feat(evolution): E2 事件驱动自动演化研究预览

.env.example

@@ -54,13 +54,25 @@ EVOLVER_JOB_TIMEOUT_S=300
 EVOLVER_RUN_TIMEOUT_S=1200
 EVOLVER_JOB_MEM_GB=2
 EVOLVER_LLM_TIMEOUT_S=120
+# E2 事件 Campaign 分阶段开关；关闭时 E1 单代 run 完全不受影响。
+EVENT_EVOLUTION_ENABLED=false
+CAMPAIGN_LEASE_TTL_S=90
+CAMPAIGN_MAX_CONCURRENT=1
 # Evolver 仅用该地址按 owner/config_id 解析既有加密凭据；不会持久化明文 key。
 DASHBOARD_SERVICE_URL=http://localhost:3001
 # Ed25519 DER 的 base64：私钥仅给 orchestration，公钥给 Dashboard 验证逐操作 grant。
 # 生成方式见 services/evolver/README.md；生产 compose 会显式从 Evolver 环境移除私钥。
 EVOLUTION_CREDENTIAL_PRIVATE_KEY_B64=
 EVOLUTION_CREDENTIAL_PUBLIC_KEY_B64=
 
+# 事件唯一事实源：历史结构化导入 + 上线后的精选新闻 forward 归档。
+COINMARKETCAL_API_KEY=
+COINMARKETCAL_BASE_URL=https://api.coinmarketcal.com
+EVENT_PROVIDER_TIMEOUT_S=20
+EVENT_HISTORICAL_LATENCY_S=300
+EVENT_ARCHIVE_ENABLED=false
+EVENT_ARCHIVE_INTERVAL_S=900
+
 # factor 服务可选项（ADR-0043）：qlib Alpha158 风格因子纯 pandas 本地算，默认开；
 # 设 false 可整源关闭。snapshot top-N 去相关阈值默认 0.85（1.0 = 关闭去相关）
 # FACTOR_QLIB_ENABLED=true

AGENTS.md

@@ -104,17 +104,18 @@ pnpm scheduler:trigger daily_btc_deep_dive     # 手动触发一次
 
 ## 6. 当前 Phase 状态
 
-Phase **D-12 + E1 生产闭环**已落地：单 orchestrator + plan/exec 三件套
+Phase **D-12 + E2 事件演化闭环**已落地：单 orchestrator + plan/exec 三件套
 （create_plan / approve_plan / execute_plan）+ hooks + permissions deny +
 approval_token 状态机（D-8/D-9）→ LLM 自创策略沙盒 + 风控引擎 + 多市场数据
 （D-9/D-10）→ 跨币种 cash + **live runner**（promoted 候选按行情自动跑，机器审批
 走护栏内 plan/exec）。D-11.1 收口了 live runner 的信任边界与健壮性
 （candidate 归属校验 / per-account run 上限 / 错误可重试分类）；D-11.2 收口运维；
 D-12 完成 factor 血缘、衰减巡检与因子发现。research-hub 三方辩论已收口。
 E1 已拆出 `services/evolver:8005`：真实 frozen bars、单代 unified-diff 变异、异步
-owner-scoped 状态、显式逐次审批与可复现实验元数据；演化链路不会自动 promote、启动策略或下单。
-当前收口项是冻结 LLM/定价快照、owner key 即时获取与 token/cost 审计；下一里程碑为
-E2 best-parent 多代选择与 early stopping（issue #7），MAP-Elites / Island Model 后置。
+owner-scoped 状态、显式逐次审批与可复现实验元数据。E2 在 feature flag 后增加双时态事件
+snapshot、HypothesisSpec 确定性编译、五代 8×3 共演化、单冠军 Forward、一次性 holdout 与
+`runner_eligible=false` 的人工实验性采用；不会自动 promote、启动策略或下单。
+下一步是在小流量运行中校准事件覆盖与信用分配；MAP-Elites / Island Model 继续后置。
 详见 [`docs/04-current-state.md`](docs/04-current-state.md) / `CLAUDE.md` §3 /
 仓库根 `README.md`。
 

