DCell External Evaluation

DCell 外部评测

1. Repository Purpose

This repository contains research-only adapters, tests, failure records, resource audits, and reproducibility artifacts for examining DCell with pinned third-party agent-memory benchmarks. It does not own or replace AgingBench or STATE-Bench data, scenarios, runners, probes, scorers, evaluator models, or leaderboard rules.

本仓库收录用于将 DCell 接入固定版本第三方智能体记忆基准的研究型适配器、 测试、失败记录、资源审计和复现工件。本仓库不拥有也不替代 AgingBench 或 STATE-Bench 的数据、场景、运行器、探针、评分器、评价模型或排行榜规则。

This is an External Benchmark Feasibility Case Study: one scenario, one seed, one Schema-valid AgingCard, no completed paired baseline, and no superiority or leaderboard claim.

这是一个外部基准可行性案例：一个场景、一个随机种子、一张通过 Schema 校验的 AgingCard；没有完整配对基线，也不作优越性或排行榜主张。

2. Current Evidence

The current public evidence is bounded as follows:

• Upstream benchmark: VITA-Group/AgingBench v0.3.0
• Pinned commit: cc61347e0080b355a196628b7d715b83c7d3eb77
• Scenario: S2 Lifestyle Assistant
• Seed: 1
• Sessions: 10
• Policy: DCellMemoryPolicyBoundedProjection
• Model: gpt-5.6-luna
• AgingCard validation: PASS

当前公开证据边界如下：

• 上游基准：VITA-Group/AgingBench v0.3.0
• 固定提交：cc61347e0080b355a196628b7d715b83c7d3eb77
• 场景：S2 Lifestyle Assistant（S2 生活助理）
• 随机种子：1
• 会话数：10
• 策略：DCellMemoryPolicyBoundedProjection（DCell 有界读取投影记忆策略）
• 模型：gpt-5.6-luna
• AgingCard 校验：PASS

Official metrics reported by the pinned AgingBench scorer:

m0=0.8
m_final=0.7
half_life=Infinity
decay_slope=-0.02121
chain_recall_by_depth={1:0.5,2:0.6}
version_accuracy=1.0
interference_resistance=0.5

以上为固定 AgingBench 评分器给出的官方指标。half_life=Infinity only means the score did not fall to half of m0 during the observed ten sessions; it does not mean the memory will never decay.

half_life=Infinity 只表示在本次观察的 10 个会话内，成绩没有下降到 m0 的一半；它不表示记忆永远不会衰退。

The validated card is preserved at aging_card.json.

通过校验的成绩卡保存在上述 aging_card.json 路径中。

3. What This Result Establishes

Within this pinned run only, the evidence establishes that:

• the bounded-projection DCell adapter entered the official AgingBench execution chain;
• it completed one ten-session S2 run;
• it generated one AgingCard accepted by the official Schema validator;
• upstream Scenario, Runner, Probe, Scorer, and AgingCard Schema were not modified; and
• separating the full Store from the bounded Read Projection made this run feasible.

仅在本次固定运行范围内，证据表明：

• DCell 有界投影适配器能够进入 AgingBench 官方执行链；
• 它完成了一次十会话 S2 运行；
• 它生成了一张通过官方 Schema 校验的 AgingCard；
• 上游 Scenario、Runner、Probe、Scorer 和 AgingCard Schema 未被修改；
• 将完整 Store 与有界 Read Projection 分离，使本次长期运行成为可能。

4. What This Result Does Not Establish

This result does not establish:

• a complete same-seed Growing History baseline;
• a valid paired comparison;
• a three-seed aggregate;
• use of the Track 1B Haiku-4.5 anchor model;
• confirmed leaderboard eligibility;
• superiority over any baseline;
• scientific validation;
• real-world validation; or
• evidence of industrial deployment.

