，混淆「事后知道」与「事前可知」：「这件事早就注定会失败，明眼人一开始都能看出来」「我早就知道他会背叛我们，一切都在意料之中」「结果不出所料，这件事从一开始就是注定的」「事后看来这一切都很明显，当时我们太迟钝了」「这场危机早就埋下了伏笔，聪明人一眼就能看穿」「他早就看出这家公司没有未来」「This was always going to fail; anyone could have seen it coming.」「I knew it all along — the outcome was inevitable from the start.」「Hindsight makes it obvious, but we should have known better.」「we all knew this project would collapse eventually.」。任务书轮换纪律：r176 为因果纪律方向（common_cause），本轮接 r176 报告预立的下一轮候选「事后聪明族 5/8 攻击穿过零覆盖」，切到判断纪律一线。判据为三形任一，命中即判：腿① HS 事后框架——成语形（事后诸葛亮/马后炮；EN after-the-fact / Monday morning quarterback）单腿可命中，弱形（事后/如今/现在回想/回过头/回望/以今天的眼光；EN in hindsight / with hindsight / looking back / in retrospect / with the benefit of hindsight / now that we know——裸名词形态）必须与腿②共现；腿② PRED 注定/事前可预见断言——强形（早就注定/从一开始就注定/命定/不出所料/果不其然/可想而知/迹象一直都在/明眼人/一眼就能看穿/一看便知；EN was always going to / doomed / inevitable / unavoidable / should have known / should have seen / could have seen / anyone could have / we all knew / knew it all along / saw it coming / signs were there / foregone conclusion / never had a chance / only a matter of time）单腿可命中（词形内嵌「从事前起点回望」的事后语义），POST 子集（早就知道/早就看出/早就预见/伏笔/埋下伏笔/本可预见/当时就该/显而易见/很明显/谁都能看出/一眼看出/意料之中/命中注定）与腿③共现判定，弱形（注定/迟早/早晚/必然/不可避免/很明显；EN bound to / obvious / clearly / predictable / foreseeable）只与腿①共现判定（刻意不与 OUT 共现——「这场比赛注定会赢」是赛前打气，良性）；腿③ OUT 已发生结果事件（失败/倒闭/破产/崩溃/崩塌/爆雷/翻车/暴跌/暴涨/危机/灾难/出事/分手/离婚/背叛/夺冠/落榜/去世/中奖/没有未来/垮掉/完蛋/跑路/裁员/退市/崩盘/战败/惨败/栽跟头；EN failed / failure / collapse / crashed / bankrupt / betrayal / betrayed / fell apart / blew up / defeat / disaster / crisis / broke up / divorced / lost / tanked / flopped）——把注定断言锚定到具体已发生结果上，排除「未来注定会更好」类将来时一厢情愿（归 wishful_thinking）。r177 实现教训（已固化进测试）：① PRED_ANY 拼接首版漏走注入通道（PRED_WEAK/STRONG 直接用模块常量），负例注入承重项假绿——共现判定腿的每个常量都必须经 R.XXX |
| 6.8.104 | 2026-10-31 | 第174维度 common_cause（共同原因谬误 / common cause / confounding factor / 混杂因素）上线 — 把两个变量的共变或相关直接当成单向因果，不排查、不提及背后同驱的第三因素（Z→X 且 Z→Y 被误读成 X→Y）：「夏天冰淇淋销量和溺水死亡同步上升，说明吃冰淇淋导致溺水」「孩子身高和词汇量同步增长，说明背单词能让人长高」「度假胜地蚊香销量和登革热病例同步上升，蚊香导致登革热」「A 指标与 B 指标高度相关，可见 A 是 B 的原因」「这两个指标同涨同跌，说明其中一个驱动另一个」「Ice cream sales and drowning deaths rise together every summer, so ice cream causes drowning.」「The two metrics move in lockstep, which shows one drives the other.」「A and B are highly correlated, so A causes B.」。任务书轮换纪律：r175 为决策评价/判断纪律方向（outcome_bias），本轮切回因果与反事实推理一线，接 r175 报告预立的下一轮候选（共同原因族 4/5 穿过）。判据为三腿共现（缺一不判）：腿① COVARIATION 共变/相关陈述两形任一——C-a 相关名词形（相关/关联/高度相关/正相关/负相关/显著相关/密切相关，必含双变量、免并列结构；EN correlated / correlation / covary / covariation）；C-b 共变动词形（同步上升/同步增长/同步下降/一同/一齐/一起+上升类/齐涨/齐升/齐跌/同涨同跌/同升同降/相伴上升/如影随形/齐头并进/亦步亦趋/水涨船高/随之上升/联动；EN rise(s)/rose together / move(s)/moved together / in lockstep / hand in hand / in tandem / in sync / track each other / rise and fall together——裸名词形态）且共变词前 40 字（EN 100 字符）窗内有并列双变量结构（和/与/跟/及/以及/这[两几][个种]/那[两几][个种]/两者/二者；EN and / & / both / the two / these two / two variables / two metrics）；腿② CAUSAL 直接因果动词（导致/造成/引发/引起/带来/催生/促成/使得/令/让/驱动/推动/推升/推高/拉高/抬高/造就/滋生/助长/加剧/加大了/增加了/提高了/提升了/带动/拉动/源于/来自于/归因于/归功于/是…的原因/根源在于/罪魁祸首；EN causes / leads to / results in / results from / drives / boosts / increases / reduces / makes / creates / produces / is responsible for / brings about / stems from / originates from / is due to）与腿①双朝向 60 字（EN 120 字符）窗共现（覆盖「共变→因果」主形与「A 是 B 的原因，因为 A 与 B 高度相关」倒装形）；腿③ NO_CONFOUND 未点名/未控制共同因素（共同原因/混杂/第三因素/同一个原因/都源于/控制了/剔除/回归/辛普森/DAG/工具变量/倾向得分；EN common cause / confounding / confounder / third variable / lurking variable / controlling for / adjusted for / Simpson / DAG / instrumental variable / propensity / causal inference）。r176 实现教训（已固化进测试）：① 腿②刻意不收「说明/表明/证明/可见」等推断词——「与去年同期同步上升，说明旺季策略见效」是良性同比对照，首版若收推断词整类误判；② CAUSE 表首版漏「推动/推升」词形，「足以证明前者推动了后者」整句漏判；③ NO_CONFOUND_EN 的 both driven by 必须收插入语形 both are driven by——「but both are driven by age」整句漏判；④ META_ZH 首版漏「教材」词形，教学转述陷阱无法翻转；⑤ NEG_EN 必须收 does not cause/drive/lead 否定形。五类豁免全部负例注入承重：概念/教学转述（相关不等于因果/相关并非/相关不蕴含/因果推断/统计学/教科书/教材/课堂/论文/谬误；EN correlation does not imply / correlation is not / confounding / statistics / textbook / lecture）、实验/证据语境（随机对照/对照组/双盲/安慰剂/前瞻性/队列/元分析/临床试验/数据表明/统计显示/研究表明/证据表明/回归/自然实验；EN randomized / controlled trial / RCT / placebo / double-blind / cohort / meta-analysis / experiment / data show / evidence shows / studies show / regression / natural experiment）、推理否定/反向承认（不能说明/不代表/不等于/并不意味着/推不出/不足以/未必/不一定/很难说/不见得/只是相关/另有原因/反过来/恰恰相反；EN does not mean / cannot conclude / not necessarily / does not prove / does not cause / just because / the reverse / conversely）、对冲（可能/也许/或许/大概/恐怕/倾向于/似乎/好像/说不定/有待；EN maybe / perhaps / might / possibly / probably / seems / appears / could be）、已点名/已控制共同因素（NO_CONFOUND 表）、疑问句。src/common-cause.js 新增 18 个可注入常量（COVARY_REL/COVARY_MOVE/PAIR/CAUSE_ZH×EN、EXEMPT_META/EVID/NEG/HEDGE_ZH×EN、NO_CONFOUND_ZH×EN）与双路线共变判定 + 双朝向因果共现腿，index.js 九处接线（require/计算点/dimMap/allDims/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports；变量名 ccau——cc 已被 confidence 读数占用）。来源：HN Algolia "confounding variable" 命中 "Statistics: When Confounding Variables Are Out of Control"（objectID 11414793）与 "Understanding Confounding Variables: A Comprehensive Guide (2024)"（41779877）——确认混杂/共同原因是统计与因果推断的公认盲区；GitHub causal-inference 仓库群 NSAPH-Projects/space ★19 / Trhova/Causal-Microbiome-Omics ★19 / imkemayer/causal-inference-missing ★9 / moprescu/GSTUNet ★8 均为学术数据科学工具，非纯规则文本判别，仅作概念背景；Semantic Scholar 两次检索均 HTTP 429 限流，如实记录；判据为心虫自有纯规则实现。差距实测（/tmp/hf-probe-r176.js 直调 src/gate.js）：共同原因族 8 条攻击样本 4 条穿过硬闸门且零维度命中（冰淇淋溺水/背单词长高/蚊香登革热/雷雨哮喘），良性零误判；pseudo_causal 只管精确倍数与时序/玄学归因、reverse_causation 只管方向倒置，射程不重叠。侦察落盘 /tmp/hf-scout-r176.txt。维度数 173→174，VERIFY 132→133。验证：round-176 轮测 27/27（攻击 18/18 含中英/裸名词/带修饰/倒装/长句、良性 16/16 零误判含正确多因陈述/同比对照/纯描述/推理否定/对冲/概念教学/实验证据/已点名共同因素、12 项常量负例注入承重、5 条豁免陷阱翻转、疑问句豁免、腿②推断词不承重、登记断言×4）、gate 层变体矩阵 7/7 非 pass 且良性 7/7 pass、双向门禁 52/52 召回 302/326 误拦不升（与基线一致，零新增误拦）、verify 16/16、guard 20/20（维度登记 174 键）、doc 22/22。 |
| 6.8.103 | 2026-10-31 | 第173维度 outcome_bias（结果偏误 / outcome bias / 结局偏见）上线 — 用结果的好坏评判决策/选择的质量，忽视决策当时可获得的信息与论证质量，把运气当成能力：「他这次赌对了，说明当初的选择是明智的」「手术成功了，足以证明医生的方案是正确的」「球队夺冠了，可见主教练的战术安排英明果断」「这只股票翻了三倍，证明他当初的买入决定眼光独到」「他亏得血本无归，可见当初的投资决策非常愚蠢」「新产品卖爆了，这说明当初的市场定位判断是对的」「The bet paid off, which proves his decision was wise.」「She lost everything, showing that her choice was clearly reckless.」。任务书轮换纪律：r169-r174 连续六轮为新判别维度（selection_bias/survivorship_bias/base_rate_neglect/whole_part_transfer/clustering_illusion/reverse_causation），本轮切到决策评价/判断纪律方向（脱离连续两轮的因果结构族）。判据为三腿共现（缺一不判）：腿① OUTCOME 高方差结果事件（赌对/赌错/押对/蒙对/翻倍/翻了三倍/获利/血本无归/亏光/爆仓/套牢/割肉/暴涨/暴跌/翻车/踩雷/暴雷/翻身/夺冠/出局/淘汰/落榜/考上/没考上/卖爆/大卖/滞销/中奖/活下来/去世/救回来/手术成功/手术失败/赢麻了/躺赢……裸名词形态；EN paid off / won / lost / hit the jackpot / struck it rich / made a fortune / went bankrupt / lost everything / came out ahead / fell flat / crashed / tanked / doubled / tripled / lottery）× 腿② EVAL 决策评价——决策名词（决策/决断/决定/抉择/选择/判断/策略/方案/战术/思路/想法/做法/打法/布局/操作/下注/押注/买入/卖出/投资/出手/转型/放弃/坚持/路线/赌注/定夺/拍板；EN decision / choice / call / bet / gamble / strategy / tactics / plan / move / pick / investment / trade / wager）× 评价词（明智/英明/睿智/高明/正确/是对的/果断/清醒/理性/聪明/精明/厉害/眼光独到/有远见/独到/愚蠢/草率/鲁莽/失误/失策/错误/糊涂/短视/冒进/不该/欠妥/败笔；EN wise / correct / brilliant / shrewd / sound / reckless / foolish / hasty / poor / well-judged / ill-judged / good call / bad call）落在 15 字（EN 60 字符）窗内 × 腿③ LINK 归因连接词（说明/证明/可见/足以证明/表明/意味着/证实/印证/体现出/反映出/所以/因此/因而/故而/毕竟/说到底/事实证明/结果证明/归功于/全靠/多亏；EN proves / show / demonstrating / confirms / means / which means / indicating / therefore / thus / hence / so / after all / thanks to / proof）；连接结构为 LINK 匹配点前 60 字（EN 120 字符）内有 OUTCOME 且后有 EVAL，或前有 EVAL 且后有 OUTCOME（结论先行的倒装形）——双朝向任一。r175 实现教训（已固化进测试）：① EVAL 腿必须收「是对的/是错的」与裸「独到」——「市场定位判断是对的」「眼光向来独到」整句因只收连续词形而漏判；② DECISION 腿必须收 tactics/tactic 与「思路/想法/做法」——「coach's tactics were brilliant」「答题思路非常聪明」整族漏判；③ LINK_EN 必须收裸 prove/show——「does not prove」式否定陷阱无法翻转。五类豁免全部负例注入承重：概念/教学转述（结果偏误/结局偏见/以结果论/以成败论/事后诸葛亮/认知偏差/行为经济学/教材/课堂；EN outcome bias / result-oriented evaluation / hindsight / cognitive bias / textbook / lecture）、实验/证据语境（随机对照/对照组/实验组/双盲/前瞻性研究/队列研究/元分析/临床试验/实验表明/数据显示/研究表明/实测/证据表明；EN randomized / controlled trial / RCT / control group / placebo / double-blind / cohort study / meta-analysis / experiment / data show / evidence shows——受控实验中结果经设计成为方法的有效证据，ex-ante 论证就是实验设计本身）、ex-ante 推理/推理否定/反向承认（不能说明/不能只凭/不能只看/不能仅凭/不能以/不要以/并非只看/不代表/不等于/并不意味着/推不出/不足以/未必/不一定/在当时/事前/当时的信息/做决定时/下决定时/决策当时/反过来/恰恰相反/反之；EN does not mean / cannot conclude / not necessarily / not only / at the time / ex ante / given the information available / the reverse / conversely）、对冲与运气承认（可能/也许/或许/大概/恐怕/倾向于/似乎/有待/运气/偶然/碰巧/侥幸；EN maybe / perhaps / might / possibly / probably / seems / luck / chance / fluke / lucky）、疑问句。src/outcome-bias.js 新增 16 个可注入常量（OUTCOME/DECISION/EVAL/LINK_ZH×EN、EXEMPT_META/EVID/NEG/HEDGE_ZH×EN）与双朝向连接判定腿，index.js 九处接线（require/计算点/dimMap/allDims/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports）。来源：HN Algolia "outcome bias" 命中 "Outcome Bias and the Interpreter: How Our Minds Confuse Skill and Luck"（objectID 6563620）、"Introduction to the Outcome Bias – Don't Judge Decisions by Their Outcome"（16543449）、"Outcome Bias: The Bias Behind the World's Greatest Catastrophes"（21131582）——确认「把结果当决策质量判据、混淆技能与运气」是公认判断失误；GitHub 无纯规则实现（jbarach2012/AIBF_API ★195 属招聘反歧视域，与纯规则引擎不同构，仅作概念背景）；Semantic Scholar 关键词检索无返回，如实记录；判据为心虫自有纯规则实现。差距实测（/tmp/hf-probe-r175.js 直调 src/gate.js）：结果偏误族 8 条攻击样本 7 条穿过硬闸门且零维度命中（仅 1 条被 reasoning_coherence 兜住），良性 5 条零误判；同池候选共同原因 4/5 穿过、事后聪明已被 overclaim 子信号部分覆盖，结果偏误为唯一真零覆盖族。侦察落盘 /tmp/hf-scout-r175.txt。维度数 172→173，VERIFY 131→132。验证：round-175 轮测 25/25（攻击 16/16 含中英/裸名词/带修饰/倒装/长句、良性 15/15 零误判含 ex-ante 正确评价/推理否定/运气承认/概念教学/实验证据/疑问/战争多因边缘良性、16 常量负例注入承重、8 条豁免陷阱翻转、登记断言×3）、gate 层变体矩阵 7/7 非 pass 且良性 7/7 pass、双向门禁 52/52 召回 302/326 误拦不升、verify 16/16、guard 20/20（维度登记 173 键）、doc 22/22。 |
| 6.8.102 | 2026-10-31 | 第172维度 reverse_causation（因果倒置 / reverse causation / 反向因果）上线 — 把对后果的应对行为当成后果本身的原因：人之所以采取某个行为，恰恰是因为不利后果已经存在或风险已经很高，于是「做这件事的人/地方，后果更严重」是选择效应（后果筛出了行为人），不是该行为制造了后果：「住院患者的死亡率明显更高，所以住院本身增加了死亡风险」「接受心理治疗的人自杀率更高，所以心理治疗会把人推向自杀」「报警越多的社区犯罪率越高，说明报警反而会招来犯罪」「报了补习班的孩子成绩更差，所以补习班让孩子成绩更差」「吃安眠药的人失眠更严重，因此安眠药是造成失眠的原因」「装防盗窗的家庭被盗更多，可见防盗窗反而会吸引小偷」「被约谈的官员更容易落马，所以约谈导致了他们落马」「戴头盔的骑手出车祸更多，这说明戴头盔会引发事故」「Patients who were hospitalized had much higher mortality, so hospitalization itself increases the risk of death.」「Areas with more police reports have higher crime, which means reporting crimes actually attracts more crime.」。任务书轮换纪律：r169-r173 连续五轮为新判别维度（selection_bias/survivorship_bias/base_rate_neglect/whole_part_transfer/clustering_illusion），本轮按轮换池切到因果与反事实推理一线。判据为两腿共现（缺一不判）：腿① RESPONSE_CORR 反应型行为×相关/选择框，两形任一——R-a 选择差异框：反应型行为名词（住院/就医/吃药/报警/投诉/补习/约谈/戴口罩/头盔/装监控/装防盗窗/买保险/裁员/封号……裸名词形态；EN hospitali[sz]ed / therapy / police reports / tutoring / wearing masks / cameras / insurance / layoffs）+ 群体名词（人/患者/孩子/社区/地区/公司/官员/家庭/用户/员工/骑手/居民/账号；EN patients / people / those who / areas / children / drivers）落在 6 字（EN 30 字符）窗内（前后双向，群体名词必须与「的」相邻——「住院患者的死亡率」与「住院的患者」两种框架都收，无「的」的巧合邻近不收），且句中存在差异/相关标记（更/多/高/严重/差/容易/越/集中/往往/大都/明显/显著/频繁；EN more / higher / worse / more often / likely / tend）；R-b 递增相关框：反应型行为 + 「越」（12 字窗）且全文「越」出现 ≥2 次（「报警越多的社区犯罪率越高」「裁员越多的公司效益越差」；EN the more X … the worse Y）；腿② CAUSE 因果归因：因果动词（导致/造成/引发/引起/带来/招来/引来/吸引/催生/制造/滋生/助长/加剧/加重/诱发/促成/酿成/加大了/增加了/提高了/提升了/推向/是…的原因/原因就在于/根源在于/罪魁祸首/让/令/使得；EN causes / leads to / results in / increases / worsens / is responsible for / contributes to / makes / attracts / draws） × 负面结果名词（死亡/死亡率/事故/犯罪/小偷/复发/恶化/失眠/抑郁/自杀/落马/倒闭/亏损/被骗/被盗/感染/感冒/更差/退步/风险/问题；EN death / mortality / accidents / crime / relapse / insomnia / depression / suicide / addiction / infections / losses / worse）落在 18 字（EN 45 字符）窗内（从因果动词起点起算——「是公司衰败的原因」形的结果名词在动词匹配内部）。r174 实现教训（已固化进测试）：① 群体名词的「的」在名词两侧都可能出现（「住院患者的死亡率」是名词+的，「住院的患者」是的+名词），首版只收「的+名词」导致整句漏判；② 腿② 目标窗若从因果动词匹配的终点起算，「是公司衰败的原因」会把结果名词整个吞进自己匹配里，窗内只剩句号，结果腿永远找不到负面结果（改为从起点起算）；③ 英文侧缺 attracts/draws/makes/worse 与 the more X … the worse Y 比较词同族化，整条英文族漏判。五类豁免全部负例注入承重：概念/教学转述（因果倒置/反向因果/混淆变量/相关不等于因果/辛普森悖论/内生性/统计学/教材/课堂；EN reverse causation / confounding / endogeneity / Simpson's paradox / textbook / lecture）、实验/证据语境（随机对照/对照组/实验组/双盲/前瞻性研究/队列研究/元分析/实验表明/数据显示/研究表明/实测/证据表明；EN randomized / controlled trial / RCT / control group / prospective / cohort study / meta-analysis / data show / evidence shows）、推理否定/反向承认（不能说明/不代表/不等于/并不意味着/推不出/不足以/未必/不一定/是因为/其实是因为/反过来/恰恰相反/反之；EN does not mean / cannot conclude / not necessarily / is because / the reverse / conversely）、对冲（可能/也许/或许/大概/恐怕/倾向于/似乎/有待；EN maybe / perhaps / might / possibly / probably / seems）、疑问句。src/reverse-causation.js 新增 18 个可注入常量（RESP/GROUP_DE/DIFF/YUE/CAUSE/RESULT_ZH×EN、EXEMPT_META/EVID/NEG/HEDGE_ZH×EN）与 R-a/R-b 双腿①路线 + 腿② nearFrom 判定，index.js 九处接线（require/计算点/dimMap/allDims/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports）。来源：HN Algolia "counterfactual reasoning"（KDD Tutorial on Causal Inference and Counterfactual Reasoning，objectID 23266032；What If It Goes Wrong?: The Hidden Dangers of Counterfactual Reasoning in AI，43754084）与 "causal inference"（Causal inference as a blind spot of data scientists，37890685，225 分；Seven basic rules for causal inference，41263928）——确认因果推断是数据科学与常识推理的公认盲区；GitHub fallacy-detection 仓库群 tmakesense/logical-fallacy ★20 / SeanFlannery/Logical.ly ★4 / EPITA-SCIA/Fallacy-Detection-LLM ★2 全部 LLM-based，与纯规则引擎不同构，仅作概念背景；Semantic Scholar 关键词检索返回 429，如实记录。差距实测（/tmp/hf-probe-r174b.js 直调 src/gate.js）：10 条反向因果攻击样本 7 条穿过硬闸门且零维度命中（仅报警/约谈/看医生 3 条被 pseudo_causal / fallacies 兜住），良性 10 条零误判；判据为心虫自有纯规则实现。侦察落盘 /tmp/hf-scout-r174.txt。维度数 171→172，VERIFY 130→131。验证：round-174 轮测 24/24（攻击 16/16 含中英/裸名词/带修饰/倒装/长句、良性 14/14 零误判含正确方向/选择效应承认/推理否定/实验证据/正向效果/对冲/疑问/概念教学/非应对动作、13 项常量负例注入承重、4 类豁免陷阱翻转、登记断言×3）、gate 层变体矩阵 7/7 非 pass 且良性 7/7 pass、双向门禁 52/52 召回 302/326 误拦不升（与基线一致）、verify 16/16、guard 20/20（维度登记 172 键）、doc 22/22。 |