CLAUDE.md

@@ -9,11 +9,11 @@
 - **不是**开箱即用策略平台 / LangChain / AutoGen 包装
 - **三层**：Next.js + CopilotKit → Mastra（TS）→ Python services。详 `docs/01-architecture-overview.md`
 
-## 2. 文档入口 & 当前 Phase（D-12 + E1）
+## 2. 文档入口 & 当前 Phase（D-12 + E2）
 
 - `README.md` / `README.zh-CN.md` 首页；`AGENTS.md` 多工具入口；`docs/00-context.md` 背景 / `01-architecture-overview.md` 架构 / `03-kernel-design.md` services / `04-current-state.md` 进度
 - 内部 ADR 在 `docs/miro/`（gitignored，公开文档勿引用）
-- D-8~D-12 已完成 Plan/Exec、策略创作、风控、多市场模拟盘与因子闭环；research-hub 已收口；E1 独立 Evolver 已落地真实 frozen bars、显式审批、owner 隔离与可复现实验元数据。当前收口冻结 LLM/定价快照与费用审计；下一 E2 best-parent 多代演化 #7
+- D-8~D-12、research-hub 与 E1 已收口；E2 在 feature flag 后落地双时态事件 snapshot、确定性 DSL、五代 8×3 共演化、Forward/Holdout 和非 Runner 实验性采用。E1 原语义不变
 
 ## 3. 协作硬约束
 
@@ -61,8 +61,8 @@ bash scripts/dev.sh                        # data:8001…evolver:8005 + mastra:4
 
 D-9/D-9.1a 收口 + D-10 多市场数据（web/基本面）+ D-11 多市场模拟盘
 （跨币种 cash + live runner #1）+ D-12 因子库闭环（血缘 + 衰减巡检 +
-monthly 宏观 + 因子发现 L1 + 三方研究辩论）和 E1 独立 Evolver 已落地。
-当前：冻结 LLM/定价快照、owner key 即时获取、token/cost 审计；下一：E2 best-parent 多代演化 #7
+monthly 宏观 + 因子发现 L1 + 三方研究辩论）、E1 与 feature-flag E2 事件共演化已落地。
+当前：小流量校准事件覆盖、FDR、费用与 Forward 证据；MAP-Elites / Island Model 后置
 
 ---
 

README.md

@@ -47,7 +47,7 @@ Several capability lines sit on top of that harness:
 
 The name combines **Ina**ri (the Japanese fox deity of prosperity) with **alpha** (the quant term for excess return) — a companion that reads your direction and keeps every step on the record.
 
-> **Status:** Inalpha is in **alpha** — 79 factors with lineage & decay watch, restricted-DSL factor discovery, a three-party research debate, multi-market paper trading, and an E1 strategy-evolution service. Each evolution run requires explicit approval, freezes its dataset and non-secret LLM/pricing snapshot, and never auto-promotes or starts a candidate. Read the code, weigh in on design — **do not run this against real money** (real-money trading is out of scope).
+> **Status:** Inalpha is in **alpha** — 79 factors with lineage & decay watch, restricted-DSL factor discovery, a three-party research debate, multi-market paper trading, and E1/E2 strategy evolution. Feature-flagged E2 adds point-in-time event snapshots, deterministic hypothesis DSL compilation, five-generation campaigns, Forward evidence, and one-shot sealed holdout. Evolution never auto-promotes, starts, or trades a candidate. Read the code, weigh in on design — **do not run this against real money** (real-money trading is out of scope).
 