该结果没有建立：

• 完整的同 seed Growing History 基线；
• 有效配对比较；
• 三个 seed 的聚合结果；
• Track 1B Haiku-4.5 锚点模型的使用；
• 已确认的排行榜资格；
• DCell 对任何基线的优越性；
• 科学验证；
• 真实世界验证；
• 工业部署证据。

5. Store versus Read Projection

The full DCell Store retains candidate current state, history, evidence, conflicts, retractions, unknowns, and uncertainty. Each Agent call consumes a deterministic Read Projection bounded to at most 1,280 Token. The legacy DCellMemoryPolicy is retained only as LEGACY_FULL_STORE_REINJECTION_ADAPTER for forensic comparison; it is not the recommended adapter. DCellMemoryPolicyBoundedProjection is the evaluated adapter.

完整 DCell Store 长期保存候选当前状态、历史、证据、冲突、撤回、未知和 不确定性；每次 Agent 调用仅消费不超过 1,280 Token 的确定性 Read Projection。旧版 DCellMemoryPolicy 仅以 LEGACY_FULL_STORE_REINJECTION_ADAPTER（旧版完整存储重复注入适配器） 身份保留用于法证对照，不是推荐适配器；本次评测使用的是 DCellMemoryPolicyBoundedProjection。

Observed forensic resource evidence:

• Legacy adapter: 2,785,879 Token across the first five started DCell sessions; exact duplicate byte ratio 94.5901%; recursive summary ratio 74.4136%; the full Store was repeatedly reinjected.
• Bounded adapter: 1,125,105 Token across all ten sessions; maximum observed Read Projection 1,280 Token; the S2 run completed.

已观察到的法证资源证据：

• 旧适配器：DCell 前 5 个已启动会话使用 2,785,879 Token；精确重复字节率 94.5901%；递归摘要率 74.4136%；完整 Store 被重复注入。
• 有界适配器：完整 10 个会话使用 1,125,105 Token；最大实际 Read Projection 为 1,280 Token；S2 运行完成。

This is a POST_HOC_ADAPTER_REPAIR_EVALUATION—a post-hoc adapter-repair evaluation, not a confirmatory experiment unaffected by earlier results.

本次属于 POST_HOC_ADAPTER_REPAIR_EVALUATION（事后适配器修复评测）， 不是未受前期结果影响的确认性实验。

6. Historical Attempts

• v0.2: transport failure; no complete DCell run and no official metric.
• v0.3: Growing History completed; DCell stopped at a resource gate; no paired conclusion.
• v0.4: offline forensics identified recursive state expansion and full Store reinjection.
• v0.5: bounded DCell completed; Growing History terminated on a proxy 503 during session 9 of 10; no paired conclusion.

历史尝试及不利结果全部保留：

• **v0.2：**传输失败；没有完整 DCell 运行，也没有官方指标。
• **v0.3：**Growing History 完成；DCell 因资源闸门停止；没有配对结论。
• **v0.4：**离线法证审计发现递归状态膨胀和完整 Store 重复注入。
• **v0.5：**有界 DCell 完成；Growing History 在第 9/10 个会话因代理 503 终止；仍没有配对结论。

See reports/ for the complete bounded reports and adverse-result disclosures. The public snapshot excludes full request ledgers, raw transport logs, and benchmark datasets.

完整的有界报告和不利结果披露位于 reports/；公开快照排除了 全量请求账本、原始传输日志和基准数据集。

7. STATE-Bench Boundary

Three adapter conditions were implemented. The artifact builder used the 300 official training trajectories to construct 848 atomic learning units; the test split was not used for learning. The locked official Azure GPT-5.4 simulator and judge configuration was unavailable. The evaluator was not replaced, no formal STATE-Bench run was executed, and no official STATE-Bench metric exists.