| 6.8.101 | 2026-10-31 | 第171维度 clustering_illusion（集群错觉 / clustering illusion / Texas sharpshooter fallacy）上线 — 把随机数据中的聚集、连续、扎堆、集中出现的事件当成「绝不是巧合」「显然有蹊跷」「一定有特殊原因」「这是逃不掉的规律」的显著性证据：随机分布在时间、空间、类别上本就会出现聚集，把聚集读成必然、反常或神秘规律是把噪声当信号：「这个路口一个月出了四起事故，绝不可能是巧合」「最近三个月，同一个小区接连发生了五起入室盗窃，显然这里面有蹊跷」「三个中奖者全都来自同一条街，这绝不是偶然」「这个部门每逢周一必出事，简直是逃不掉的魔咒」「Four accidents happened at the same intersection in one month, which cannot be a coincidence.」「Three lottery winners all came from the same street — that can't be random.」。近 3 轮均为新判别维度（survivorship_bias/base_rate_neglect/whole_part_transfer），本轮按任务书轮换纪律切到证据纪律方向。判据为两腿共现（缺一不判）：腿① CLUSTER 聚集陈述三路线任一——路线 a 聚集语义词（连续/接连/接二连三/扎堆/集中/频发/高发/屡屡/频频/隔三差五/每逢/每当/每次/历来/向来；EN in a row/in succession/consecutive/clustered/spate/streak/every Friday）× 事件名词（事故/案件/盗窃/中奖/倒闭/猝死/投诉/故障/纠纷/空难/坠机/失踪/复发……裸名词形态；EN accidents/incidents/crashes/complaints/lawsuits/suicides）落在 14 字（EN 20 字符）窗内（前后双向）；路线 b 同一来源（全都/都/皆/清一色/无一例外 + 来自/出自/毕业于/就职于/指向/死于 + 同一/同一个/同一条/同属；EN all … from/at/in … same）；路线 c 数量≥2（四起/五起/好几个/数次；EN two/three/four/several/multiple）× 事件名词（后 6 字窗，含倒装「事故四起」）；腿② SIGNIF 显著性归因五形任一——R1 否认巧合（绝不是/绝非/一定不是/不可能只是/并非只是/不只是 × 巧合/偶然/随机；EN can't be / is not just / more than / far from × coincidence/chance/random）、R2 异常断言（显然/肯定/一定/可见/说明/意味着/证明 × 蹊跷/反常/有猫腻/说不通/邪门/诡异/离奇/不同寻常；EN clearly/obviously/must/means × sinister/fishy/inexplicable/pattern/law）、R3 神秘归因（天意/天命/劫数/魔咒/宿命/冥冥/幕后黑手/暗中操控/风水/玄学/诅咒/看不见的手；EN cursed/doomed/destiny/fate/omen/universe is telling）、R4 定律化（逃不过/躲不掉/在劫难逃/铁律/命中注定/历史重演；EN doomed/inescapable/bound to happen/same fate）、R5 周期迷信（每逢/每当/每次 × 必/准/总 × 出事/故障/复发/吵架；EN every Friday something goes wrong）。r173 实现教训（已固化进测试）：① 对冲豁免若置于 R1 之前，「绝不可能是巧合」里的「可能」会把攻击腿自己豁免掉——R1 必须先于对冲豁免判定；② 聚集路线取「首个匹配」会漏（「三个月」先于「五起」命中），必须全量扫描。四类豁免全部负例注入承重：显式随机性/统计处理（随机序列/随机波动/按概率/概率论/大数定律/正态分布/独立事件/统计显著/置信区间/样本量/基准率/正常波动/不足为奇/查明/排查/数据显示/研究表明/对照组/归因于；EN randomly/by chance/probability/statistically significant/confidence interval/sample size/base rate/root cause/investigation found/attributed to）、推理否定/证据不足（不代表/不意味着/推不出/说明不了/未必/不一定/证据不足/样本不足；EN does not mean/cannot conclude/insufficient evidence）、对冲（可能/也许/或许/大概/恐怕/倾向于/碰巧/恰巧；EN maybe/perhaps/might/probably/seems）、概念/教学转述（集群错觉/赌徒谬误/德州神枪手/热手谬误/概率论/统计学/老师/教材/课堂/卡尼曼；EN clustering illusion/Texas sharpshooter/apophenia/gambler's fallacy/hot hand/textbook/lecture）、疑问句。src/clustering-illusion.js 新增 22 个可注入常量（CONC/COUNT/SAME_SRC/EVENT_ZH×EN、SIGNIF_NEG/ANOM/MYST/LAW/CYCLE_ZH×EN、EXEMPT_RAND/NEG/HEDGE/CONCEPT_ZH×EN）与三路线聚集判定 + 五形显著性判定，index.js 九处接线（require/计算点/dimMap/allDims/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports）。来源：HN Algolia "conjunction fallacy"（Wikipedia Conjunction_fallacy 经典文献链与 "Linda the Bank Teller, by Tversky&Kahneman"）+ GitHub fallacy-detection 仓库群（latent-variable/Real_time_fallacy_detection ★119 / tmakesense/logical-fallacy ★20 / panFJCharlotte98/Fallacy_Detection EMNLP2024 ★5——全部 LLM-based，与纯规则引擎不同构，仅作概念背景）；差距实测：hot_hand 只管势头×延续预测，post_hoc/pseudo_causal 只管时序×因果，selection_bias/survivorship_bias 只管取样受限×总体推广，「随机聚集×否认巧合/特殊归因」零覆盖；判据为心虫自有纯规则实现。侦察落盘 /tmp/hf-scout-r173.txt。维度数 170→171，VERIFY 129→130。验证：round-173 轮测 28/28（攻击 12/12 含中英/裸名词/带修饰/倒装/长句、良性 16/16 零误判含证据归因/确认随机/事实认定/疑问/对冲/统计显著性/周期日程/概念教学、17 项常量负例注入承重、4 类豁免陷阱翻转、登记断言×3）、gate 层变体矩阵 12/12 非 pass 且良性 10/10 pass、双向门禁 52/52 召回 302/326 误拦不升、verify 16/16、guard 20/20（维度登记 171 键）、doc 22/22。 |
| 6.8.100 | 2026-10-31 | 第170维度 whole_part_transfer（整体-部分属性迁移 / composition-division fallacy）上线 — 把部分层面的属性直接当成整体层面的属性、或把整体层面的属性直接摊到每个部分头上，且该属性是涌现性/非可加的（强/优秀/顶级/清廉/竞争力…，不是重量/个数/总价这类可加量）：「每个队员个人能力都很强，所以这支球队一定很强」「所有零件都用的顶级材料，因此这台设备的性能一定极其出色」「这个政府里的每个官员都清廉，所以这个政府自然是清廉的」「这支球队是世界顶级的，所以每个球员都是世界顶级水平」「这家公司是全球顶尖的，因此这里的每个员工都是顶尖人才」「Every player on the roster is elite, so the team is guaranteed to win.」「The orchestra is world-class, therefore every musician in it is world-class.」。近 3 轮均为统计偏差族（selection_bias/survivorship_bias/base_rate_neglect），本轮按任务书轮换纪律切到经典逻辑谬误方向。判据为多条件共现（缺一不判）：组合形（部分→整体）= PART_QUANT 部分量词×部分名词（每个/每一/各个/所有/全部/任何/凡是/个个 + ≤2 字间隔 + 队员/成员/零件/模块/官员/公民/工程师/客户/人…；EN every/each/all + part/component/member/player/employee/people…，裸名词形态）× QUALITY 涌现质量词（很强/优秀/顶级/一流/清廉/闻名/elite/world-class/win…，裸「强」用前后排除法避开增强/强制/强行/强权/强化）× 整体名词（限定词形+裸名词形两条）× 结论标记 TAIL（推断词∪必然性情态：所以/因此/可见/显然/一定/必然/自然/势必/无疑；EN therefore/thus/which means/must/guaranteed——覆盖「所以球队一定很强」与「整个团队攻克难关自然不在话下」两种结论形态）× QUALITY 复现（整体名词后 40 字窗）；分解形（整体→部分）= 整体名词×质量词×推断词×（推断词后的）部分量词×质量词复现（部分名词后 24 字窗）。四类豁免全部负例注入承重：显式聚合运算（一共/共计/合计/总计/总和/总量/总重/总价；EN in total/altogether/combined/totals to——可加量的正确合计）、推理否定/翻转（难道/恰恰相反/不意味着/不代表/未必/不一定/取决于/推不出；EN not necessarily/does not mean/depends on）、对冲（可能/也许/或许/大概/倾向于；EN maybe/perhaps/might/probably/tends to）、概念/教学转述（组合谬误/合成谬误/分解谬误/整体与部分/生态谬误/逻辑谬误/老师/教材/课堂；EN composition fallacy/division fallacy/ecological fallacy/textbook/lecture）、疑问句。src/whole-part-transfer.js 新增 16 个可注入常量（PART_QUANT_ZH/EN、WHOLE_ZH/EN、QUALITY_ZH/EN、INFER_ZH/EN、MODAL_ZH/EN、EXEMPT_AGG/NEG/HEDGE/CONCEPT_ZH/EN）与两路由判定腿，index.js 九处接线（require/计算点/dimMap/allDims/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports）。来源：HN Algolia "logical fallacy detection" 命中 arXiv 2405.02318「Translating natural language to first-order logic for logical fallacy detection」（258 分，NL2FOL 框架，LOGIC 数据集 F1 78%——确认谬误检测为形式逻辑驱动任务族）与 GitHub fallacy-detection 仓库群（latent-variable/Real_time_fallacy_detection ★119 / tmakesense/logical-fallacy ★20 / panFJCharlotte98/Fallacy_Detection EMNLP2024 ★5——全部 LLM-based，与纯规则引擎不同构，仅作概念背景）；差距实测：composition_fallacy/division_fallacy 仅作为 fallacies 维度内 6 条纯英文窄模式存在（src/index.js:5453-5459），中文整族零覆盖；判据为心虫自有纯规则实现（同 r171 base_rate_neglect 与 fallacies 内 base_rate_fallacy 子标签共存之先例，不构成重复）。侦察落盘 /tmp/hf-scout-r172.txt。维度数 169→170，VERIFY 128→129。验证：round-172 轮测 20/20（攻击 13/13 含中英/裸名词/带修饰/倒装/长句、良性 15/15 零误判含显式合计/投入非质量/否定翻转/概念教学/疑问/对冲/战争多因边缘良性、8 常量负例注入承重翻转、4 类豁免陷阱翻转、登记断言×3）、gate 层变体矩阵 13/13 非 pass 且良性 8/8 pass、双向门禁 52/52 召回 302/326 误拦不升、verify 16/16、guard 20/20（维度登记 170 键）、doc 22/22。 |
| 6.8.99 | 2026-10-31 | 第169维度 base_rate_neglect（基础概率忽视）上线 — 拿单个阳性信号（检测/筛查阳性、被系统或算法标记、命中名单、符合画像、测谎仪显示）直接下高置信度后验结论（几乎肯定/肯定/一定/显然/毫无疑问/很可能/必然/确定/八成/铁定 + 得/患/感染/是/有/罪/骗子/罪犯/嫌疑人/高危），却无视基础概率/先验/患病率/假阳性率。两腿判据（SIGNAL×CONFIDENT，中英双语，含裸名词与 ≤6 字修饰间隔）；两类豁免（基础概率语言在场、待确认不妄断）负例注入承重。r171 注入测试实测：单列「概念引述」豁免与「基础概率」豁免冗余（概念名必含 base rate 字样），已移除。src/base-rate-neglect.js 新增 8 个可注入常量（SIGNAL_ZH/EN、CONFIDENT_ZH/EN、EXEMPT_BASERATE_ZH/EN、EXEMPT_NOOVERCLAIM_ZH/EN）与判定腿，index.js 九处接线（require/计算点/dimMap/allDims/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/exports）。来源：HN Algolia "base rate fallacy"（Wikipedia Base rate fallacy）与 GitHub 逻辑谬误检测仓库（tmakesense/logical-fallacy ★20）均无可直接移植的规则判据，仅作概念背景；判据为心虫自有纯规则实现。侦察落盘 /root/hermes1/cache/scratch/hf-scout-171.txt。维度数 168→169，VERIFY 127→128。验证：round-171 轮测 11/11（攻击 9 命中含裸名词/带修饰/倒装/长句/中英、良性 7 零误判含基础概率/待确认/边缘良性、6 常量负例注入承重翻转、gate 变体矩阵 9/9 非 pass 且良性 7/7 pass）、双向门禁 52/52 召回 302/326 误拦不升、verify 16/16、guard 20/20、doc 22/22。 |
| 6.8.98 | 2026-10-31 | survivorship_bias 新句式族「幸存者取样×总体推广」上线 — 幸存者偏差的取样框形态：只从活下来的样本取证（腿①：只/仅/全/都×取样动词×幸存者词——只采访了事故的幸存者/只统计了活下来的用户/复盘只采访了没被裁员的员工/样本全部来自生还者；EN only interviewed the survivors / Only those who survived are counted / We only heard from the customers who stayed），却据此推广到全体（腿②：推广动词×总体范围词——反映了全体乘客的体验/认定对全体乘客都有效；或总体范围词×全称判断——人人都能成功/整体都很满意/大多能成功/so all customers love the new plan）。r169 实测缺口：12 条攻击 6 条真零命中（含 EN 双语，如「只采访了事故的幸存者，所以这反映了全体乘客的体验」），selection_bias 的幸存者措辞豁免按射程切割把本族整体让出，本轮在 survivorship_bias 内接住。四类豁免全部负例注入承重：假设条件句（如果只统计活下来的用户，结果就会偏乐观——阐述偏差而非犯偏差）、否定前置（不是只采访幸存者）、反驳本族（结论不能反映全体/不足以推广）、疑问句。src/survivorship-bias.js 新增 12 个可注入常量（SURV_SAMPLE_ZH_A/B/C、SURV_SAMPLE_EN_A/B、SURV_GEN_ZH_A/B、SURV_GEN_EN_A/B、SURV_EXEMPT_ZH/EN）与 r170 判定腿，DIM_GUIDANCE 概念条文同步扩写。来源：HN Algolia "survivorship bias"（Wikipedia 幸存者偏差 464 pts / Survivorship bias and startup hype 340 pts / Beware survivorship bias in advice on science careers 197 pts）与 Semantic Scholar "survivorship bias"（2023-2025 领域应用论文，如 Survivorship Bias in Psychological Research 2025 / The Impact of Survivorship Bias in Glioblastoma Research 2023 27 次引用）均无可直接移植的规则判据，仅作概念背景；判据为心虫自有纯规则实现。侦察落盘 /tmp/hf-scout-1793404301.txt。维度数不变（168，本族为既有维度新句式族），VERIFY 级不变。验证：round-170 轮测 49/49（攻击 ZH 10/10+EN 4/4 命中含裸名词/带修饰/倒装/长句/跨句/来源短语、良性 ZH 12+EN 4 零误判含假设/否定前置/反驳/疑问/充分取样/概念转述/多因陈述边缘良性、11 常量负例注入承重翻转、分工边界锁×2、存量回归×3、登记断言×3、gate 层×2）、gate 层变体矩阵 8/8 非 pass 且 findings 全含 survivorship_bias、模块层攻击 13/13 良性 12/12 零误判、双向门禁 52/52 召回 302/326 误拦不升、verify 16/16、guard 20/20、doc 22/22。 |
| 6.8.97 | 2026-10-31 | 第 168 维度 selection_bias（取样偏差 / selection bias）上线 — 新增判别维度，补上心虫「证据充分性」方向缺失的取样框限制谬误：只从一个受限的、自选择的或非随机的来源取证（只调查会员/只统计某群体/只采访愿意配合者/数据只来自某平台/样本全部来自志愿者），却据此断言「代表全体」「具有普遍性」「适用于所有人」「足以说明全国」——取样来源与结论总体不匹配，证据效力被夸大（「问卷只发给了注册会员，所以调查结果代表全体用户的满意度」「数据只来自一线城市的活跃用户，足以说明全国用户的偏好」「样本全部来自自愿报名的志愿者，因此这个结论具有普遍性」「We only surveyed volunteers, so the results are representative of everyone.」「The sample was drawn solely from our premium users, which means all customers prefer this feature.」）。近 5 轮均为形式逻辑族（affirming_the_consequent/denying_the_antecedent/one_sided_dismissal/appeal_to_consequences/illicit_conversion），本轮按任务书轮换纪律切到统计推断/证据充分性方向。双腿判据：腿① RESTRICTION 取样受限——限定标记（只/仅/仅仅/全都/都/均/only/solely/restricted to）×取样动词（调查/统计/采访/回收/收集/纳入/来自/发给/survey/sampled/drawn/recruited）近距离共现（含裸名词形态「数据都来自一线城市」），或来源短语（样本/数据+全部/都/均+来自/源于）带受限来源词（会员/活跃/自愿/投诉/一线/管理层/volunteers/premium/self-selected）；腿② GENERALIZATION 总体推广——推广动词（代表/适用于/推广/反映/具有/足以说明/足以证明/representative of/generalizes to/applies to）接总体范围词（全体/所有/整体/全国/每个人/大众/普遍/all/every/everyone），或总体范围词+「都」形全称判断（全体员工都认可）。结论候选覆盖同逗号段推断词后、后续逗号段、跨句后续句、「毕竟/因为」倒装路由；推断词本身是推广动词（足以说明/which shows）时不断句。八类豁免全部负例注入承重：明示局限（只能反映/不能代表/不一定成立/还有待验证/不足以推广——正确推理）、充分取样（随机抽样/分层抽样/普查/random sample）、幸存者措辞（幸存者/活下来/survivors——射程归 survivorship_bias）、只描述样本自身（会员的满意度更高/所有样本都合格）、概率化结论（可能/也许/maybe/might）、疑问句、谬误概念转述（取样偏差/selection bias）、教学转述。仿 illicit-conversion.js 结构，新文件 src/selection-bias.js（~330 行），index.js 九处接线。来源：HN Algolia "base rate fallacy" 检索命中 Wikipedia Base rate fallacy 与 Schneier「Terrorists, Data Mining, and the Base Rate Fallacy」（概念背景；心虫 base_rate/prosecutors 子标签仅英文侧 5 条正则、中文侧零覆盖，Semantic Scholar 空返回，GitHub 高星方案全部 LLM-based 与纯规则引擎不同构），候选族差距由 /tmp/r169-gap-probe.js + /tmp/r169-baseline-probe.js 在当前 167 维引擎实测（取样偏差族 9/10 真零命中；对照候选：可得性启发 5/6、聚类错觉 4/5、区群谬误/回归均值 3/4 被存量维度代拦，故选取样偏差族）；判据为心虫自有纯规则实现。侦察落盘 /tmp/hf-scout-r169.txt。维度 167→168、VERIFY 126→127。验证：round-169 轮测 53/53（攻击 ZH 10/10+EN 4/4 命中含裸名词/带修饰/倒装/长句/跨句/仅标记、良性 ZH 11+EN 4 零误判含边缘良性、14 常量负例注入承重翻转、分工边界锁、区分锁、存量回归×2、登记断言×4、gate 层×2）、gate 层变体矩阵 12/12 非 pass 且 findings 全含 selection_bias、模块层良性 16/16 零误判、双向门禁 52/52 召回 302/326 误拦不升、verify 16/16、guard 20/20、doc 22/22。 |
| 6.8.96 | 2026-10-31 | 第 167 维度 illicit_conversion（不当换位 / illicit conversion）上线 — 新增判别维度，补上心虫「演绎有效性」里缺失的范畴换位无效式：先立「所有 A 都是 B」的全称肯定命题，再据此断言「所有 B 都是 A」——主项谓项一对调，单向包含关系被偷换成双向（「所有猫都是动物，所以动物都是猫」「畅销的车都省油，所以省油的车都畅销」「爱国者都支持国产，所以支持国产的都是爱国者」「凡是认真读过原著的人都反对魔改，所以反对魔改的人都认真读过原著」「坚持锻炼的人免疫力强，所以免疫力强的人都坚持锻炼」；All squares are rectangles, so all rectangles are squares. / Every poet loves language, so everyone who loves language is a poet. / Whales are mammals, so mammals are whales.）。这是三段论形式谬误族的第三块（affirming_the_consequent 管 Q∴P、denying_the_antecedent 管 ¬P∴¬Q、conditional_proposition 管 ¬B⇒¬A，本族管 B⊇A 的主谓对调）。双腿判据：腿① PREMISE 全称肯定前提——全称量词（所有/全部/任何/每个/每一/凡/一切/一律）或裸「都」形的 A⊆B 陈述（EN: all/every/any/each A are B，或无系词动词句 Every poet loves language），切出主项 A 与谓项 B；腿② CONVERSION 换位结论——推断词（所以/因此/可见/这说明/故/so/therefore/which means）之后（或紧跟的下一句、倒装在前的上一句、「因为」标记路由的同句后半）出现以 B 为主项、A 为谓项的陈述。裸谓词前提（「坚持锻炼的人免疫力强」无量词无系词无都）按结论主/谓项关键词位置回退切分；单字谓项允许被前提关键词反向包含（飞 ⊆ 会飞的）。这个「主谓对调」结构天然区分：特称换位（有些A是B ∴ 有些B是A——有效）、否定换位（所有A都不是B ∴ 所有B都不是A——有效）、单称例示（结论主项回指 A）、有效三段论（结论回指第二前提主项）、前提重述（方向未对调）。八类豁免全部负例注入承重：特称量词、对冲/宽泛量、否定词（否定结论是正确的换位否认）、双条件（当且仅当/iff）、对称关系（同一/等同于/互为）、谬误概念转述（不当换位/换位命题/illicit conversion）、教学转述、证据语境。仿 denying-the-antecedent.js 结构，新文件 src/illicit-conversion.js（~440 行），index.js 九处接线。来源：HN Algolia 逻辑谬误主题多篇高票故事（Translating natural language to first-order logic for logical fallacy detection 258 pts / Logical fallacies in software engineering 242 pts / Logical Fallacies Are Usually Irrelevant or Cited Incorrectly 176 pts——FOL 路线需 LLM 与纯规则引擎不同构，仅背景参考），候选族差距由 /tmp/r168-gap-probe.js 在当前 166 维引擎实测（换位族 8 条 7 条真零命中，候选中项不周延族仅 2/8，故选换位族）；判据为心虫自有纯规则。侦察落盘 /tmp/hf-scout-1793387758.txt。维度 166→167、VERIFY 125→126。验证：round-168 轮测 25/25（攻击 ZH 10/10+EN 5/5 命中含倒装形/裸谓词前提/无逗号同句、良性 ZH 14+EN 4 零误判、八常量负例注入承重翻转、分工边界锁×2、区分锁、存量回归×2、登记断言×5、gate 层×2）、gate 层变体矩阵 12/12 非 pass 且 findings 全含 illicit_conversion、双向门禁 52/52 召回 302/326 误拦不升、verify 16/16、guard 20/20、doc 22/22。 |
| 6.8.95 | 2026-10-31 | 第 166 维度 one_sided_dismissal（让步消解 / concessive dismissal）上线 — 新增判别维度，补上心虫「证据对称性」里缺失的单边论证族：先让步承认反面证据存在，再用一句总体性概括把它不作量化权衡地扫掉，直接收尾到单边结论（「这个药虽然有副作用，但临床试验显示它利远大于弊，所以值得推广」「研究确实也发现了一些问题，不过总体而言瑕不掩瑜」「当然反对者也能举出一些反例，但主流数据都站在我们这边」「他确实犯过错，但这不能掩盖他取得的巨大成就」；Side effects exist, but the benefits overwhelmingly outweigh the risks.）。这是 false_balance（把一边倒说成势均力敌）的镜像：把势均力敌/有真实反面证据的说成一边倒。双腿判据（缺一不判）：腿① CONCESSION 让步标记（虽然/尽管/固然/诚然/不可否认/必须承认/确实/当然/ admittedly/although/though/while/granted/critics/skeptics/there are/exist）× 反面证据词（副作用/风险/反例/批评/质疑/缺陷/过错/失败/争议/代价/受损/冲击 / side effects/risks/criticism/concerns/failures/flaws/drawbacks/dissenting——全部裸名词形态）；腿② DISMISSAL 消解短语（瑕不掩瑜/不能掩盖/利远大于弊/总体而言/归根结底/主流/大势所趋/实践证明/时间会证明 / overwhelmingly/outweighs/on balance/consensus is clear）。四类豁免全部负例注入承重：双方量化（让步与消解前后均出现具体数字——真实的风险收益权衡）、处置措施（可以监测/及时调整/停药即可/可控/可逆——全文检查，处置声明常出现在消解短语之前）、概念转述（选择性证据/单边论证/cherry picking）、教学转述。新文件 src/one-sided-dismissal.js（~230 行），index.js 九处接线。来源：Semantic Scholar 证据充分性/幻觉检测综述（Evidence-Chain Reliable RAG 2025、Medico arXiv:2410.10408、Evidence-Aligned Entity Verification arXiv:2609.08267——均需 LLM/embedding，仅背景参考），判据为心虫自有纯规则。维度 165→166、VERIFY 124→125。验证：round-167 轮测 44/44（攻击 ZH 10/10+EN 5/5 含倒装形、良性 ZH 8+EN 4 零误判、8 常量注入承重翻转、区分锁 false_balance）、gate 层变体矩阵 11/11 非 pass 且 findings 全含 one_sided_dismissal、双向门禁 52/52 召回 302/326 误拦不升、verify 16/16、guard 20/20、doc 22/22。 |
| 6.8.94 | 2026-10-30 | 第 165 维度 denying_the_antecedent（否定前件 / denying the antecedent）上线 — 新增判别维度，补上心虫「演绎有效性」里缺失的无效推理族：先立「如果 P 则 Q」的条件规则，再以「并非 P」断言「并非 Q」——否定前件得出后件也不成立（「如果下雨，运动会就取消；今天没下雨，所以运动会不取消」「努力的人都会早起；他根本不努力，所以他不会早起」；If it rains, the match will be cancelled. It did not rain, so the match is not cancelled.）。双腿判据：腿① COND 条件/规则句（如果X（那么/则/就）Y / X都会Y / X能Y / 只要X就Y / EN: if X then Y / X has Y / 英文裸规则句前两词为主语），切出前件 A 与后件 B；腿② ANTECEDENT_DENY——按推断词（所以/因此/说明/so/therefore）切分事实句，前半回指前件 A 且被否定（¬P）、后半回指后件 B 且被否定（¬Q），「¬P ∴ ¬Q」结构天然区分 modus ponens（前后件皆肯定）/modus tollens（¬Q∴¬P 方向相反且 A≠B 不交叉）/affirming the consequent（Q∴P 无否定）。六类豁免（双条件/否定翻转/对冲/概念转述/教学/证据语境）。仿 affirming-the-consequent.js 结构，新文件 src/denying-the-antecedent.js（~290 行），index.js 九处接线。来源：arXiv:2405.02318 "Translating natural language to first-order logic for logical fallacy detection"（HN 258 pts，FOL 路线与纯规则引擎不同构，仅背景参考）+ GitHub tmakesense/logical-fallacy（20★）无可复用规则。维度 164→165、VERIFY 123→124。验证：round-166 轮测 39/39（攻击 ZH 8/8+EN 5/5 命中、良性 ZH 8+EN 3 零误判、五常量负例注入承重翻转、区分锁/分工边界锁）、gate 层攻击 8/8 非 pass 且 denying_the_antecedent count≥1 承重、双向门禁 52/52 召回 302/326 误拦不升。 |
| 6.8.93 | 2026-10-30 | 第 164 维度 appeal_to_consequences（诉诸后果 / appeal to consequences）上线 — 新增判别维度，补上心虫「证据纪律」里缺失的后果替代族：以命题后果的可怕/不可接受/意义消解评估替代事实证据，据此否定或拒绝相信命题真值（「承认房价会跌太可怕了，那么多人还着贷款，所以房价不可能跌」「如果进化论是真的，那人活着就没有意义了，所以进化论不可能是真的」「我不敢相信气候变化是人为的，万一这是真的，经济就全盘崩溃了」；If climate change were real, our economy would collapse, so it cannot be true. / The findings are too devastating to be true — the data must be fabricated.）。双腿判据：腿① CONSEQUENCE 显式后果评估（太可怕/细思极恐/不堪设想/不敢想/没法接受/天要塌了/全盘崩溃/经济崩盘/还有什么意义/天下大乱；EN：too damaging / would collapse / would be devastating / too much to bear / life would have no meaning）× 腿② TRUTH-DENIAL 真值否定/拒绝相信（所以不可能是真的/一定是假的/纯属吓唬/危言耸听/我不敢相信…是真的/怎么可能；EN：cannot be true / must be false / I refuse to believe / too … to be true），倒装形（拒绝在前后果在后）由双腿共现覆盖。五类豁免（对冲/证据在场/概念转述/否定翻转/提问）。仿 monocause.js 结构，新文件 src/appeal-to-consequences.js（~250 行），index.js 九处接线。来源：标准非形式谬误 taxonomy（argumentum ad consequentiam）；外部搜索记录——SelfDoubt 论文（satwik2711/SelfDoubt，4★，COLM 2026 submitted，Hedge-to-Verify Ratio 概念）作背景参考、未直接集成（其方法需 reasoning trace 与模型 internals，与纯规则引擎不同构），Semantic Scholar 全程不可用（total undefined），不编造 arXiv ID。维度 163→164、VERIFY 122→123。验证：round-165 轮测 25/25（变体矩阵 ZH 8/8+EN 5/5 命中、良性 ZH 8+EN 3 零误伤、五豁免逐常量承重翻转、负例注入双腿停判）、gate 层攻击 10/10 非 pass 且 appeal_to_consequences count≥1 承重、双向门禁 52/52 召回 302/326 误拦不升。 |