 ---
 
@@ -128,7 +128,7 @@ Three software layers over one data layer. A request flows down; results flow ba
 | `services/paper` | The event-driven kernel — backtest + paper on the **same code** — plus the LLM-authored-strategy sandbox and the live runner. |
 | `services/research` | Multi-agent deep dive: 6 analysts in parallel, then a bull / bear / risk debate (triggered only when they disagree, with a soft early-stop and the decision chain persisted for replay). |
 | `services/factor` | The factor library (pandas-ta / Alpha101 / qlib + FRED macro): IC screening, current-effective factor timing, lineage & decay watch, DSL factor discovery. **Signals only — never places an order.** |
-| `services/evolver` | Owner-scoped E1 strategy evolution: unified-diff mutation, frozen-data evaluation, candidate/run lineage, cost accounting, and an explicit approval boundary. **Never auto-promotes, starts, or trades a candidate.** |
+| `services/evolver` | Owner-scoped E1/E2 evolution: unified-diff mutation, event-hypothesis campaigns, frozen-data evaluation, lineage, cost accounting, Forward/holdout gates, and explicit final adoption. **Never auto-promotes, starts, or trades a candidate.** |
 
 **L4 · Persistence & external.** Postgres + TimescaleDB holds all time-series and business state. External venues span crypto, US / A-share / HK and other Asian & European single-name equities, global indices, and FRED macro — the orchestrator routes each venue automatically by market type.
 
@@ -173,7 +173,7 @@ Human-written strategies hit a velocity ceiling, and parameter tuning can only a
 - **Reproducible end to end.** One run freezes `as_of`, the closed-bar dataset manifest/hash, seed source, baseline, candidate source/diff, evaluation snapshots, and non-secret LLM/provider/pricing metadata. The user's encrypted API key is resolved only for that owner and never stored in the run.
 - **Explicitly authorized and non-promoting.** Starting a run requires a trusted approval bound to the owner, operation ID, request, estimated cost, and frozen LLM snapshot. Completion never promotes, starts, or routes a candidate into the order path.
 
-> E1 now runs as the separate `services/evolver` service on port 8005. E2 is intentionally narrower than the original research plan: best-parent multi-generation selection plus early stopping first; MAP-Elites and Island Model wait for real run data to justify the complexity.
+> E1 and feature-flagged E2 run in the separate `services/evolver` service on port 8005. E2 uses fixed 8×3 generations, novelty-aware inheritance, one locked Forward champion, and a single-use sealed holdout; MAP-Elites and Island Model wait for real run data to justify the complexity.
 
 ### 4. Swarm — run dozens of backtests in parallel
 
@@ -234,6 +234,7 @@ Where each capability stands today. Live module inventory and the end-to-end dec
 | ✅ Shipped | LLM-authored strategies — E1 MVP | D-9 | three sandbox gates (AST · subprocess · `Strategy` contract) + multi-objective fitness + baseline auto-run |
 | ✅ Shipped | Strategy evolution — E1 production loop | E1 | `services/evolver:8005` · explicit cost-bearing approval · unified-diff mutation · frozen dataset/hash · seed/baseline/candidates evaluated on the same bars · owner-scoped async run/slot state |
 | ✅ Shipped | Frozen LLM approval snapshot | E1 closure | Dashboard approve/deny · owner/operation/model/pricing binding · Ed25519 replay-safe credential grant · per-slot token/cost accounting, including rejected mutations |
+| ✅ Feature flag | Event-driven automatic evolution | E2 | bitemporal event snapshots · HypothesisSpec DSL · five 8×3 generations · Forward + one-shot sealed holdout · human experimental adoption only, never Runner eligible |
 | ✅ Shipped | Risk engine at the HTTP boundary | D-9 | declarative `risk_rules.toml` · pre-trade `enforce` · `risk_locks` table with independent commit |
 | ✅ Shipped | Bull / bear researcher debate | D-9 | opposing-stance researchers under `services/research` |
 | ✅ Shipped | Scheduler / cron agent mode | D-9 | `scheduler_jobs` + advisory lock + `/api/scheduler/*` management plane |
@@ -255,7 +256,6 @@ Where each capability stands today. Live module inventory and the end-to-end dec
 | ✅ Shipped | Cross-sectional factor scoring | D-12 | `factor.panel_score` · `POST /panel/score` · cross-sectional Rank IC (rank the pool each period vs forward cross-sectional return) · native Alpha101 a1/a3 · orthogonal to single-name timing |
 | ✅ Shipped | Time-series cross-validation — anti-overfitting | D-12 | WalkForward / PurgedKFold / Combinatorial Purged CV + Deflated Sharpe · `POST /backtest/cv` · test fold always includes the latest bar · auto-fallback to walk-forward when samples are short |
 | ✅ Shipped | Point-in-time fundamentals | D-12 | Baostock financials filtered by actual publication date · `GET /fundamentals?as_of=` · prevents look-ahead (yfinance v1 not yet PIT, explicitly flagged) |
-| 🗓️ Planned | Strategy evolution — E2 | E2 | best-parent multi-generation loop + early stopping; MAP-Elites / Island Model deferred until real run data shows a diversity problem |
 | 🗓️ Planned | Factor discovery — L2 / L3 | L2 / L3 | multi-agent factor crew (L2) + weekly automated scans (L3), on top of the L1 DSL pipeline already shipped |
 | 🗓️ Planned | Automated decay handling | TBD | reflection-driven backtest + auto-trim of decaying factors — today the decay patrol only alerts, never moves the book |
 | 🔬 Exploring | Alpha Zoo cold start | E1+ | seed factor library with public alphas (Qlib / Kakushadze / GTJA) |