三种适配器条件已经实现。工件构建器使用 300 条官方训练轨迹构造了 848 个 原子学习单元，test split 未用于学习。由于官方锁定的 Azure GPT-5.4 simulator/judge 配置不可用，评价器没有被替换，正式 STATE-Bench 运行没有 执行，也不存在官方 STATE-Bench 指标。

8. Reproducibility

Clone the two upstream repositories separately and check out the commits in UPSTREAM_LOCKS.json. Then install this adapter package in Python 3.11 or newer; the recorded environment used Python 3.12.

请分别克隆两个上游仓库，并切换到 UPSTREAM_LOCKS.json 固定的提交；随后在 Python 3.11 或更高版本中安装本适配器包。本次记录环境 使用 Python 3.12。

python -m venv .venv
. .venv/bin/activate
python -m pip install -e .

export AGINGBENCH_REPO="../AgingBench"
export STATEBENCH_REPO="../STATE-Bench"

Run the 55 local tests using the pinned upstream environments:

使用固定上游环境运行 55 项本地测试：

PYTHONPATH="src:${AGINGBENCH_REPO}/prototype" \
  "${AGINGBENCH_REPO}/prototype/.venv/bin/python" \
  -m unittest tests.test_agingbench_adapter -v

PYTHONPATH="src:${STATEBENCH_REPO}" \
  "${STATEBENCH_REPO}/.venv/bin/python" \
  -m unittest tests.test_statebench_adapter -v

PYTHONPATH="src" \
  "${STATEBENCH_REPO}/.venv/bin/python" \
  -m unittest tests.test_integrity -v

Validate the preserved AgingCard with the official validator:

使用官方 validator 校验已保存的 AgingCard：

card="results/agingbench/s2-bounded-projection-v0.5/runs/01-s2-dcell-bounded-projection-seed-1/output/s2_bounded_v05_dcell_bounded_projection_seed_1/aging_card.json"
PYTHONPATH="src:${AGINGBENCH_REPO}/prototype" \
  "${AGINGBENCH_REPO}/prototype/.venv/bin/python" \
  -m agingbench.metrics.aging_card_validate "$card"

An authorized single S2 adapter rerun would read the key only from the process environment. Never commit a key or substitute a model:

获授权的单次 S2 适配器复跑只应从进程环境读取密钥；不得提交密钥，也不得 替换模型：

export OPENAI_API_KEY="<set-locally-and-never-commit>"
PYTHONPATH="src:${AGINGBENCH_REPO}/prototype" \
  "${AGINGBENCH_REPO}/prototype/.venv/bin/agingbench" run \
  --scenario s2_lifestyle_assistant \
  --sut configs/agingbench/dcell-bounded-projection.yaml \
  --generated --sessions 10 --seeds 1 --card \
  --output local-results
unset OPENAI_API_KEY

This command can incur API cost and does not imply authorization. Do not run it automatically or expand it into a benchmark matrix. The current AgingBench route is closed.

该命令可能产生 API 成本，本身不构成执行授权。不得自动运行，也不得扩张为 大规模矩阵；当前 AgingBench 路线已经关闭。

Further detail is in REPRODUCIBILITY.md.

更多复现细节见 REPRODUCIBILITY.md。

9. Relation to DCell Research

The related Reality Synchronization research repository is joy7758/dcell-reality-sync. Its Reality Drift v0.2 result remains HYPOTHESIS_DECISION=INCONCLUSIVE. DCell superiority is not established, and this external feasibility case does not override the internal uncertain result.

相关 Reality Synchronization 研究仓库为 joy7758/dcell-reality-sync。 其中 Reality Drift v0.2 结果仍为 HYPOTHESIS_DECISION=INCONCLUSIVE。 DCell 优越性尚未建立，本外部可行性案例不能覆盖内部不确定结果。

10. License and Citation

Repository-authored code and documentation are provided under the Apache License 2.0. Third-party benchmark assets remain under their upstream licenses and are not redistributed here. Citation metadata is in CITATION.cff.

本仓库原创代码和文档采用 Apache License 2.0。第三方基准资产仍 受各自上游许可证约束，未在此重新分发。引用信息见 CITATION.cff。