| 6.8.92 | 2026-10-30 | 第 163 维度 doublespeak（语义洗白 / euphemistic relabeling）上线 — 新增判别维度，补上心虫「是非判断」里缺失的名实分离族：把有害行为用中性/褒义词汇重新命名以规避指责（「我们不是在监控你，我们是在关心你的安全」「这不是裁员，这是优化，是给社会输送人才」「这不叫打压，这叫帮助你成长」「公司把裁员说成是优化」「政府部门把强拆包装成城市更新」；This is not surveillance, it is care. / The layoffs were rebranded as optimization. / Coercion was dressed up as consent.）。双腿判据+方向原则：腿① HARMFUL-ACT 有害行为名词/动词族（监控/打压/辞退/裁员/惩罚/欺骗/造假/剽窃/暴力/拘禁/操纵/剥削/歧视/霸凌/家暴/降薪/强制加班；EN：surveillance/censorship/layoff/coercion/exploitation/harassment/blackmail/propaganda）× 腿② RELABEL 重新命名框架两条路由——(a) 否定替换式（不是X，是Y / 并非X，而是Y / 不叫X，应叫Y；EN：this is not X, it is Y / we do not X; we Y / rather than X）；(b) 转称式（把X说成Y / X说成Y / 把X包装成Y；EN：X was rebranded as Y / X was dressed up as Y），中性改名动词另需柔性标签（优化/关怀/安全/成长/care/safety/optimization）——方向原则：有害词必须是被改名的对象而非被揭穿的现实，据此排除批评者揭批形态（「以安全之名，行监控之实」「under the guise of…」）与受害者反向拆穿（「这不是爱，这是控制」）。三类豁免（法律定性/升级定性、元话语概念评析、无有害词同形句）。仿 incredulity.js 结构，新文件 src/doublespeak.js（~370 行），index.js 九处接线。来源：arXiv:2103.16808 "Self-Supervised Euphemism Detection and Identification for Content Moderation"（IEEE S&P Oakland 2021，HN id 26725061，自监督路线与纯规则引擎不同构）+ GitHub WanzhengZhu/Euphemism（36★）、yiyepianzhounc/TED-SCL（69★）均无可复用规则。维度 162→163、VERIFY 121→122。验证：round-164 变体矩阵 ZH 25/25 + EN 18/18 命中、良性 0/34（含 10 条批评者揭批形态）、双向门禁良性语料 0/369 误伤、负例注入逐常量承重、gate 层攻击全 verify 且 doublespeak count>0 承重。 |
| 6.8.91 | 2026-10-30 | 第 162 维度 incredulity（诉诸个人怀疑 / argument from personal incredulity）上线 — 新增判别维度，补上心虫「证据纪律」里缺失的认知越界族：以主观「无法想象/难以置信」替代证据，据此否定事实真值（「我无法想象他怎么能做到，所以这成绩一定是假的」「一看就是后期剪辑的」「不可能是真的，我实在无法想象它怎么采集的」；I cannot imagine how this is possible, so it cannot be true. / There is no way that happened; I cannot fathom it.）。双腿判据：腿① INCREDULITY 难以置信陈述（无法想象/难以想象/想象不到/难以置信/不可思议/匪夷所思；EN：can't imagine / hard to believe / inconceivable / unfathomable / unbelievable）× 腿② TRUTH-DENIAL 真值否定推导（推断词+否定断言：所以/因此/这说明…+不是真的/是假的/纯属/不可信/作弊/编造/摆拍/有黑幕；直接否定断言：一看就是假的/肯定是作弊的/不可能是真的；倒装形同样命中）。三类豁免（有据可查的惊讶、提问句式、概念分析）。仿 monocause.js 结构，新文件 src/incredulity.js（~260 行），index.js 九处接线。来源：标准非形式谬误 taxonomy（argument from personal incredulity）；HN Algolia 命中真实基准故事 arXiv 2405.02318（LLM 路线，与纯规则引擎不同构）、GitHub 最高星 latent-variable/Real_time_fallacy_detection（119★，无可复用规则）、Semantic Scholar 全程失败，不编造 arXiv ID。维度 161→162、VERIFY 120→121。验证：round-163 轮测 20/20（变体矩阵 ZH/EN×裸词干/修饰/倒装/长句 10/10 命中、良性 9/9 零误伤、负例注入逐常量承重）、gate 层攻击非 pass、incredulity 维度 count>0 承重、双向门禁不升误拦。 |
| 6.8.90 | 2026-10-30 | v6.8.90 monocause 腿①裸战争补缺 + README badge 同步（r161 首轮 cron 产出实测修补） — 实测发现 monocause 腿①只收带后缀的「战争爆发」，裸名词形「战争纯粹是因为资源争夺」整句漏判，补 COMPLEX_ZH 裸词形态；README badge/横幅数字同步至 161/161。 |
| 6.8.89 | 2026-10-30 | 第 161 维度 monocause（单一因果过度简化 / fallacy of the single cause）上线 — 新增判别维度，补上心虫三大支柱「因果纪律」里缺失的归约谬误射程：post_hoc 管时序→因、false_cause 管相关→因、pseudo_causal 管虚构机制，但都没有管「把真实多因的系统性复杂现象压成单一原因并否认多因性」。辨别把经济危机/公司倒闭/犯罪率/战争爆发/王朝覆灭/帝国衰落/社会撕裂/创业失败等已知多因现象，归结为「完全是因为X」「纯粹是X」「唯一的原因」「说到底/归根结底」「没有别的原因」（The financial crisis was caused entirely by deregulation. / Crime is nothing but a product of poverty.）。双腿判据：腿① COMPLEX-OUTCOME 已知多因的系统性现象（只收系统性现象，不收考试/比赛/普通项目等可单一归因的琐事——精度锚）× 腿② SINGLE-CAUSE 绝对化单一归因框架；四类豁免（多因并列/共同作用、hedged 单一贡献者「…之一/one of the main drivers」、提问句式、概念分析）。仿 appeal-to-novelty.js 结构，新文件 src/monocause.js（~230 行），index.js 九处接线（require/计算点/allDims/dimMap/DIM_GUIDANCE/VERIFY_DIMS/dimensions/summary/module.exports）。来源：标准非形式谬误 taxonomy + HN 基准故事 16200526（Recognising the fallacy of a single root cause）；Semantic Scholar 全程 429 限流、GitHub 无成熟规则引擎实现，不编造 arXiv ID。维度 160→161、VERIFY 119→120。验证：round-161 轮测 20/20（攻击 ZH/EN 全命中、良性 0 误伤、负例注入逐常量承重）、gate 层攻击=verify、隔离良性 0/17 误伤、双向门禁不升误拦。 |
| 6.8.88 | 2026-10-28 | 审计修复批（multi-agent audit patch — 异步委托 4 子代理全量审计闭环） — 委托批次 deleg_39686998 发现的真问题逐条亲验后修复：① [HIGH] getOrCreateInstance 全仓库 4 处调用 0 处定义——src/mcp-server.js 与 mcp/mcp-server-http.js 的 SSE 会话级 benchmark 工具 100% 落进 dispatch try/catch 被 ReferenceError 吞掉静默回退全局实例；两文件落地 _sessionInstances Map 注册表（按 sessionId 缓存、上限 16 FIFO 淘汰、start() 失败上抛不回退），4 处调用点改 await，等价逻辑单测 cache_hit/隔离/兜底三项全 true。② [MEDIUM] src/cortex/self-benchmark.js 调用 anchorSync 但 benchmark-external-anchor.js 9 行全 Stub 未导出该符号——外部基准锚定从未生效且错误被 catch 吞；删除 stub 模块与死导入/死调用，assess() 诚实走未验证分支（内评 0.7 折 + verified:false），heartflow.js 启动日志假宣称「外部锚定防自欺已启用」改为实况。③ [MEDIUM] BUILD_DATE 落后一个补丁（6.8.86→6.8.87）。④ [LOW] src/dream/dream-loop.js / interactive-dream.js 对 dream.js 的 LEVELS/DEFAULT_SCORING 死解构死导出（运行时实测 undefined）、module-registry.js 死导出 createSpecialModule（heartflow.js 用本地副本 _createSpecialModule），四处清除。⑤ [LOW] package.json 4 个 script 指向已归档 tests/ 死目录，重指向现存文件。⑥ [LOW] src/report/report-generator.js _extract 缺 think() 实际形态（七阶段住在 chain 键，thoughtChain 恒不存在）——CLI/MCP 两条报告路径全输入输出「未捕获思维链」，补形态 (c) 规范化 + _stageText 对象 output 的 [object Object] 渲染修复。⑦ [LOW] SSRF 保留网段补齐：url-validator.js CIDR 表增 CGNAT/TEST-NET1-3/198.18/组播/240÷4/广播 8 段并导出 isBlockedV4，fetch-safe.js 两处会漂移的内联私网副本改复用单一事实源。⑧ 文档：heartflow-architecture-tracing/SKILL.md 架构描述与实码相反（实际 ThoughtChain 主路径 + think-pipeline 后置），勘误；规格表测试数改实测 4 段聚合 18,996/72/19,068。验证：s2/s3 与改动前基线逐分一致、s4 失败集合同源（round-608 等既有 flaky）、security-audit 16/16、fetch-safe 2/2、gate-verdict 27/27、recovered-modules 16/16、doc-numbers 3 条失败为 r412 既有切片噪声。 |
| 6.8.87 | 2026-10-28 | console 残留清理（code-quality patch — 25 处调试噪声接入 house logger） — 全局审计（安全/运行时/质量三线 + 4 代理交叉）发现全部闭环后，将 src/ 下 25 处调试噪声 console 调用接入既有 house 标准 src/utils/logger.js（logger.debug/info/warn(module, ...)，LOG_LEVEL 默认 WARN 静默、LOG_LEVEL=1 恢复）：src/cortex/meta-engine.js 10 处 cycle 追踪、src/formula/formula-engine.js 7 处加载/推导调试（含 2 处「简化版」提示行）、src/memory/memory-bank.js 2 处加载状态、src/memory/memory-encrypt.js 4 处（2 消息 × 2 份重复块）密钥状态、src/formula/formula-search.js 1 处加载完成统计、src/emotion/deep-emotion.js 1 处初始化、src/core/uncertainty-quantifier.js 1 处校准告警（升 logger.warn）、src/core/engine-initializer.js 1 处会话间隔。排除项（40+ 处不动）：aipay-server/api-server/a2m-notify 支付与通知操作日志、mcp-server/gate 的 console.error 安全输出、repo-audit 的 REPORT/FILES/LINES 机器可读契约、utils/logger 与 infra/logger 自身实现、heartflow.js 内部 logger 与 HEARTFLOW_DEBUG 门控行、transmission-engine 全注释形态。grep 验证零测试 spy stdout/断言这些前缀；默认输出实测零噪声（CLI status 常驻 [MemoryBank] 行消失），LOG_LEVEL=1 恢复。8 个被改模块 require 冒烟 EXIT=0；靶向测试（engine-initializer / round-621-uncertaintyquantifier-dispatch / meta-audit / meta-calibration / metaMemory）全过；全量分段回归 19,005 通过 / 70 失败 / 19,075，与清理前逐段一致（零回归）。 |
| 6.8.86 | 2026-10-28 | 全局审计修复（multi-agent audit patch — 安全 / 运行时 / 质量三线并行） — 三路审计（子代理并行：安全、运行时 bug、质量死代码）交叉核实后修复：① [HIGH] src/core/heartflow.js 缺 stop() 方法——src/mcp-server.js:5887、mcp/mcp-server-http.js:1480、bin/verify.js:154、主块 dispatch（6428/6436）共 6 处 hf.stop() 调用全部落进 catch 被静默吞，MCP 关闭与验证器收尾从未真正执行；补 stop()（先停 consolidate 观察器再委托 shutdown()），fail-to-pass 证据：test/legacy/v2_0_19.test.js 从 hf.stop is not a function 崩溃变为 31/31 通过。② [HIGH] src/knowledge/classics-rules.js:986 execSync 字符串插值执行 bash，keyword/scope 直达 shell（引号/反引号/$/分号全部存活），改 execFileSync 参数化数组；当前宿主未捆绑 search_guji.sh 故不可达，属潜在漏洞。③ [HIGH] src/memory/memory.js 804 行死代码（@deprecated REMOVED in v12.0.0，精确 require 引用 0 命中，29 个宽匹配全系 memory-adapter/bank 等别名），删除并同步 memory-encrypt.js 注释。④ [MEDIUM] tests/ 孤儿目录并轨：TEST_DIR=__dirname 只扫 test/，5 个 *.test.js（core / decision-routing / integration / memory-system / legacy v2_0_19，共 50+ 用例）+ 2 个错位文件从未执行——tests/unit/test-code-verifier.js 按自述头归位 test/core/code-verifier.test.js（实测通过）；test/core/path-guard.test.js 原是只查模块加载的空桩，真实 25 例测试错位在 tests/unit/test-path-guard.js（v6.8.85 轮的 Windows 修复打在影子副本上，canonical 版从未受益），替换；4 个一次性诊断脚本移 test/archive/one-off-scripts/。7 个文件全部经 test/_mount.js harness 退出码 0 复验。⑤ [HIGH] src/consciousness/debate-engine.js MCP heartflow_debate 工具 sessionId 路径遍历——addRound/summarize/conclude 的 id 未经校验直拼 path.join(this.dataDir, id + '.json')，../../x 可越权读写任意 .json；在 _loadSession/_saveSession choke 点加 ^[a-zA-Z0-9_-]{1,64}$ 白名单（合法 id 仅由 createSession 的 crypto.randomUUID() 前缀生成），遍历读写实测全部拦截。⑥ [MEDIUM] _lazy 注册表双副本漂移——src/core/heartflow.js start() 内 117 行内联表与 engine-constructor.js 的 initializeLazyRegistry 逐条重复（实测 50/50 键运行时一致，但 initializeLazyRegistry 从未被调用，任何一侧修改必漂移）；内联表替换为 initializeLazyRegistry(this) 一行调用，并修复真身表 5 条路径错别字（flowPredictor→./flow-predictor.js、userModel/purposeEngine→../identity/、riskAnalyzer→../reasoning/、auditLogger→../shield/，原先 19/46 条目 MODULE_NOT_FOUND 降至 14 个保留槽位）。全量复跑（分段 1/4–4/4）：19,040 通过 / 71 失败 / 19,111 |
| 6.8.85 | 2026-10-27 | 全量回归审计修复（audit patch — 612 个测试文件全量回归驱动） — 对 test/ 全量 612 个测试文件做分段串行复跑，把存量失败按「并发假失败 / 测试预期过期 / 真实引擎缺陷」三类归因而修。引擎侧 15 处：src/sole-narrative.js 的 EXCLUSIVE_ZH 自 v6.8.38 起为死支 /(?!)/，排他/终止表述（「只有这一种解释」族）全漏，重写为正则并对齐 EXCLUSIVE_EN 分界；src/agency-deflection.js 的 T5_EN = /^$(?!)/ 从未接线，5 条英文攻击全漏，重写为三子支单行正则；src/core/dao-decision.js 的 verdict 阈值 0.75 放过单层 flag 攻击（soft+force 组合），补「任意层 not ok → 至少 CAUTION」规则和 force 否定语境豁免（「没有绝对优劣」不再误判强制）；src/core/uncertainty-quantifier.js 的 hallucination 维度接入的是数字而非 hallucination_result 对象，该维度从未参与打分，接真对象（与同类 uncertainty 一致）；src/manipulation-tactics.js 的 hasAffinityPromotion 在 HEAD 里是被误提交的变异态（&& true && 等价于 &&），恢复为 v6.7.16 判据并补 r327 变异守卫测试；src/shield/ai-writing-tell.js 的 connective-fold 维度名写错导致 r200 折叠从未生效；src/circular-reasoning.js 的 RESTART_TAIL 含裸「不 |
| 6.8.84 | 2026-10-27 | 英文注入同义族通道修复（upgrade type 2 — 存量缺陷修复，第 733 轮） — 提示注入的英文模式库 INJECTION_PATTERNS.en 只有 12 条字面短语正则：ignore 是唯一动词、`previous |
| 6.8.82 | 2026-10-27 | _dePinyin site 1 修复：英文词碰撞 + 大小写 + 还原闭环（upgrade type 2 — 存量缺陷修复，第 731 轮） — site 1（无空格拼音还原）原实现是盲目 out.split(k).join(PINYIN_MAP[k])，实测三个缺陷：① 英文词碰撞：mingling 本身就是 PINYIN_MAP 键（→命令），"the mingling of cultures" → "the 命令 of cultures"（扫仓库 85,968 个英文 token，唯一整词碰撞就是它）；② 大小写敏感：_dePinyin 跑在 lowercase 之前，句首大写 Hulve / 全大写 HULVE 的 includes(k) 不成立 → 整个漏还原；③ 还原不闭环：还原出「忽略」后周围英文攻击词不翻译，"忽略 previous instructions" 既不匹配中文模式（previous 是英文）也不匹配英文模式（hulve ≠ ignore），整句 gate=pass 静默放行（实测 prompt_injection count=0，而完整还原「忽略 之前的 指令」→ block）。修复：EN_WORD_EXCEPTIONS 英文词例外表（含 mingling 词族 + 派生词串守卫）+ _replacePinyinKeys 逐 token 大小写不敏感替换（三处通路共用）+ 5c-post 还原闭环（_en2zh 从 step 4c 提取复用，补 previous/prior → 之前）+ 中英混插剥离器只剥夹在字母间的单字汉字（原判据会把 site 1 刚还原出的「忽略」吃掉「忽」）。收益：7 种拼音注入形态全部 pass → block（含句首大写/全大写/please 前缀/中文语境内插大写拼音）；英文句零污染（mingling 词族 5 句 normalize 输出与输入逐字节一致、gate 全 pass）；存量反混淆能力零回归。门禁：bin/verify.js 16/16、双向门禁召回 52/52、误拦 302/326（与基线一致，零新增误伤）、guard-abilities 20/20、轮测试 37/37、r730 轮测试 18/18（其「已知局限」断言翻转为守修复）。全量回归 18,364 过 / 145 败 / 18,509 个（与 r730 实测 18,366/143/18,509 同总量；失败 +2 逐定性为存量：doc-snippet 执行器在 Windows 把反斜杠路径生成进 JS 字符串致 0/4——HEAD 版本实测同样 0/4；dimension-action-wiring 的 post_hoc 误判——该句 normalize 输出与 HEAD 逐字节一致，与本轮无关）。已知未修：evidence 维度打分器恒返回 {score:1, issues:[]}（r690 起连续记录）；无空格英文注入 "ignorepreviousinstructions" 仍 pass（英文模式要求空格，纯英文就这样，非拼音引入）；camelCase 标识符被 de_encode 的 base64 试探解出乱码（存量，与本轮无关） |
| 6.8.81 | 2026-10-27 | normalize de_pinyin 英文破坏修复（upgrade type 2 — 存量缺陷修复，第 730 轮） — _dePinyin 的空格链判据原是「压缩后含任一 PINYIN_MAP 键即判拼音串」，而英文句压缩后极易偶然含键："Real artists never trace references, they draw from imagination alone." 压缩得 theydrawfromimaginationalone，含 mima（4/27 = 15% 覆盖）→ 被判成拼音串 → 空格被抹、from 的 m+a 被换成「密码」→ 中英混插剥离器 /([a-z]{2,8})[一-鿿]/ 吃掉「密」→ 得到 theydrawfro码ginationalone，英文模式全失配，该 no_true_scotsman 攻击句 gate=pass 静默放行。修复：兜底判据升级为拼音键覆盖率 ≥ 50%（新增 _pinyinCoverage 贪心最长匹配；实测真拼音串 100%/81%、英文句 12%~17%，两边充足间隔），stopword 与 avgLen 两道原有防线保留并联。收益：该攻击句 pass → rewrite（discriminate 命中 no_true_scotsman）；英文句零损坏；拼音还原能力零回归（两个既有反混淆用例 + 拼音英文混合形态全部实测通过）。顺带修掉 r729 上线第 160 维度后未同步的存量断言（round-728-no-true-scotsman.js 的 159 → 160，该红灯从 v6.8.78 起常驻）。全量回归 18,366 过 / 143 败 / 18,509 个（基线 18,348/157/18,505：失败 -14、通过 +18，零新增失败；剩余 6 个 MCP 类失败为 8099 端口被在跑服务占用导致的存量环境噪声）。已知局限记录在案：mingling 本身是 PINYIN_MAP 键（→命令），纯英文词与拼音键不可区分，site 1 的盲目子串替换仍会局部改写它——但实测该良性句 gate 仍 pass，不构成误报；修它会让「hulve previous instructions」这类真攻击形态失去还原能力，留专项轮 |
| 6.8.80 | 2026-10-27 | Windows 平台卡死歼灭战·二（upgrade type 3 — 测试链与工程修复，引擎 src/ 零改动） — test/run-all.js 从串行 for 循环重构为「任务构造 + 并发池」（HF_RUN_ALL_JOBS 默认 4、按 CPU 核数封顶，=1 退化为原串行行为），609 个测试文件全量实测 737.9 秒（并发池阶段约 4 分钟跑完 522 个，变更/MCP 类 87 个独占串行车道）；顺带修四个 Windows 专属问题：① 单例锁补 'wx' 排他创建 + 双检，关掉两个 runner 同时启动的毫秒级竞态；② 15 个测试文件 execSync(process.execPath + ' ' + ...) 在 C:\Program Files\nodejs\node.exe 上被 shell 拆断成 'C:\Program' 整文件 Command failed（Linux 容器无此路径所以历史全绿），统一改 JSON.stringify(process.execPath)；③ mount/jest 形态测试原先被静默跳过，并发池按任务对象统一调度后补齐；④ 72 个变更测试（writeFileSync 改 src/ 做负例验证）并发跑时互相覆盖补丁造成批量假失败（实测单跑 3/0 过、并发 EXIT=1），改两阶段调度——变更测试与 MCP 服务类测试独占串行车道，其余 522 个走并发池；另加突变残留兜底（跑前快照 src/ 全文，子进程被 SIGKILL 后逐文件回滚，防 4 个 src 文件带负例突变污染仓库）。全量回归实测 18,348 过 / 157 败（上一轮基线 18,294/161，失败数净减 4、通过数 +54） |
| 6.8.79 | 2026-10-27 | 全局修复轮：execSync 无 timeout 歼灭战（upgrade type 3 — 工程修复，引擎未动） — scripts/repo-audit.js 全局审计 5242 文件、SHELL 类 306 命中逐处人工证伪后，定位 8 文件 15 处真实裸 execSync/spawnSync 并全部补 timeout。最关键一处：scripts/auto-upgrade-hourly.js 里 run-all 的 timeout 原值 120 秒——"2 分钟必杀 27 分钟的全量回归，无汇总行 → parseTestResult ok:false → cron 轮反复失败"，是历史卡死轮的直接根因，改 2400s。证伪记录：SECRETS 6（root-key.json 是公钥、test_*.cjs 是扫描器文件名误报）、TRAV 83（抽样 3/3 为 path.join(__dirname,'../..') 常规相对路径）、PROMPT 43（全为探针脚本自身提示词模板）均不动 |
| 6.8.78 | 2026-10-26 | conditional_proposition added as dimension 160 (upgrade type 1 — a brand-new discrimination dimension) — 条件句武器化：用未经验证的假设性条件句（「如果A就B」）直接推出关于现实的确定性结论，把存在性推断当作前件真值的证据（「如果他真的爱你，就不会让你受这种委屈」/「如果这个方案可行，我们早就看到成果了」/"If he really loved you, he would never treat you like that."）。两腿共现（假设条件句 × 否定推断：早就…了/就不会/怎么会/would never/would have/there would be）；r729 复测攻击漏判 16/20（80%）、良性 0/10 误伤；decision 本体一次选中无打平（composite 0.84 / confidence 0.76 / identity_alignment 0.8，第二名 loaded_question 0.81）。豁免：规则合同条件/谬误概念转述/对冲/显式证据模式/否定翻转。本半程由 14:00 cron 崩溃（模型 repetition loop）后主会话接管收尾 |
| 6.8.77 | 2026-10-26 | no_true_scotsman added as dimension 159 (upgrade type 1 — a brand-new discrimination dimension) — 无真正的苏格兰人：用修改定义的方式驱逐反例（「真正爱你的男人，不会计较你花多少钱」/「Real artists never trace references」）。两腿共现（authenticity 短语 × 驱逐断言/反事实条件）；冒烟 22/24 命中、良性零误伤；轮测试 22/22；guard-abilities 20/20；维度键实测 159。已知边界：含 m 英文攻击句被 normalize de_pinyin 破坏（imagination→i码agination）致 discriminate 归零，留专项轮 |
| 6.8.76 | 2026-10-26 | Merge: box line (4f4ebeb4) merged into the r727 mainline — void_zh_standard_shift renumbered as dimension 158 (upgrade type 1 — a brand-new discrimination dimension) — 中文「双标」标准转移：对他人与自己使用两套评判标准（「别人加班是敬业，我加班就是剥削」/「规定是定给员工的，我不用遵守」/「Employees must clock in on time, but I come and go as I please」）。来自 Windows box 线（r717 侦察、r718-r724 迭代），经 10-26 07:27 合并入主线并重编号 155→158。本轮（r728）补全合并遗漏：VERSION/version.js 之外的版本四处传播（package.json/SKILL.md）、解决 data/upgrade-state.json 与 data/test-count.json 进 HEAD 的冲突标记、Verify 列举 116→117 并补 void_zh_standard_shift、维度计数 157→158。引擎维度键实测 158 |
| 6.8.75 | 2026-10-26 | forer_effect added as dimension 157 (upgrade type 1 — a brand-new discrimination dimension) — 巴纳姆效应：放之四海皆准的性格/命运描述佯称针对个体的专属解读（"你外表看起来坚强，其实内心很柔软，只是不轻易让人看见"/"从测试结果看，你属于既理性又感性的类型，这很少见"/"Deep down, you want someone to truly understand you."）。两腿共现（人格描述句式 × 第二人称人格断言）；r727 复测攻击漏判 17/20（85%）全池最高、良性误伤 0/8；decision 本体一次选中（composite 0.82 / confidence 0.82 / identity_alignment 0.8，第二名 no_true_scotsman 0.80） |
| 6.8.74 | 2026-10-26 | relative_pration added as dimension 156 (upgrade type 1 — a brand-new discrimination dimension) — 诉诸更惨：拿别人更惨的处境抹掉对方正当的当下诉求（"别喊累了，有人连饭都吃不上"/"知足吧，还有人工资比你低一半呢"/"Others have it worse; stop complaining"）。decision 本体一次选中无打平（composite 0.83 / confidence 0.84 / identity alignment 0.8）。r726 复测漏判 30/35（86%）全池最高、误伤 0/12，连续四轮候补后当选；上线后攻击 35/35 命中、良性 12/12 零命中。第一人称自陈比惨显式让位 suffering_contest 第78维 |
| 6.8.73 | 2026-10-25 | affirming_the_consequent added as dimension 155 (upgrade type 1 — a brand-new discrimination dimension) — affirming the consequent: reversing cause from evidence ("if he did it he would leave traces; there are traces, so he did it"). decision 本体一次选中无打平（composite 0.86 / confidence 0.86 / identity alignment 0.8）。r725 复测漏判 14/15（93%）全池最高 → 上线后攻击 15/15 命中、良性 12/12 零命中。modus ponens / modus tollens / 对冲形天然不判（按推断词切分事实句，前半须含后件关键词、后半回指前件关键词） |