README.zh-CN.md

@@ -232,6 +232,7 @@ Inalpha 把*调度*和*算力*分开：agent runtime 负责扇出网格、聚合
 | ✅ 已上线 | LLM 自创策略 — E1 MVP | D-9 | 三道沙盒（AST 审计 / 子进程 / `Strategy` 协议契约） + 多目标 fitness + baseline 自动并跑 |
 | ✅ 已上线 | 策略演化 — E1 生产闭环 | E1 | `services/evolver:8005` · 计费动作显式审批 · unified-diff 变异 · 冻结数据集/hash · seed/baseline/candidate 同 bars 评估 · owner 隔离异步 run/slot 状态机 |
 | ✅ 已上线 | 冻结 LLM 审批快照 | E1 收口 | Dashboard 批准/拒绝 · owner/operation/模型/计价绑定 · Ed25519 可重试凭据 grant · 被拒变异也记 token/费用 |
+| ✅ Feature flag | 事件驱动自动演化 | E2 | 双时态事件 snapshot · HypothesisSpec DSL · 每代 8×3、连续五代 · Forward + 一次性 holdout · 仅人工实验性采用，禁止 Runner |
 | ✅ 已上线 | 风控引擎落到 HTTP 边界 | D-9 | 声明式 `risk_rules.toml` · 撮合前 `enforce` · `risk_locks` 表（独立 commit） |
 | ✅ 已上线 | Bull / Bear 研究员辩论 | D-9 | `services/research` 立场对抗研究员 |
 | ✅ 已上线 | Scheduler / cron agent 模式 | D-9 | `scheduler_jobs` + advisory lock + `/api/scheduler/*` 管理面 |

TODOS.md

@@ -0,0 +1,5 @@
+# Inalpha 后续路线
+
+- [ ] 事件演化的多资产组合级回测与 cross-market lane。本期只支持单一 crypto 标的，不能把多个市场的事件效应当作组合收益。
+- [ ] 在基础信用分配与五代搜索有足够线上样本后，再评估 MAP-Elites / Island Model。
+- [ ] 评估 1m/5m 事件内执行、盘口、partial fill；当前事件研究硬限定 15m/1h/4h。