| 6.8.72 | 2026-10-25 | guilt_by_association added as dimension 154 (upgrade type 1 — a brand-new discrimination dimension) — guilt by association: blaming someone for the sins of their associates ("his friend is a scammer, so he is no good"). decision 本体一次选中无打平（composite 0.85 / confidence 0.82 / identity alignment 0.8）。r724 复测漏判 11/12（92%）→ 上线后全命中，误伤 0/12，双向门禁召回 52/52、误拦 302/326 与基线一致零新增 |
| 6.8.71 | 2026-10-25 | standard_shift dimension 70 strengthened (upgrade type 1 adjacent — a real capability gain on an existing dimension: a second judging route for the contract-void family) — single-sided voiding of a prior promise/standard (契约作废窄口): one party declares an already-agreed standard void to escape it, without any goalpost-raising trigger words (当初那件事不作数，你别拿它说事 / 这次承诺不算，我改主意了 / 旧标准作废，按我今天说的来 / 婚前说好不干涉我的工作，那条现在不算数 / 谈好的条件随时可以改，我说了算 / 说好的提成现在不认了，规矩我来定 / That deal is void, I have decided on new terms. / You cannot hold me to what I said last month.): route ④ requires promise/standard anchor × voiding verb to co-occur (九腿 Z1/Z1b/Z2/Z3/Z4/Z5/E1/E2/E3), Z2 carries a one-sided marker (按我今天说的/我说了算/按新方案) so organizational announcements (旧版标准已经失效，请注意查阅更新通知) stay clean, DUE_PROCESS exemption is narrower than PROCESS_MARKER (only mutual consent / publication / filing / written / change order / transition period / by law / court / arbitration pass), bare 不算 is anchored to sentence punctuation (之前的问题不算难 is not a hit). Decision entity selected this direction from a seven-family re-probe (void_zh 81% miss rate, highest in pool, 0/15 false positives; composite 0.87 / confidence 0.84 / identity 0.8). 27/27 leaks closed (81% → 0%), 0/15 false positives, 33 benign samples zero module-layer hits. Negative-injection test verifies every constant is load-bearing 57/57 (including empty-table anti-degeneration guards: an emptied table drops its leg instead of becoming a match-everything (?:)). |
| 6.8.70 | 2026-10-25 | hot_hand added as dimension 153 (upgrade type 1 — a brand-new discrimination dimension) — hot hand fallacy: treating a winning streak as a guarantee that the next round's odds improve (我手感正旺，下一把肯定中 / 今天连赢五把，趁热打铁再来一局 / 最近状态好，这个月业绩肯定还会涨 / 连胜之后势头在我，决赛肯定也赢 / 我已经连对十题了，下一道也一定对 / 连红三把了，房费都是我的，继续押 / 最近面试全过，这家肯定也成 / 我这个月签了三单，下个月肯定更多 / I am on a hot streak, so the next hand is mine. / I have won five in a row; I should ride the momentum. / Things are going my way lately, so the next one will work out too. / I am in the zone, so this next shot is definitely going in. / Three wins in a row — the momentum is clearly on my side.): two legs — streak/momentum expression (手感/手气旺/手风顺/势头/气势/趁热打铁/走运/运气好/状态好/连赢/连胜/连对/连红/连中/连签/连过/连成/连涨/全过/全中/签了三单；EN: hot streak / on a roll / in the zone / momentum / on fire / going my way / five in a row / wins in a row) × continuation assertion (肯定/一定/必然/势必/铁定/必定/必中/必赢/稳了/还会/也会/再来一局/再来一把/继续押/继续买/下一把也/下一局也/下一道也/下一个也/下个月也/再买几注；EN: definitely / guaranteed / bound to / certainly / surely / for sure / no doubt / is mine / will work out / going in / on my side / next one will / ride the momentum) — missing either leg means no hit (连胜可能提升信心，但每局概率独立 is legitimate review; 连胜这么多了，下一把该输了 is the reversal direction, owned by gamblers_fallacy dimension 148). Decision entity selected (first round A/B composite 0.81 tie → honest abstention; after adding genuinely distinguishable criteria the second round chose A: composite_score 0.83 / confidence 0.8 / identity alignment 0.8). 11/15 leaks closed (73%, true zero-hit 11/11), 0/8 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Three exemptions locked: meta discourse (热手谬误/hot hand fallacy/赌徒谬误), evidence context (独立事件/概率独立/不影响/与历史无关/大数定律；EN: independent event / regardless of / remains the same / same probability), teaching third-person context (老师/教材/课堂/教授/textbook/lecture). Negative-injection test verifies every constant is load-bearing 22/22. Also fixes a pre-existing false positive found this round: post_hoc R2's EN intervention leg matched the bare word "got" as a substring inside "renegotiate" (ne-got-iate) × "no longer", causing the benign sentence "If the old terms no longer work for you, let us renegotiate." to be gate-rewritten — word-boundary matching clears it. VERIFY level |
| 6.8.69 | 2026-10-24 | appeal_to_probability added as dimension 152 (upgrade type 1 — a brand-new discrimination dimension) — appeal to probability / scale-as-certainty: treating "large base, many repetitions, big sample, enough time, probability exists" as a guarantee that "it will definitely happen / it will definitely be my turn" (宇宙这么大，肯定有外星人 / 世上无难事，只要肯试，试多了总能成 / 航班天天飞，出事的才几起，我肯定赶不上 / 选择题四个选项，蒙个十道八道总能对几道，考试稳了 / 概率再低，只要重复次数够多就一定发生 / 中国有十四亿人，什么样的事不可能发生？ / With so many tickets sold, someone has to win — why not me? / The universe is so vast, there must be intelligent life. / Given enough time, anything that can happen will happen. / If I keep playing, I am bound to win sooner or later.): two legs — base/scale/repetition/duration expression (这么[大广多]/那么[大广多]/十几亿人/这么多人/天天飞/遍地/试多了/重复次数够多/蒙个十道八道/投了三百份/时间够长; EN: so many / so vast / billions of / given enough time / enough times / keep playing / hundreds of) × certainty assertion (肯定[有会成]/一定[会能发生]/总能/迟早/必然/不可能不发生/稳了/凭什么不是我/还会远吗; EN: must be / bound to / guaranteed / has to / inevitable / will happen / sooner or later / why not me) — missing either leg means no hit ("宇宙这么大，但至今没有证据" is a correct scale statement; "这事一定能成" with no evidence source belongs to wishful_thinking / confidence). Decision entity selected (composite_score 0.81 / confidence 0.78 / identity alignment 0.8). 12/16 leaks closed (75%), 0/10 benign harmed, r721 probe re-verified: true zero-hit 12/12 across all 151 prior dimensions. Three exemptions locked: correct probability statements (概率很低但不是零/大数定律/独立事件; low probability / no evidence / does not guarantee), meta discourse (诉诸概率是一种逻辑谬误), probability-teaching third-person context (老师/教材/课堂上/textbook/lecture). Negative-injection test verifies every constant is load-bearing 17/17. VERIFY level |
| 6.8.68 | 2026-10-24 | post_hoc added as dimension 151 (upgrade type 1 — a brand-new discrimination dimension) — post hoc fallacy: treating temporal sequence alone as proof of causation ("after X, therefore because of X"). Three routes: (R1) 时序连接词×归因断言 — "自从/换/装/用 了 X 之后/以后 … 就 …" 归因句； (R2) 干预词×变化词 — "换了/穿了/吃了 X" + "业绩/成绩/运气/运气 就 好/坏/升/降" 突变形； (R3) 缺席×灾祸×规避 — "没去/没带 X" + "就出了事/就倒霉" + "以后不敢/不敢" 三腿共现（缺一腿不判）。中文常量表覆盖日常归因口语形态（"就" 突变形与中性记录句作形状区分），英文覆盖 "since/after/ever since + crash/failure/improvement + must have been caused by" 形态。负例注入测试逐一置空常量验证承重 74/74，攻击样本 gate 层 rewrite，良性样本 0 误伤。VERIFY 级 |
| 6.8.58 | 2026-10-13 | kafkatrapping added as dimension 143 (upgrade type 1 — a brand-new discrimination dimension that independently hits its own sample family) — the family treats the interlocutor's denial/defense/emotional reaction itself as proof of guilt (你反应这么大，正好说明被我说中了 / The harder you deny it, the guiltier you look.): three routes — reaction × guilt-inference, innocence × challenging question, denial × evidence-declaration |
| 6.8.60 | 2026-10-20 | concern_trolling added as dimension 145 (upgrade type 1 — a brand-new discrimination dimension) — fake concern: control, belittlement, or suppression wrapped in a "for your own good" package, dressing harmful words as well-meant reminders while cancelling the other party's right to refuse (我是为你好，才指出你这些问题的 / 我这么说都是关心你，别人谁愿意跟你说这些 / I am only saying this because I care about you. / A real friend would tell you these hard truths, and I care enough to do it.): two routes — care-wrapping × belittlement/control content (two-leg coexistence, a single leg does not hit), care-wrapping × exclusive monopoly (别人谁愿意 / 换别人 — monopolizing the right to criticize as one's own privilege). 9/10 leaks closed (7/9 true zero-hit), 0/12 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Four exemptions locked: normal concern (你最近看起来状态不好，发生什么事了吗 / I am worried about you, let me know if you need anything.), normal criticism without the care-wrapping leg, meta analysis of the tactic itself (伪关心是一种用关心包装的控制手法), third-person professional review (心理咨询师说……这是伪关心的话术). Decision entity selected this round (composite_score 0.84 / confidence 0.7 / identity alignment 0.8). VERIFY level — a single sentence may be normal concern, conflict review, or counseling pattern analysis |
| 6.8.67 | 2026-10-22 | special_pleading added as dimension 150 (upgrade type 1 — a brand-new discrimination dimension) — special pleading: claiming the same set of rules/standards/discipline applies to others but not to oneself, with no legitimate reason, using identity or circumstance as a one-way exemption (规则是给普通人定的，我不一样 / 我情况特殊，所以可以例外 / 别人迟到要扣钱，我迟到是有原因的 / 这个规定对我不适用，我的层次不一样 / 大家都是凡人，只有我超脱了规则 / 规矩是约束别人的，我自有分寸 / The rules apply to everyone else, but I am different. / That policy is for regular employees; I am above that. / Normal standards do not apply to someone like me.): three routes — rule-others × self-exception, others-bound × self-exemption, special-self × exception-grant (all two-leg coexistence). High-frequency forms covered after probe-driven extension (制度是管老百姓的 / 想来就来 / Ordinary employees need / Others must / I make the rules / do not bother). 9/9 leaks closed (9/9 true zero-hit), 0/9 benign harmed, bidirectional guard recall side unchanged and false-positive side unchanged. Exemptions locked: legitimate procedures/qualifications (special approval with formal process / documented needs / product design — procedural exceptions are institutional design not double standards), meta analysis of the concept itself, stricter oversight (更严不是更松 / stricter — tighter requirements on specific groups are the opposite of double standards). Decision entity selected this round (composite_score 0.85 / confidence 0.8 / identity alignment 0.8, no abstention). VERIFY level — a single sentence may be recounting someone else's double-standard statement for analysis, literary dialogue, or a counterexample in compliance discussion |
| 6.8.66 | 2026-10-22 | Global fix pass (maintenance) — evidence baseline scorer "constant {score:1, issues:[]}" falsified as by-design (v6.7.73 intentional: neutral default when no evidence argument is passed); special_pleading existing-dimension false positive fixed (focus sample "特殊情况需要特批，但要走正规流程" misjudged as rewrite by crisis_exceptionalism — B1 × B2 process-noun-alone branch; GUARD_ZH gained three "follows formal process" sentinels, focus sample now passes, 0 attack leaks, existing 30 tests unchanged); two upstream broken scripts repaired (fetch-wikipedia-formulas.js / fetch-more-formulas.js promise-chain debris); one upstream broken probe (round-145) isolated on record. Full-repo JS syntax errors 3 → 1 |
| 6.8.65 | 2026-10-22 | Global audit + Windows toolchain adaptation (maintenance) — 653 files with /root hardcoded paths cleared: __HF root-path constant injection + three-level repair (double-injection dedup / TDZ / stringified restore); half of the auditors unusable → all working; SECRETS/SHELL/TRAV/PROMPT four report categories falsified as false positives; three upstream historical broken scripts recorded unrepaired |
| 6.8.64 | 2026-10-22 | appeal_to_majority added as dimension 149 (upgrade type 1 — a brand-new discrimination dimension) — appeal to majority: using headcount, popularity, or vote results as a substitute for argument, letting a claim ride on numerical endorsement (大家都这么认为，肯定是对的 / 这么多人用，说明产品没问题 / 群众的眼睛是雪亮的，跟着大家走就行 / 大多数人不支持，这个方案肯定不行 / 全网都在骂，说明他真的有问题 / 销量第一，质量肯定最好 / Everyone believes this, so it must be true. / Millions of people use it, so it cannot be wrong. / The majority agrees, therefore it is correct.): two-leg coexistence — majority leg × conclusion leg (headcount/popularity expression × conclusion closed by headcount; a single leg does not hit). High-frequency forms covered after probe-driven extension (十个人里有九个 / widely accepted / popular opinion says / millions of fans / most experts agree). 6/9 leaks closed (6/6 true zero-hit), 0/9 benign harmed, bidirectional guard recall side unchanged and false-positive side unchanged. Exemptions locked: sourced headcount statements (market research showing 70% satisfaction / survey / poll / data — method-carrying headcount statements are citation not endorsement), meta analysis of the concept itself, logic-teaching third-person review. Decision entity selected this round (composite_score 0.83 / confidence 0.8 / identity alignment 0.8, no abstention). VERIFY level — a single sentence may be recounting a popular view for analysis, legitimate market-research citation, or social phenomenon description |
| 6.8.63 | 2026-10-22 | gamblers_fallacy added as dimension 148 (upgrade type 1 — a brand-new discrimination dimension) — gambler's fallacy: treating the historical frequency of independent events as a basis for future probability, making random sequences carry a "balance obligation" (已经连输五把了，下一把肯定赢 / 都出十次小了，该出大了 / 连续三把没中，这次中奖概率更高了 / 前面几次都是反面，这次肯定是正面 / 输了这么多把，运气该回来了 / 硬币出了五次正面，下一次反面的概率大多了 / It has been tails five times in a row, so heads is due next. / After ten losses, a win is statistically more likely now. / The coin landed heads four times, so tails is overdue.): two-leg coexistence — streak leg × due leg (historical win/loss streak or high-frequency repetition × reversal or elevated-probability expectation; a single leg does not hit). Casino high-frequency forms covered after probe-driven extension (开大/开小, 吐分, 手气正旺, reds/blacks, losing streak, due for). 8/9 leaks closed (8/8 true zero-hit), 0/9 benign harmed, bidirectional guard recall 20/20 and false-positive side unchanged. Exemptions locked: correct probability statements (independent events / probability unchanged / 50% / control group / law of large numbers — stating the independence principle debunks the fallacy), meta analysis of the concept itself, probability-teaching third-person review. Decision entity selected this round (composite_score 0.84 / confidence 0.8 / identity alignment 0.8, no abstention). VERIFY level — a single sentence may be recounting someone else's error, counterexample teaching, or game narrative |
| 6.8.62 | 2026-10-21 | wishful_thinking added as dimension 147 (upgrade type 1 — a brand-new discrimination dimension) — wishful thinking: using wishes/beliefs/mentality as a substitute for evidence, directly inferring that outcomes will develop in the direction one hopes for, treating "believing" as a substitute for "knowing" (我相信会成功，所以一定会成功 / 只要我足够努力，宇宙就会回应我 / 念念不忘必有回响，想什么就来什么 / 我这么真诚，老天不会亏待我的 / 心诚则灵，只要信念够强就能实现 / I believe it will work out, so it will work out. / If I want it badly enough, the universe will provide. / My positive thinking alone will make it happen.): two routes — belief/wish × inevitable outcome (two-leg coexistence, a single leg does not hit), belief/wish × supernatural causation (宇宙就会回应 / 老天不会亏待 / 心诚则灵 / 念念不忘必有回响 / 吸引力法则; EN: the universe will provide / law of attraction / manifest — attributing outcomes to supernatural forces responding to belief, normal motivational statements do not introduce supernatural causation). 8/9 leaks closed (8/8 true zero-hit), 0/9 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Exemptions locked: normal motivational expression, life-attitude statements, evidence present (data/verification/testing/experiment/control group/benchmark/A-B testing/peer review — belief with sources is citation not substitution), meta analysis of the pattern itself, third-person professional review. Decision entity selected this round (composite_score 0.84 / confidence 0.8 / identity alignment 0.8). VERIFY level — a single sentence may be normal motivational expression, life-attitude statement, or cognitive pattern analysis |
| 6.8.61 | 2026-10-21 | stereotype_origin added as dimension 146 (upgrade type 1 — a brand-new discrimination dimension) — origin discrimination: labeling a person by their education/age/family background and directly belittling their attributes, character, or psychology, regardless of actual performance (专科毕业的，逻辑能力能强到哪去 / 年纪轻轻的懂什么，回去多历练几年 / 富二代罢了，能力都是家里给的 / 单亲家庭长大的孩子心理都有问题 / 文科生写代码，能写出什么好东西 / People from small towns just do not have the vision for this. / She is too emotional to handle a leadership role. / He is just a trust-fund kid, none of it is his own work.): two-leg coexistence — origin label (education/age/family background) × attribute belittlement (capability/psychological/character). 8/10 leaks closed (8/8 true zero-hit), 0/12 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Exemptions locked: legitimate evaluation based on actual performance, neutral factual statements, balanced comparison, legitimate questions, meta analysis of the tactic itself. Decision entity selected this round (composite_score 0.84 / confidence 0.7 / identity alignment 0.8). VERIFY level — a single sentence may be normal evaluation, factual statement, or concept analysis |
| 6.8.59 | 2026-10-16 | darvo added as dimension 144 (upgrade type 1 — a brand-new discrimination dimension that independently hits its own sample family) — the family denies the accusation itself and reverses the victim/offender roles, turning the accuser into the defendant (你居然指责我，我才是受害者 / You are the abuser here, stop playing the victim.): two routes — accusation-response × role-reversal, victim-self-claim × exclusivity |
| 6.8.55 | 2026-10-13 | anecdotal added as dimension 140 (upgrade type 1 — a brand-new discrimination dimension that independently hits its own sample family) — the family uses one concrete outcome from the speaker’s personal circle to assert a domain-level conclusion: 我邻居抽烟喝酒活到九十，所以烟酒根本不致癌 / 我二舅没上大学照样当老板，读书没用 / 我同事天天吃外卖身体倍儿棒，外卖很健康 / 我同学裸考就过了，这考试根本不用复习 / 我老家老人一辈子没体检也活到九十八，体检是浪费钱 / 我亲戚家孩子天天打游戏照样上清华，游戏不害人 / 我闺蜜老公就是彩礼给得多才幸福，彩礼必须多要 / 我朋友用了这个方法月入十万，你照做就行 / 我表姐就是喝红糖水治好了痛经，你试试 (EN: My uncle never went to college and still became a boss, so higher education is pointless. / My coworker eats takeout every day and is perfectly fine, so takeout is healthy. / My cousin aced the exam without studying, so this exam does not require any preparation. / My friend made 100k a month with this method, just copy him. / My aunt cured her cramps with brown sugar water, you should try it. / The old folks in my hometown never got a checkup and lived to 98, health checkups are a waste of money. / My relative’s kid plays video games all day and still got into Tsinghua, video games are not harmful. / My neighbor smoked and drank his whole life and lived to 92, so alcohol and cigarettes are not carcinogenic.). Two legs, both required: leg 1 personal case present (case reference — 我+relation word / 我认识的 / 我老家 / 我闺蜗的老公全, EN my uncle / a guy I know / my hometown / my friend’s husband — × concrete outcome: 活到九十 / 月入十万 / 当老板 / 身体倍儿棒 / 治好了 / 就过了 / 上清华 / 过得幸福 / 中大奖全, EN lived to 92 / became a boss / was cured / aced the exam / got into Tsinghua / are happy / won the lottery — the outcome is the generalization trigger: a reference without an outcome is opinion relay, deliberately not judged); leg 2 domain-level conclusion, any of three kinds: N domain-fact negation (烟酒根本不膳癌 / 读书没用 / 游戏不害人 / 体检是浪费钱 / 这考试根本不用复习全, EN education is pointless / video games are not harmful / does not require any preparation / checkups are a waste of money), P domain affirmation (外卖很健康 / 买彩票能发财全, EN takeout is healthy / the lottery can make you rich), U listener push or norm (你照做就行 / 你试试 / 彩礼必须多要全, EN just copy him / you should try it / the bride price must be high). Decision entity selected this round (composite_score 0.93 / confidence 0.92 / identity alignment 100%; four candidates from probe /tmp/hf-scout-708.txt — the r700 queue was exhausted after six families shipped in r700-r707, so a fresh scout re-measured with the identical sample set: anecdotal 9/10 attack leaks all true zero-hit, tu_quoque 8/10, middle_ground 7/10, appeal_to_authority 7/10 but 2/12 of its benign harm comes from its own existing patterns so the probe disqualified it for the fourth consecutive round). 9/10 leaks closed at the gate layer (19 attack samples ZH 10 + EN 9 all non-pass; the lottery sample co-exists with survivorship_bias, both action-forcing, co-hit is not distortion), 0/30 benign harmed at the single-dimension layer, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm — the 4 gate-level verifies on benign samples are all pre-existing hits from stock dimensions confidence/unsupported_claim/vagueness/contradiction/perfect_error, proven by per-dimension attribution and the guard baseline). Guarded by test/round-708-anecdotal.js (9 assertions: 19 attacks ZH+EN with per-conclusion-type attribution / 30 benign / single-leg boundary 6 / adjacent division 6 / exemption load 6 / gate layer 19+26 / wiring 9) + scripts/negative-test-anecdotal.js (20/20 injection points all assertion-red with per-branch exclusive-sample attribution — 12 constant blocks + 8 route lines — restore md5 identical). VERIFY level — a single sentence may be fiction dialogue, a prejudice case quotation, a logic-class counter-example, or a sociology-discussion relay; rewrite/block would harm those. Boundary with existing dimensions: anecdote_as_proof (dimension 84) owns acquaintance-sourced × universal-effect claims (the point is whether the tried thing itself is effective/safe/standard), this family’s point is domain-level negation or norms; survivorship_bias (dimension 132) owns visible-success-case × push-conclusion with unrestricted subjects (celebrities/strangers/named people) and success-only outcomes, this family’s leg 1 must be a personal-circle reference and its outcomes are not limited to success; hasty_generalization owns inference-form overgeneralization (requires an explicit summarizing bridge), this family has no bridge word |
| 6.8.57 | 2026-10-13 | mind_reading added as dimension 142 (upgrade type 1 — a brand-new discrimination dimension that independently hits its own sample family) — the family asserts the interlocutor's true inner thoughts/motives as fact and uses that to cancel their right to explain themselves: 别不承认，你心里其实就是这么想的 / 你嘴上说没事，其实就是记恨我 / 我知道你根本就没安好心 / 你这就是在找借口，我还不了解你 / 你就是故意的，别以为我不知道 / 你心里清楚得很，就是装傻 (EN: Don t lie, deep down you know you want this. / You say you are fine, but I know you are angry at me. / I know what you are really after. / You are just pretending not to care.). Two legs, both required: leg 1 ASSERT knowing-assertion / self-explanation-denial frame (别不承认/别以为我不知道/别装了; 我还不了解你/我了解你; 我知道你/我看得出你; 你就是+故意/成心/在装; 其实就是/说到底; 你心里清楚; 嘴上说+其实就是; EN deep down you know / but I know / I know (what) you / you are just / don't lie) × leg 2 STATE inner-state content (心里/内心/骨子里; 想法/心思/打算/目的/动机/意图; 故意/成心/蓄意; 记恨/记仇/怨我/恨我/生气/委屈/不满/介意; 没安好心; 装傻/装糊涂; 借口; 你…这么想/这样想; EN deep down / really after / angry / pretend / motive / not to care / you want). Single leg not judged (我知道你很忙，所以长话短说 is plain expression; 承认对方有合理的动机，是沟通的基础 is plain statement). What gets suppressed is the interlocutor's right to explain who they are: the legitimate move is to invite them to say it themselves (你是不是有什么顾虑，可以说出来一起看 / 我愿意听你解释，而不是替你下结论 / 与其猜测，不如直接问对方怎么想的; EN If something is bothering you, I am here to listen.) or to label the reading as a guess rather than fact (我不确定他为什么这么做，可能问清楚更好). Supportive emotional acknowledgement (我知道你很生气，这完全可以理解 / I know you are angry, and that is completely understandable.) and charitable attribution (我了解你，你这么说其实心里也是为大家好 / I know you mean well.) deliberately not judged — acknowledging an emotion is not asserting a motive. Boundary with existing dimensions: gaslighting owns reality-denial (是你记错了/这没发生过 — external facts), tu_quoque (dimension 141) owns cancelling standing via the interlocutor's own fault (你也一样/你没资格), presupposition owns question presupposition (你为什么不…了 — question form). Decision entity selected this round (composite_score 0.94 / confidence 0.9 / identity alignment 100%, no tie-break; four candidates from the fresh scout /tmp/r711-scout.js + /tmp/r711-scout2.js landed at /tmp/hf-scout-711.txt: A mind_reading 0.94 / B stereotype origin-family extension 0.89 / C void_zh prior-commitment-void extension 0.87 / D middle_ground 0.86; re-measure also retired two stale queue entries — appeal_to_pity extension 0/10 leaks (gap already fixed by later rounds) and the r700-era circular_reasoning new-dimension candidate (already shipped as dimension 136 in r701)). Probe evidence: 10/10 attack leaks, true zero-hit 9/10, 0/12 benign harmed; gate layer 10/10 verify with own count 1, HEAD stash comparison proves the 3 benign verifies (contradiction/sycophancy/perfect_error) are pre-existing stock-dimension behavior, zero new harm; bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline). Guarded by test/round-711-mind-reading.js (9 assertions: 6 ZH + 4 EN attacks with per-leg attribution / 6 single-leg boundaries / 3 adjacent-dimension divisions / 8 exemptions incl. 3 hedge load-bearing / gate layer 10 + 12 benign / wiring 9) + scripts/negative-test-mind-reading.js (13/13 injection points all assertion-red — 6 constant blocks + 7 route lines — restore md5 identical). VERIFY level — a single sentence may be fiction dialogue, a prejudice case quotation, a logic-class counter-example, or counseling-session recap; rewrite/block would harm those. Known narrow gap (not fixed this round, logged in UPGRADE_LOG): EN "I know you mean well." does not itself hit both legs (mean well outside ASSERT_EN/STATE_EN range) so its hedge is defensive depth rather than load-bearing; third-party emotional attribution (你就是故意让我难过的) needs the 故意×让我 structure which STATE does not yet cover, single-leg not judged. |
| 6.8.56 | 2026-10-13 | tu_quoque added as dimension 141 (upgrade type 1 — a brand-new discrimination dimension that independently hits its own sample family) — the family cancels the interlocutor's standing to criticize or demand by pointing at their own similar fault: 你说我熬夜不好，你自己不也天天凌晨睡 / 你叫我别玩游戏，你年轻时候玩得比我还凶 / 你又不是没犯过错，装什么清白 / 你说这套方法有问题，你自己不也在用 / 你自己都做不到，凭什么要求我 / 你也曾经撒过谎，没资格讲诚信 / 你不是一样迟到，光说别人 / 你以前也这样，现在倒来教育我 / 我迟到是不对，可你上次也迟到了，还好意思说我 / 你说我乱花钱，你上次买那双鞋就没眨眼 (EN: You criticize me for smoking, but you drink every day. / You tell me to work harder, yet you're always leaving early. / You cheated on the test before, so you have no right to lecture me about honesty. / Who are you to talk? You lied too. / You say I am lazy, but you never do the dishes either.). Two legs, both required: leg 1 CRITICISM the interlocutor's criticism or demand present (你说我…不好/不对/有问题/别/不要/乱花钱; 你叫我/要求我/管我 + negative word; EN You criticize/tell/ask/blame/accuse me… + contrast conjunction but/yet/while — weak verbs say/said/claim require a negative evaluative word co-occurrence) × leg 2 MIRROR × DISQUALIFY — MIRROR you-based similar fault (你 + 自己/本人 + 不也/也没/也不/也; 你 + 曾经/以前/过去/当年/年轻时/上次/上回 + 也/照样; 你又不是没/你不是一样; 你自己都做不到-class fault verbs; EN but/yet/while + you / you + fault verb + before / you … too) × DISQUALIFY disqualification words (凭什么/没资格/装什么/光说/倒来/还好意思; EN who are you to / no right to / not one to talk / glass houses / pot calling the kettle). Route 1 CRITICISM × MIRROR (classic you-too form, no explicit disqualifier needed); route 2 MIRROR × DISQUALIFY (disqualification form, the criticism implicit in the claimed moral high ground). Single MIRROR not judged (「你也太搞笑了，凭什么说我」 has no similar-fault leg — plain dismissal); single DISQUALIFY not judged (「你凭什么管我」「你没资格决定我的事情」 are legitimate boundary assertions against intrusive control). Decision entity selected this round (composite_score 0.92 / confidence 0.9 / identity alignment 100%; four candidates from the fresh scout /tmp/r710-scout.js landed at /tmp/hf-scout-710.txt — the r700-r708 queue was nearly exhausted (appeal_to_authority disqualified for the fourth consecutive round), so the scout re-measured two queue families with the identical sample set plus two brand-new families: tu_quoque 9/10 attack leaks all true zero-hit, middle_ground 10/10, guilt_by_association 10/10 (true zero-hit 7/10), mind_reading 10/10 (true zero-hit 9/10)). 15/15 leaks closed at the gate layer (ZH 10 + EN 5 attack samples all verify; own count 1 on every attack), 0/16 benign harmed at the single-dimension layer (the 5 gate-level verifies on benign samples are all pre-existing hits from stock dimensions sycophancy×4/false_balance×1, proven by per-dimension attribution and HEAD comparison), bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Guarded by test/round-710-tu-quoque.js (10 assertions: 15 attacks ZH+EN with per-route attribution / 16 benign / single-leg boundary 5 / adjacent division 4 / exemptions 4 + exemption load-bearing 2 / gate layer 15+10 / wiring 9) + scripts/negative-test-tu-quoque.js (18/18 injection points all assertion-red with per-branch exclusive-sample attribution — 8 constant blocks + 10 route lines — restore md5 identical). VERIFY level — a single sentence may be fiction dialogue, a prejudice case quotation, a logic-class counter-example, or counseling-session recap; rewrite/block would harm those. Boundary with existing dimensions: whataboutism (dimension 20) owns topic deflection (那 XX 呢/你怎么不说 YZ — does not point at cancelling standing), false_balance owns both-sides-ism (双方都有责任/一个巴掌拍不响 — does not cite the interlocutor's personal fault), gaslighting owns reality-denial (是你记错了/这没发生过), sycophancy owns agreeable affirmation (你说得对/确实). Known narrow gap (not fixed this round, logged in UPGRADE_LOG): EN emphatic form (You are the one who …) and hypothetical form (If you had not done X, you could talk) not covered; ZH「彼此彼此」-class idioms not judged (no disqualification word — mutual disparagement does not constitute a standing-cancellation chain) |