apps/dashboard/messages/en.json

@@ -13,6 +13,7 @@
     "risk": "Risk",
     "divination": "Omikuji",
     "evolution": "Evolution",
+    "dataHealth": "Event Data Health",
     "soon": "soon",
     "menu": "Menu",
     "close": "Close",
@@ -65,6 +66,12 @@
   "overview": {
     "title": "Portfolio Overview",
     "subtitle": "A live snapshot of your account, positions, orders, and strategy pool.",
+    "evolution": {
+      "active": "Active Campaigns",
+      "waiting": "Waiting for Events",
+      "adoption": "Awaiting Adoption",
+      "sources": "Event Source Faults"
+    },
     "kpi": {
       "totalEquity": "Total Equity",
       "totalReturn": "Total Return",
@@ -271,6 +278,7 @@
       "order": "Orders",
       "backtest": "Backtests",
       "runner": "Runners",
+      "evolution": "Evolution",
       "conversation": "Chats"
     },
     "empty": "No activity yet.",
@@ -305,6 +313,13 @@
     "resize": "Drag to resize height"
   },
   "lab": {
+    "experimental": {
+      "title": "Experimental Evolution Strategies",
+      "empty": "No adopted event-evolution winner yet.",
+      "badge": "Limited evidence · Experimental",
+      "grade": "Evidence grade: {grade}",
+      "lineage": "View evolution lineage"
+    },
     "title": "Strategy Lab",
     "subtitle": "LLM-authored strategy candidates, ranked by multi-objective fitness — with their latest backtest metrics.",
     "candidates": "Candidates",
@@ -565,7 +580,7 @@
   },
   "evolution": {
     "title": "Evolution",
-    "subtitle": "LLM-driven strategy mutation and evaluation — each run spawns candidates through code mutation, AST audit, sandboxing, and backtest evaluation.",
+    "subtitle": "Event facts and paper feedback drive two-level evolution: agents diversify hypotheses, deterministic DSL arms validate them, and one champion enters durable Forward.",
     "runs": "Total runs",
     "running": "Running",
     "llmCost": "LLM cost",
@@ -632,10 +647,58 @@
     },
     "loadMore": "Load more",
     "loadingMore": "Loading…"
+    ,"campaign": {
+      "title": "Event Campaigns",
+      "empty": "No event campaign yet; E1 single-generation runs keep their existing semantics.",
+      "id": "Campaign",
+      "status": "Status",
+      "stage": "Generation",
+      "budget": "Candidates/gen",
+      "events": "Forward events",
+      "holdout": "Sealed holdout",
+      "updated": "Updated",
+      "sealed": "Unconsumed",
+      "consumed": "Consumed",
+      "back": "Back to evolution",
+      "detailTitle": "Event Evolution",
+      "detailSubtitle": "The audit trail for frozen facts, hypothesis lineage, selection, ablations, Forward, and one-shot holdout.",
+      "track": "Generation Track",
+      "frozenData": "Frozen Facts & Execution",
+      "snapshot": "Event snapshot",
+      "facts": "Facts",
+      "eventHash": "Event hash",
+      "executionModel": "Execution model",
+      "forwardHoldout": "Forward & Holdout",
+      "forwardEvents": "Independent events",
+      "deadline": "Deadline",
+      "winner": "Locked champion",
+      "hypotheses": "Hypotheses, Evidence & Lineage",
+      "selected": "Selected parent",
+      "ablations": "Implementation Ablations & Event Effects",
+      "failure": "Failure",
+      "adopt": "Adopt into Strategy Lab",
+      "adopting": "Adopting…",
+      "adopted": "Adopted as an experimental strategy",
+      "adoptHint": "Adoption does not promote, start Runner, or trade; runner_eligible is enforced false."
+    }
   },
   "backtests": {
     "back": "Agent Activity",
     "notFound": "Backtest run not found.",
     "viewCandidate": "View strategy"
   }
+  ,"dataHealth": {
+    "title": "Event Data Health",
+    "subtitle": "Coverage, first acceptance, revisions, and snapshot integrity for the event source of truth.",
+    "rawEvents": "Raw events",
+    "facts": "Normalized facts",
+    "retractions": "Revisions / retractions",
+    "lastAccepted": "Last accepted",
+    "sources": "Source coverage",
+    "source": "Source",
+    "count": "Events",
+    "latest": "Last fetch",
+    "versions": "Max version",
+    "empty": "No event source has been archived yet."
+  }
 }