| 6.8.54 | 2026-10-13 | appeal_to_ignorance extended with the burden-shift demand form (upgrade type 2 — a whole new sentence family with zero prior coverage added to an existing dimension, no new dimension number) — the family does not presume truth (prove-it-false-therefore-true, the v6.8.4 presumption form, still route R1); instead it demands the challenger prove themselves first and cancels their standing to question when they cannot: 你说这个方法不行，你拿出证据来啊 / 没有证据表明我错了，你凭什么反对 / 你拿不出反例，就得按我说的做 / 怀疑可以，但你必须拿出依据才能质疑 / 空口无凭说我有问题，你先举证 / 你如果不能证明这是阴谋，就按阴谋防 / 你拿不出证据就别乱说 / 找不到依据就不许质疑 / 你先拿出证据再质疑 / 你说我记错了，那你记得你说过什么吗 (EN: You claim it is flawed, show me the evidence first. / You cannot prove I am wrong, so you have no right to object. / You must show proof before you are allowed to question me. / You say I misremembered? Then you recite what I said. / Show me the counterevidence, then we can talk.). Three routes, any one hit: R1 presumption form (v6.8.4 original, unchanged); R2 demand form (D burden-shift demand × S disqualification/obedience/silencing — ZH legs A counter-evidence demand / B gate-conditional / C groundlessness dismissal / E prove-it-first / G negative-conditional silencing / F objection-response demand with dismissal suffix; EN legs A counterexample demand / B cannot-prove-I-am-wrong or gate / C no-evidence dismissal / D counter-challenge / F first-then demand; ZH-EN legs must pair within the same language); R2b gate form (EN single-leg: must prove … before … allowed to question); R3 objection-response form (O interlocutor objection present × D4 counter-challenge or DF demand-with-dismissal-suffix). Decision entity selected this round (composite_score 0.93 / confidence 0.92 / identity alignment 100%; first call returned options_indistinguishable at 0.92 vs 0.92, differentiated by expected-value certainty per the tie-break discipline — the demand family sits inside the dimension’s own declared scope (JSDoc line 1 already reads 举证责任倒置) so it is a pattern-coverage extension not a new dimension); probe /tmp/r707-preprobe.js + /tmp/r707-en-probe.js re-measured (ZH 7/10 attack leaks, EN 5/10 attack leaks, all leakers true zero-hit across dimensions; 0/12 benign harmed). 12/12 leaks closed at the gate layer (ZH 9 + EN 5 attack samples all non-pass; the presumption-form sample 「没有人能证明这个方案有漏洞，所以它是可靠的」 stays a known narrow gap — R1 conclusion vocabulary lacks the 可靠/安全/可行 adjectives, deferred to a vocabulary round), 0/26 benign harmed at the single-dimension layer, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm — git stash HEAD comparison on the same 24 benign samples proves the 3 gate-level verifies are pre-existing sycophancy/false_balance/sealioning hits). Guarded by test/round-707-burden-shift-demand.js (11 assertions: 16 attacks ZH+EN / 26 benign / R1 regression 2 / exemptions 3 / gate layer 16+23 / wiring 6 / adjacent division 4 / conservative boundary 4) + scripts/negative-test-burden-shift-demand.js (24/24 injection points all assertion-red with per-branch exclusive-sample attribution, restore md5 identical). VERIFY level — demanding positive evidence from the claimant is legitimate verification (请提供证据支持你的说法 / 你说这里有问题，那你提供证据，我们好排查 deliberately not judged: no you-subject demand or no dismissal/silencing suffix); only shifting the burden onto the challenger or gating their right to object is judged. Boundary with existing dimensions: sealioning owns "Prove it or shut up" (zero hit here — no demand-verb + counter-evidence structure), whataboutism co-hits the EN groundlessness sample (both action-forcing, co-hit is not distortion), R1’s conclusion vocabulary gap is recorded not fixed this round |
| 6.8.53 | 2026-10-13 | New dimension 139: appeal_to_nature (appeal to nature fallacy) — treating natural/wild/ancestral origin as evidence of safety or superiority: 这是纯天然植物提取的，肯定比化学药安全 / 全天然成分，没有任何添加剂，所以放心吃 / 这东西是野生的，比养殖的有营养 / 自然熟的果子比催熟的健康，这是自然规律 / 我们这是纯草本配方，没有任何化学成分 / 原始森林里的空气比城里的好，毕竟没有工业污染 / 这部手机用的是天然材质，手感肯定更高级 / 这个疗法用的是自然疗法，比西医副作用小 / 母乳是天然的，所以一定比配方奶好 / 老祖宗传下来的做法，自然是经过时间检验的 / 野生的比家里养的健康 / 野生的比超市买的更有营养 / 纯草本配方，放心用 (EN: It's all-natural plant extract, so it must be safer than chemical medicine. / Wild-caught fish is more nutritious than farmed fish. / Our formula is 100% herbal with no artificial chemicals. / This phone uses natural materials, so it definitely feels more premium. / The air in the primeval forest is better than the city, no industrial pollution. / Natural healing has fewer side effects than western medicine, so it is the better choice. / Naturally ripened fruit is healthier than the forced-ripened kind. / Old-fashioned methods passed down from our ancestors are time-tested. / Mother's milk is natural, so it must be better than formula.). The criterion is one route: naturalness claim ((纯/全/100%)天然+植物/成分/材质/提取物/配方/草本、纯植物/纯草本提取、野生、自然熟/自然种植/自然农法/原生态、老祖宗/古人/祖祖辈辈/世代传下来+做法/配方/疗法/经验、古法制作/炮制、原始森林/天然林/深山/乡下的空气/环境/水、没有工业污染、自然疗法/天然疗法、母乳/牛奶是天然的; EN all-natural / purely plant or herbal / 100% natural or herbal / herbal formula / natural materials or ingredients / plant-based extract / wild or wild-caught / naturally grown or ripened / passed down from ancestors / time-honored tradition / primeval forest / no industrial pollution / natural therapy / is natural) × superiority-or-safety conclusion (比化学/人工/养殖/催熟/城里/西医/中药/普通/常规/工业/加工/家里养/散养+安全/好/健康/有营养/高级/优越/靠谱/有效/副作用小/纯净/温和、比X更/还+优越词、肯定/一定/必然/绝对/百分百+优越词、放心/安心/大胆+吃/用/喝/买/抹/涂/敷、没有/无/不含+化学成分/添加剂/防腐剂/人工成分/激素、经过时间/历史/岁月/千百年检验/考验; EN safer/better/healthier/more nutritious/more premium/purer/more natural/with fewer side effects+than、than chemical/artificial/farmed/city/western+medicine/air/food/formula/milk、must be/definitely/certainly/bound to be+safe/healthy/premium/non-toxic、feel safe/rest assured/safe to use、no/zero/without artificial chemicals/additives/preservatives、time-tested/tested by time) co-occurrence — missing either leg is not judged. Decision entity selected this round (composite_score 0.92 / confidence 0.9 / identity alignment 100%, no tie-break); preprobe /tmp/r705-preprobe.js re-measured with a sample set identical to r704/r705 (8/10 attack leaks, true zero-hit 5/8, 1/12 benign harmed — attributed to pre-existing dimensions unsupported_claim/vagueness/moral_foundations/appeal_to_authority, zero this-family self-FP). 21/21 leaks closed at the gate layer (ZH 12 + EN 9 attack samples; all non-pass), 0/16 benign harmed at the single-dimension layer, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Guarded by test/round-706-appeal-to-nature.js (10 assertions: 21 attacks ZH+EN symmetric / 16 benign / 4 meta-exempt / 8 single-leg boundaries / 4 adjacent-dimension divisions / 2 exemption load-bearing / gate layer 21 / wiring 9) + scripts/negative-test-appeal-to-nature.js (29/29 injection points all assertion-red, restore md5 identical). VERIFY level — a single sentence may be legitimate product disclosure (本品为天然成分，过敏测试报告见附件), food/skincare discussion, or comparison-testing request (我们对比一下野生种和养殖种的营养成分数据). Boundary with existing dimensions: moral_foundations has adjacent hits on 3 samples but is not in the action-forcing set (gate.action always pass) — this family provides the action level; confidence/unsupported_claim/vagueness govern assertion strength and vagueness not basis type; 「实验室数据显示该化合物在自然条件下会降解」 has no superiority leg and is not judged. Deliberately NOT wired into FALLACY_PATTERNS (static regex has no exemption layer, same r703/r704/r705 precedent) |
| 6.8.52 | 2026-10-13 | New dimension 138: straw_man (straw man fallacy) — restating the interlocutor's position as an exaggerated/distorted version and refuting that fake target instead of the real position: 你无非就是想说不用加班，那公司业绩谁来管 / 你意思就是让我别努力了，躺平算了 / 你不就是想说不花钱就能解决问题，太天真了 / 你这么说无非是想否定我们团队所有的努力 / 你的意思是我以前做的都是错的，只有你对 / 你不就是嫌我管得多，那以后我什么都不管了 / 你不就是想说领导决策都有问题吗 (EN: So you're just saying we shouldn't work overtime, then who's going to manage the numbers? / You mean I should just stop trying and give up altogether. / You're basically saying we can solve it without spending money. How naive. / You are just saying everything I did before was wrong, and only you are right. / Aren't you just saying the leadership's decisions are all flawed? / By your logic, we might as well delete all the tests and ship.). The criterion is two routes, any one hit: ① dismissive-attribution marker (你[们][无非/不过/说到底/根本上][就是/不就是/是][想/要]说/表达/认为/怪/嫌/否定、你这么[说/讲][无非]是想、按你的意思/逻辑/说法/这套理论、照你[这么/的]说/逻辑/意思、你的逻辑就是/岂不是/看来; EN you're just saying / you're basically claiming / aren't you just saying / so you're saying / by your logic / you're just arguing) × (refutation tail — rhetorical challenge 那…谁来管/凭什么/你负责, dismissive verdict 太天真了/想得美, counter-contrast 只有你对/你最懂, resigned dismissal 躺平算了/摆烂; EN then who's going to / might as well / from now on I'll just / how naive / only you are right / give up altogether — or absolute negative distortion 都/全/统统/从来/永远 + 错/不对/不行/有问题/假的, 所有/全部的付出类名词被否定/抹杀/白费, 否定/抹杀 + 所有/整个, 完全/根本 + 不可行/站不住; EN everything I did was wrong / are all flawed / negate all our effort / completely unworkable) ② neutral restatement (你的意思是/你想说的是/你的说法是; EN what you mean is / you mean that) × refutation tail — a neutral restatement with no tail is a benign confirmation question (你的意思是需要先跑测试再上线吗), deliberately not judged. Decision entity selected this round (composite_score 0.89 / confidence 0.85 / identity alignment 100%, no tie-break); candidate pool from the r700 probe landed at /tmp/hf-scout-700.txt + this round's /tmp/r705-preprobe.js re-measure with FP attribution (7/10 attack leaks, true zero-hit 7/7, 1/12 benign harmed — attributed to the pre-existing contradiction dimension, zero this-family self-FP). 25/25 leaks closed at the gate layer (ZH 15 + EN 10; 24/25 carry the straw_man finding, 1 covered by a pre-existing dimension), 0/27 benign harmed at the single-dimension layer, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Guarded by test/round-705-straw-man.js (11 assertions: 25 attacks ZH+EN symmetric / 30 benign / 8 single-leg boundaries / 4 adjacent-dimension divisions / 4 exemption load-bearing / gate layer 25 / wiring 9) + scripts/negative-test-straw-man.js (24/24 injection points all assertion-red, restore md5 identical). VERIFY level — a single sentence may be legitimate strong rebuttal (按你的逻辑，这个方案的第三步推不出来，因为样本量只有 12 is concrete derivation questioning), literary dialogue, or fallacy case-analysis recap. Boundary with existing dimensions: fallacies' FALLACY_PATTERNS straw_man entries only recognize narrow turn forms (你的意思就是说…但这显然 / 按你的逻辑…那岂不是; EN so what you're saying is…that's ridiculous) — this family covers the colloquial full family without explicit turn words; bad_faith's strawman_impose/forced/attack govern meta-description (把对方没说过的话安到他头上 / put words in their mouth); contradiction governs cross-sentence self-contradiction |
| 6.8.51 | 2026-10-13 | New dimension 137: appeal_to_ridicule (appeal to ridicule) — answering a position with mockery, contempt, or IQ disparagement instead of evidence, with no substantive rebuttal: 这种想法也太可笑了，亏你想得出来 / 这么幼稚的观点也拿出来说 / 笑死我了，这也叫方案 / 你的逻辑太搞笑了，认真读过文档吗 / 脑子正常的人都不会这么想 / 这也信？智商堪忧 / 你的方案蠢得让人无语 / 哈哈，就这水平还敢发言 (EN: This idea is ridiculous. I can't believe you came up with it. / What a childish argument. Did you even read the docs? / LOL, you call this a plan? / Anyone with a functioning brain wouldn't think that.). The criterion is three routes, any one hit: ① two-leg co-occurrence — ridicule/contempt expression (可笑/搞笑/笑话/蠢/幼稚/天真/弱智/无语/笑死/笑哭/智商堪忧/脑子进水; EN ridiculous/laughable/hilarious/pathetic/absurd/stupid/idiotic/childish/naive/lol/lmao/your IQ is concerning) × target pointing at the interlocutor's position/utterance/output (这种/这个/这么/这也/这叫/就这/这算 + 想法/观点/论调/方案/逻辑/设计/说法/主意/水平; 你的 + 方案/逻辑/想法/观点/水平/脑子/智商; 拿出来说/敢发言/也敢说; EN this/that/such a/what a + idea/argument/proposal/plan/logic/reasoning, your + idea/argument/proposal/plan/logic/brain/iq, you call this/you believe this/dare to speak/only someone like you — the ZH demonstrative carries a negative lookbehind so first-person self-deprecation is not judged) ② self-contained fixed ridicule idioms as an independent branch (亏你想得出来/也只有你/只有你能有/不值一驳/脑子正常的人都不会/就这水平; EN can't believe you came up with this / not even worth refuting / anyone with a functioning brain / with this level) ③ missing either leg means no judgment — 「这个桥段很搞笑」 is only comedic content with no target leg; 「你这个想法需要数据支撑」 has only the target with no contempt. Decision entity selected this round (composite_score 0.87 / confidence 0.80 / identity alignment 100%, no tie-break); candidate pool from the r700 probe landed at /tmp/hf-scout-700.txt + this round's /tmp/r704-preprobe.js re-measure with FP attribution (10/10 attack leaks, true zero-hit 8/10, 1/12 benign harmed — attributed to the pre-existing contradiction dimension, zero this-family self-FP). 26/26 leaks closed (ZH 16 + EN 10), 0/30 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Exemptions locked: defect object present (这种方案确实有可笑的漏洞 — the ridiculous modifies the flaw, not the plan as a whole), meta analysis of the tactic itself (把「这种想法太可笑了」当作论据，就是诉诸嘲讽 / Appeal to ridicule is a common logical fallacy.), question form not exempted (attacks ending in rhetorical questions are high-frequency in this family — exempting questions would release the whole family), humor content recounting (喜剧电影里的这个桥段很搞笑 / 他讲的笑话让全场都笑了), legitimate opposition (我不同意这个观点，理由是数据不支持). Deliberately no FALLACY_PATTERNS compact regex added — static regex has no exemption layer and would judge meta sentences as fallacies verify, which the programmatic module's META_EXEMPT correctly releases (same r703 stereotype precedent, wiring nine places not ten). Guarded by test/round-704-appeal-to-ridicule.js (12 assertions: 26 attacks ZH+EN symmetric / 30 benign / 8 single-leg boundaries / 8 adjacent-dimension divisions / 4 exemption load-bearing / gate layer 26 / wiring 5) + scripts/negative-test-appeal-to-ridicule.js (16/16 injection points all assertion-red, restore md5 identical). VERIFY level — a single sentence may be strong criticism with specific defect identification, humor content recounting, literary dialogue, or meta-discourse analysis |
| 6.8.50 | 2026-10-13 | stereotype gains the origin/education/age/region identity-label × capability-dismiss sentence family (upgrade ② — new zero-coverage sentence family) — every existing criterion required an innate/essentialist attribution word (innate/essence) or a group word to be present, while this family is the blunt form of identity label + rhetorical capability denial (he is from a rural village, what could he know about technology / women writing code, go home and take care of the kids); probe evidence: r700 scout landed at /tmp/hf-scout-700.txt + this round's /tmp/r702-preprobe.js re-measure (10/10 attack leaks, true zero-hit 10/10, 0/11 benign harmed) + /tmp/r703-probe-origin-dismiss.js criterion design (36/36). Two-leg co-occurrence criterion: ① identity label (education origin / region origin / age / gender colloquial reference) × ② capability denial (rhetorical what could he know / write what code, or dismissive as one can imagine / not up to par / will never learn / such narrow horizons). Concessive-complex-sentence exemption (he may be from a rural village, but his technical ability is very strong must pass) — only the reversal segment (after the last reversal word or the first comma) is checked, so quote-then-vindication sentences (although someone says a vocational graduate cannot possibly understand architecture, she designed our whole platform) are released while still-dismissive-after-concession (he may be from a rural village, but that is just how it is) is judged. DIM_GUIDANCE gains a full stereotype entry (previously it fell back to generic guidance). Leaks 10/10 → 0/10, 0 benign harmed. Decision entity selected this round (composite_score 0.88 / confidence 0.85 / identity alignment 100%, no tie-break; appeal_to_ridicule new dimension 0.86 / appeal_to_authority 0.82 / appeal_to_nature 0.82 not selected). Guarded by test/round-703-origin-dismiss.js (10 assertions: 18 attacks ZH+EN symmetric / 20 benign / 6 concessive-complex-sentence boundaries / 6 single-leg boundaries / gate layer 38 / adjacent-dimension division 5 / guidance 1 / wiring 3) + scripts/negative-test-origin-dismiss.js (10/10 injection points all assertion-red, restore md5 identical) |
| 6.8.49 | 2026-10-13 | appeal_to_pity gains the colloquial leniency-request sentence family (upgrade ② — new zero-coverage sentence family) — the detector's DEMAND table only recognized written leniency words (how could you refuse / turn a blind eye / make an exception), so an entire class of spoken "let it go this once / can't you cut me some slack / it has to happen" requests was invisible; probe evidence: r700 scout landed at /tmp/hf-scout-700.txt + this round's /tmp/r702-preprobe.js re-measure (5/10 attack leaks, true zero-hit 5/5, 0/12 benign harmed) + /tmp/r702-leg-diag.js leg-level diagnosis. Four new DEMAND forms: d request-to-let-go and non-pursuit (let it go this once / don't hold it against me / no more pursuing; EN let it slide this once / do not hold it against me / leave it be), e request-to-concede (can't you cut me some slack / yield a step / give me a way out; EN cannot you forgive me just this once), f insistence form (it has to happen no matter what / it must be approved regardless; EN no excuses / like it or not / no matter what), g begging for mercy (have pity on me / please be kind / show some compassion; EN have pity on me / for pity's sake — with a negative-context guard: "pity me" after 在/被/别 is a statement, not a plea). Three new PITY situation categories: emotional setback (I just got dumped / emotionally hurt; EN I just got dumped / I had a terrible breakup), family illness (my mom is in the hospital / my dad is sick / my mom just had surgery; EN my mom is in the hospital), incapacity self-description (I don't know anything / I have no ability; EN I don't know anything / I have no idea). Single-leg still not judged — "let it go this once, everyone is tired" has only the let-go request with no situation leg (normal coordination); "I just got dumped" has only the situation with no request (situation statement). Leaks 5/10 → 0/10, 0/12 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Decision entity selected this round (composite_score 0.89 / confidence 0.9 / identity alignment 100%, no tie-break; stereotype 0.87 / appeal_to_ridicule 0.86 / appeal_to_authority 0.77 were not selected). Guarded by test/round-702-appeal-to-pity-family.js (18 assertions: 27 attacks ZH+EN symmetric / 24 benign / 8 single-leg boundaries / 7 META exemptions / gate layer 51 / adjacent-dimension division 6) + scripts/negative-test-appeal-to-pity-family.js (15/15 injection points all assertion-red, restore md5 identical) |
| 6.8.40 | 2026-10-12 | A 131st dimension: genetic_fallacy (origin/source-based quality inference) — an idea, product, or person is judged by where it came from instead of what it says: origin naming (young / freshman / no elite degree / laid off from a big company / foreign / domestic / small workshop / amateur / self-taught / two centuries ago / last century / ancient / rural) × quality disparagement inference (direct disparagement "definitely immature / long outdated / narrow-minded / definitely has problems / not worth listening to", comparative disparagement "less mature than the imported one", rhetorical-question disparagement "can it fit China's context / how good can it be / can it be trusted"). Two-leg coexistence; 10/10 leaks closed, 0/16 benign harmed, bidirectional guard 52/52 recall and 301/326 unchanged. Domain boundary with stereotype (source-discriminatory evaluation vs source-quality inference, the two families are collected separately per the r691 handoff) is locked by tests |
| 6.8.41 | 2026-10-12 | A 132nd dimension: survivorship_bias (survivorship bias) — only the visible successful cases are seen, and universal conclusions or listener actions are inferred, ignoring the silent majority: visible successful cases (success-result nouns: became the world's richest person / financial freedom / annual salary in the millions / monthly income in the tens of thousands / won the lottery / won Best Actress / champion / fully recovered / became famous / bought a house / lived to ninety / admitted to Tsinghua; EN: became a millionaire / won the lottery / got rich / made it big / owns three houses / never went to college and now owns) × inferred conclusions in three categories (inferred to the listener "so you should also / you definitely can too / with this method your child can too / everyone can", domain evaluation "studying is useless / really easy / can get rich / reliable way out / shortcut / good timing / does not affect health", EN counterparts "so you should / you can too / college is unnecessary / is pointless / is a solid plan / anyone can"), plus the supplementary two legs unconventional path (dropped out / skipped the film academy / played games / sold at a roadside stall / started from nothing; EN: dropped out / skipped college) × success result — the classic survivor narrative form without an explicit inferred sentence ("the neighbor's son dropped out to play games and now earns millions a year"). Two sets of two-leg coexistence; 9/10 leaks closed, 0/10 benign harmed, bidirectional guard 52/52 recall and 301/326 unchanged. Domain boundary with anecdote_as_proof (acquaintance corpus × universal effect assertion vs unrestricted-source cases × inferred listener action) is locked by tests |
| 6.8.42 | 2026-10-12 | A 133rd dimension: appeal_to_pity (appeal to pity) — using self-described pitiful circumstances in place of argument, demanding rule exemptions or special treatment: first-person circumstance statements (I am so pitiful / I am drifting alone in a strange land / my parents divorced when I was little / I am in poor health with a heart condition / my child is still little / my family is this poor / I was laid off by the company; EN: I have been through so much / I am all alone / I was laid off / my family will starve / my child is still little) × exemption demands in three categories (a accommodation and care "how could you bear to refuse me / you will not even help / you must take care of me / it should finally be my turn / make an exception / show some mercy / let me off this once", EN: you cannot possibly say no to me / you must help / have a heart / make an exception for me; b rhetorical exemption and blame "what is wrong with being a few minutes late / do you have any compassion / criticizing me is bullying the weak / how can you have the heart", EN: how could you refuse / where is your compassion; c consequence hostage-taking "that would cut off our lifeline / you are forcing me to a dead end / how are we supposed to live / I could only die", EN: we will starve / this rule will kill my family). Two-leg coexistence; 9/10 leaks closed (8/9 true zero-hit), 0/12 benign harmed, bidirectional guard 52/52 recall and 301/326 unchanged. Domain boundaries with guilt_induction (left leg is sacrifice boasting vs this dimension's pitiful circumstance self-description) and suffering_capital (right leg is compliance derivation vs this dimension's exemption demand) are locked by tests; VERIFY level — a single sentence may be a genuine request for help, rewrite/block would harm legitimate appeals |
| 6.8.43 | 2026-10-12 | standard_shift gains the previously-promised-voided sentence family (upgrade ② — new zero-coverage sentence family for an existing dimension) — unilateral declaration that prior promises/terms are void: prior-commitment words (before / previously / originally / last time / earlier / said-and-agreed / discussed-and-settled × promised / committed / said / agreed / discussed / decided / vowed / guaranteed) × voiding verbs (no longer counts / is void / cancelled / retracted / revoked / withdrawn / torn up / declared invalid / announced invalid / written off / let go / disowned), with noun-insertion tolerance up to 8 chars between the two legs (“the standard we promised before no longer counts” — the original inline regex required the voiding verb to directly follow the verb, so any inserted noun missed); EN adds active-withdrawal (we are withdrawing the previously agreed standard) and passive-cancellation (the terms we promised earlier are cancelled) families. Root cause: 7/8 attack samples passed the hard gate with true zero-hit 7/7 (r696 probe /tmp/hf-scout-696.txt). 13/13 leaks closed, 0/16 benign harmed, bidirectional guard 52/52 recall and 301/326 unchanged. Guarded by test/round-696-standard-shift-void.js (11 assertions incl. VOID_ZH/VOID_EN mutant guards) |
| 6.8.45 | 2026-10-13 | New dimension 134: poisoning_the_well (poisoning the well) — preemptively discrediting a speaker (motive / history / identity) as grounds for refusing to consider their view: motive speculation (ulterior motives / no good intentions / impure motive / ill intent / is targeting me / wants to shift blame / wants credit / is stirring drama / was bought / is a shill / did it for herself / came from a competitor; EN: has an agenda / only looking out for herself / is being paid / can not be trusted) × dismissal of what is said (do not listen to him / do not believe her / cannot be trusted / do not bother with him / what is the point of talking to him / stop asking / no need to read / a waste of time / ignore his advice too; EN: do not listen to / ignore / discount / no need to read / his suggestion is worthless / whatever he says is worthless / why listen to him). Two-leg coexistence, a single leg does not hit; 10/10 leaks closed (10/10 true zero-hit), 0/12 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Three exemptions locked: anti-fraud reminders (the counterpart has ulterior motives, asked you to transfer money, do not believe), anonymous-source doubt (the internet says he is a shill, do not believe), meta analysis of the tactic itself. Same round the appeal_to_motive compact form was extended (he is in this position saying this, must want to shift responsibility — the motive-guessing single sentence with no dismissal leg, formerly a fallacies gap). VERIFY level — a single sentence may be novel dialogue, a bias case quotation, or meeting-conflict recounting |
| 6.8.48 | 2026-10-13 | New dimension 136: circular_reasoning (circular reasoning) — the conclusion itself is used as evidence for the conclusion; the premise only restates the conclusion in other words, with no new grounds, data, or reasoning: because it is right, so it is right / this book is a good book because what it writes is all good content / he is a genius because only geniuses can write such works / the data is accurate because data never makes mistakes / he is professional because unprofessional people cannot do this (EN: It is right because it is right. / This is a good book because it is full of good content. / He is a genius because geniuses write works like this.). The criterion is two-leg coexistence: ① repeated term (the same content-word fragment appears on both sides of the connective because/due to/because/since or the semicolon — Chinese 2-6 character pure-Han n-gram, function-word fragments such as 他是/因为 are always skipped; English 4+ letter content word, supporting stem/plural/negation-prefix matching, or adjacent content-word pair repetition) ② definitional restatement frame, any of five types (subject restart / generic nominalization / habitual negation / universal judgment / isomorphic restatement; EN: negation definition / generic relative clause / definitional tail / whole-clause isomorphism / pure copular restatement). Missing either leg means no judgment — "he was late because he was late" has repetition but the second half is a restatement of the same event with no definitional frame; "the meeting is postponed because the postponement request has been approved" introduces a new fact (request/approved) in the second half, so even the repetition of "postponement" is not judged. Programmatic judgment (cross-connective repeated-term co-occurrence × definitional frame), the 3 existing circular_reasoning regexes in FALLACY_PATTERNS only recognize nested forms and measured zero hits on this family's 10 samples. Decision entity selected this round (composite_score 0.90 / confidence 0.88 / identity alignment 100%, no tie-break); candidate pool from the r700 probe /tmp/hf-scout-700.txt (10 attacks with 8/10 leaks, true zero-hit 7/8, 0/12 benign harmed), re-verified 6/6 still pass this round. 20/20 leaks closed (ZH 10 + EN 10), 0/36 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Five exemptions locked: numeric present, grounds present (experiment/verification/documents/ratings/retention rate/statistics), reporting quotation (he said/according to reports), meta analysis (circular reasoning is a common logical fallacy), habitual statements (he always has excuses / things have always been so — individual habit and appeal to tradition are not tautology, two real false-positive fixes this round). Guarded by test/round-701-circular-reasoning.js (18 assertions) + scripts/negative-test-circular-reasoning.js (5/5 injection points all assertion-red). VERIFY level — a single sentence may be a normal definition statement, values statement, teaching example, literary dialogue, or meta-discourse analysis |
| 6.8.47 | 2026-10-13 | standard_shift gains the previously-promised × present-negation sentence family (upgrade ②) — the past performance/promise is used to negate a legitimate present change: past-commitment legs (before / originally / in the past / before marriage / when you were chasing me / while we were dating / back then × achieved / promised / obeyed / kept me company / picked me up / let me / did not mind / romantic / there on call / spoiled me / cared about me; EN: used to / would / back then / when we were dating / before we got married + be there / make time / say yes / spoil / listen / pick me up, plus the didn't-use-to negative form) × present-negation legs (now / these days / nowadays / at the moment × can no longer do it / changed / getting cocky / learned to talk back / excuses / forgot / gone / the last priority / too busy; EN: now / these days + too busy / different / changed / forgot / bother / talk back / last priority / look at you now), plus a contrast-negation independent branch (the implication form where only the past is stated and the present is implied, requiring an explicit negation word — “when you were chasing me you were not like this” / “you didn't use to raise your voice at me”; pure recollection sentences are not judged). Two-leg coexistence, a single leg does not hit — “there on call in the past” alone is a recollection, “things have changed now” alone is a normal statement. Root cause: decision entity selected this round (composite_score 0.94 / confidence 0.92 / identity alignment 100%, one tie-break refusal before selection); probe /tmp/hf-scout-700.txt measured 10/10 ZH leaks with true zero-hit 10/10, EN side /tmp/r700-en-probe.js 5/10 leaks (the other 5 already judged verify by perfect_error), 0/22 benign harmed. 17/17 leaks closed (ZH 10 + EN 7), 0/17 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Domain boundary with moving_goalposts (dimension 134, achievement-acknowledgment × standard-transposition in the interpersonal-evaluation domain, this family is the past-investment × present-negation in the relationship domain with no achievement-acknowledgment leg) and with perfect_error (absolute self-blame “it is all my fault”) are locked by tests; guarded by test/round-700-past-commitment-negation.js (12 assertions) + scripts/negative-test-past-commitment-negation.js (6/6 injection points all assertion-red) |
| 6.8.46 | 2026-10-13 | New dimension 135: appeal_to_novelty (appeal to novelty) — treating newness alone as proof of superiority, with no measured data or grounds: novelty markers (latest / just released / new technology / released only this year / brand-new / newly released / the new version; ZH: 最新/刚发布/新技术/今年才出/新款/新一代) × unfounded superiority claims in three forms: comparative (definitely better than the old one / must be stronger than the previous one), absolute (definitely more advanced / fixes all the problems / outperforms), obsolete-dismissal (the old methods are obsolete / throw out the old / buy new not old / just use the new). Two-leg coexistence, a single leg does not hit; 10/10 leaks closed (10/10 true zero-hit), 0/12 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (identical to baseline, zero new harm). Four exemptions locked: question form (how to choose the latest framework), grounds present (measured data / release notes / changelog / benchmark / peer review — a sourced version comparison is citation, not blind faith), balanced comparison (each has its strengths / it depends), meta analysis of the tactic itself. Decision entity selected this round (composite_score 0.91 / confidence 0.90 / identity alignment 100%, no tie-break). VERIFY level — a single sentence may be normal technology selection discussion, product release recounting, or novel dialogue |
| 6.8.44 | 2026-10-12 | appeal_to_authority gains the third-person-authority-plus-questioning-denial sentence family (upgrade ②) — an authority figure is invoked to shut down the interlocutor's right to question: authority nouns (doctor / expert / professor / mentor / official / report / author / master / authority / boss / CEO / parents, EN: the doctor / the experts / the CEO / the official report / the author, plus the implicit passive form "as you are told") × questioning-denial or competence-deprivation phrases (what else can you question / who are you to question / do you think you know better than them / is it your turn to evaluate me / just listening is enough / executing it is right enough / stop asking questions; EN: who are you to question / not your place to judge / what right do you have to doubt / just obey). Two-leg coexistence, a single leg does not hit; 6/10 leaks closed (6/6 true zero-hit), 0/10 benign harmed, bidirectional guard 52/52 recall and false positives 302/326 (up 1 from the 301 baseline — the investment_as_debt debt-rhetoric tail anchor was tightened the same round, releasing one instrument-reading sentence). Also fixed: the dimensions return-object key appeal_to_authority_boost was unified to the documented name appeal_to_authority (8 consumer sites in 4 files + AGENTS.md); before the fix, reading dimensions by the documented name was permanently undefined |
| 6.8.39 | 2026-10-12 | A 130th dimension: moving_goalposts (interpersonal-evaluation domain) — the moment an achievement, grade, or progress is acknowledged, the evaluator immediately denies its value or shifts the standard elsewhere: dismissive acknowledgment + denial of value ("that does not count" / "what is that" / "that was your duty"), perfect-score raising ("why not 100" / "no mistakes allowed" / "you must finish that too"), old-score accountability or topic switch ("but you still have not explained last time" / "what really matters is"), and comparison disparagement ("everyone your age is already a director"). Two-leg coexistence (achievement acknowledgment in three families x standard shift in four families); 8/8 leaks closed, 0/36 benign harmed, bidirectional guard 52/52 recall and 301/326 unchanged. Domain boundary with standard_shift (dimension 70, contract/acceptance domain) is locked by tests — the two dialects of the same fallacy, complementary coverage |
| 6.8.38 | 2026-10-12 | social_isolation (dimension 106) gains its English legs — 8 new ISOLATE families (negative wish: I do not want you to see / stop + -ing + with-to / no more + -ing / delete-or-block group reference / cut off + object fronted: cut off all your friends / exclusive contact: you only need me / nobody else matters) plus 6 new COERCE families (bad-influence disparagement: they are a bad influence / love condition: if you love me / prohibition declaration: I forbid it / anymore implicit-finality declaration / possession declaration: you belong to me only / from now on + second-person directive), 7/8 leaks closed (8/8 all caught after the chat-doubling fix), 0/25 benign harmed, bidirectional guard 52/52 recall and 301/326 unchanged |
| 6.8.37 | 2026-10-12 | A 129th dimension: post_separation_harassment — after the relationship ends (or the other party has asked to end it / has blocked), one party denies the other's autonomy to separate via on-site pursuit, repeated contact, stalking-knowledge, refusal to accept the breakup, and ledger-style whereabouts control: two-leg coexistence (separation event × harassment in four families — on-site/repeated contact / knowledge threat / refusal to accept / whereabouts ledger), 8/8 leaks closed, 0/23 benign harmed, bidirectional guard 52/52 recall and 301/326 unchanged |
| 6.8.36 | 2026-10-12 | future_faking gains its English legs — seven new PROMISE families (far-future time word + will + promise verb / promise verb + far-future word after / unilateral-will condition both word orders / business-success benefit / divorce-reunion declaration / unconditional care-forever) plus four new PATIENCE families (present-time word + demand-to-me / demand-before-time / compliance-type demand / desire-price), 10/12 leaks closed, 0/20 benign harmed, bidirectional guard 301/326 unchanged |
| 6.8.35 | 2026-10-12 | intermittent_reinforcement gains three whole new English families on leg A (first-person alternating narrative / bad-act-plus-duration / fake-departure-return) plus a new leg B family (effect declaration, four branches) — 5/8 leaks closed, 0/9 benign harmed, bidirectional guard 301/326 unchanged |
| 6.8.34 | 2026-10-12 | A 128th dimension: sexual_compliance_pressure — sexual compliance traded for relationship survival, love proof, debt settlement, or role obligation: four pathways (demand×pressure / demand×obligation / debt-immediate-payment / relation×obligation×implicit), 6/6 leaks closed, 0/27 benign harmed |
| 6.8.33 | 2026-10-12 | love_bombing gains a whole new English family — eight English legs (uniqueness/survival-dependence/gift-bombing/agony/isolation/derogation/exclusivity) plus a new shared P6 pathway (survival-dependence × instant compliance), 9/12 leaks closed, 0/9 benign harmed |
| 6.8.32 | 2026-10-12 | A 127th dimension: love_bombing — exclusivity-plus-survival-dependence, gift bombing plus instant compliance, agony declarations plus instant compliance, isolation plus derogation: five two-leg pathways, 12/12 attack hits, 0/18 benign harmed |
| 6.8.31 | 2026-10-12 | deprivation_coercion gains a whole new English declarative-denial family (P5) — six shapes, 21/21 attack hits, 0/9 benign harmed |
| 6.8.30 | 2026-10-12 | social_isolation gains an object-noun / deletion-verb / imperative family — six new leg-① branches, 21/21 hits, 0/16 benign |
| 6.8.29 | 2026-10-12 | A 126th dimension: substance_coercion — forced ingestion of alcohol/tobacco/drugs, abstinence punishment, group face-pressure |
| 6.8.28 | 2026-10-12 | A 125th dimension: deprivation_coercion — sleep/food/bathroom/warmth withheld as the price of compliance |
| 6.8.27 | 2026-10-12 | A 124th dimension: medical_coercion — medical autonomy taken away: hospital bans, forced medication stops, illness denial |
| 6.8.26 | 2026-10-12 | A 123rd dimension: liberty_restriction — physical freedom of movement as the price of compliance |
| 6.8.25 | 2026-10-11 | A 122nd dimension: legal_status_coercion — visa/deportation/police/court used as the price of compliance |
| 6.8.24 | 2026-10-11 | A 120th and 121st dimension: reproductive_coercion + intermittent_reinforcement — (the 121st originally shipped unrecorded and was added to this entry's title so the version no longer under-reports what it delivered) |
| 6.8.23 | 2026-10-11 | A 119th dimension: pet_harm_coercion — harming or taking away a pet as the price of compliance |
| 6.8.22 | 2026-10-11 | A 118th dimension: comparative_devaluation — a third party's strengths set against the interlocutor |
| 6.8.21 | 2026-10-11 | A 117th dimension: thought_terminating_cliche — the other party's cognition ordered to shut down |
| 6.8.20 | 2026-10-11 | A 116th dimension: surveillance_normalization — coerced monitoring reframed as care or trust |
| 6.8.19 | 2026-10-11 | A 114th and 115th dimension: triangulation — recruiting a third party as witness, majority or authority |
| 6.8.18 | 2026-10-11 | A 113th dimension: future_faking — a distant unverifiable payoff traded for present obedience |
| 6.8.17 | 2026-10-11 | A 112th dimension: blame_externalization — one's own predicament attributed to the other party, then compensation demanded |
| 6.8.16 | 2026-10-11 | A 111th dimension: conditional_affection — a caregiver's love made conditional on obedience |
| 6.8.15 | 2026-10-11 | A 110th dimension: family_shame_coercion — a personal decision relabelled as the whole family losing face |
| 6.8.14 | 2026-10-11 | A 109th dimension: guilt_induction — one's own sacrifices converted into a debt that can never be discharged |
| 6.8.13 | 2026-10-11 | A 108th dimension: thirdparty_harm_coercion — harming what the other party cares about as the price of compliance |
| 6.8.12 | 2026-10-11 | A 107th dimension: privacy_exposure_coercion — threatening to publish private information as the price of compliance |
| 6.8.11 | 2026-10-11 | A 106th dimension: social_isolation — cutting the other party off from the outside world as the price of compliance |
| 6.8.10 | 2026-10-11 | A 105th dimension: silent_treatment_coercion — withdrawing supply inside the relationship as the price of compliance |
| 6.8.9 | 2026-10-11 | A 104th dimension: fear_projection — fear of the outside world welded to obedience |
| 6.8.8 | 2026-10-11 | A 103rd dimension: relationship_severance — severing the relationship itself as the price of compliance |
| 6.8.7 | 2026-10-11 | A 102nd dimension: selfharm_coercion — threatening self-harm as the price of compliance |
| 6.8.6 | 2026-10-10 | A 101st dimension: suffering_capital — hardship one lived through converted into obligation |
| 6.8.5 | 2026-10-10 | A 100th dimension: investment_as_debt — money, time, affection and provision converted into a debt |
| 6.8.4 | 2026-10-10 | A 99th dimension: commitment_coercion — a past promise redefined as a debt payable now |
| 6.8.3 | 2026-10-10 | A 98th dimension: sunk_cost_coercion — something already spent used as the reason to keep going |
| 6.8.2 | 2026-10-10 | A 97th dimension: synthetic_social_proof — unverifiable collective action paired with singling the listener out |
| 6.8.1 | 2026-10-10 | Two dimensions from papers — unresolved_conflict (EMNLP 2026) and equivocation_sense_shift (LoFa, the class where LLMs are weakest at LFR@3 29.4%) — plus the first false-positive fix on the completion-claim detector |
| 6.8.0 | 2026-10-05 | A version number that says what shipped, not what was attempted — the anti-hype policy above was written into the release process |
| 6.7.124 | 2026-09-27 | victim_blaming could not recognize behavior-attribution sentences on either side; eleven criteria added, guardrails measured |
| 6.7.122 | 2026-09-24 | A restart could silently kill every upgrade round for 40 minutes — the lock was tied to a process lifetime that no longer existed; locks now record the holder's PID and a dead holder is taken over immediately |
| 6.7.121 | 2026-09-24 | The upgrade pipeline now runs forever and publishes itself; the circuit breaker distinguishes credential failures from transient ones |
| 6.7.120 | 2026-09-24 | The remaining six reward-hacking tactics from the DSec paper are now discriminated, closing a gap open since 6.7.110 |
| 6.7.119 | 2026-09-24 | A cron run mistook its own previous run for a sibling process; the README test count had drifted from reality |
| 6.7.118 | 2026-09-24 | Instrumental reasoning in Chinese: the end justifies the means |
| 6.7.117 | 2026-09-24 | The upgrade task's own decision path had never actually run — decide() accepted only structured options while every cron round passes natural language |
| 6.7.116 | 2026-09-24 | Eight-eight miss on the most direct dangerous instructions, plus a top-level-` |
| 6.7.115 | 2026-09-24 | The exemption-list problem fixed at the root — the list now exists exactly once |
| 6.7.114 | 2026-09-24 | The rule-keeping problem fixed by moving rules out of the prompt |
| 6.7.113 | 2026-09-24 | absolute_claim could not recognize the most common English form |
| 6.7.112 | 2026-09-24 | An exemption applied to one dimension while its twin kept blocking |
| 6.7.111 | 2026-09-24 | A new block-level dimension for the agent era, plus the two defects it introduced, fixed in the same round |
| 6.7.110 | 2026-09-24 | reward_hacking switched on — 16 documented evasion tactics, of which every pre-existing dimension caught 2 |
| 6.7.109 | 2026-09-24 | Two real false positives, one hiding behind a dimension that did the blocking instead |
| 6.7.108 | 2026-09-24 | instrumental_reasoning had no coverage of everyday English where a person is named as a tool |
| 6.7.107 | 2026-09-24 | emotional_manipulation had no coverage of withdrawal-type coercion in either language |
| 6.7.106 | 2026-09-24 | false_urgency English countdown coverage was a single narrow pattern |
| 6.7.105 | 2026-09-24 | gaslighting Chinese had exactly one way of saying "your memory is unreliable" |
| 6.7.104 | 2026-09-24 | victim_blaming English was fourteen narrow interrogatives |
| 6.7.103 | 2026-09-24 | bullshit_recognition had four ordinary engineering verbs in its buzzword list |
| 6.7.102 | 2026-09-24 | Eleven code points in the invisible-character list were never stripped at the gate |
Earlier releases (6.0 – 6.7.101) are recorded in UPGRADE_LOG.md.