Conformance — 2Origin Compatible checklist

A machine is 2Origin Compatible only when it satisfies all checks below. Each check is executable, so "conformance" is testable, not aspirational.

Status legend: ✅ passed (with evidence) · ⏳ in progress · ❌ not yet

C1 — Cross-Session state ✅

Close the session, reopen, and the task resumes via State — not transcript replay.

Test:

Start a task, create a task.origin.json.
Open a fresh harness session in the same directory (zero prior conversation).
Check it receives the state (e.g. via SessionStart hook) and can state the task title/goal without re-asking.

Pass = the fresh session reports the correct task without user explanation.

Evidence (2026-08-08): fresh claude -p session auto-received task.origin.json summary and reported the title. Extended to a simulated second machine: copying the environment (CLAUDE.md, schemas, hooks, a task.origin) to a new directory and opening a fresh session there still auto-loads the state and reports the title — the credentials "transferred" with the environment (environment-is-image / AI can transfer schools).

C2 — Cross-Harness ✅

Switching harness keeps the state usable — no re-asking.

Test:

Harness A does half the task, updates task.origin.json.
Harness B (codex exec, etc.) continues from only that state + facts.
Confirm B never asks "what is the task?"

Pass = B continues without re-asking, and can also persist via the Southbridge write path.

Evidence (2026-08-08): Claude Code → Codex handoff, zero re-asking (session logs). Full loop: Codex persisted its output itself via the Southbridge MCP tool (southbridge_write → demo/southbridge-e2e.md, 25 bytes), closing the cross-harness handoff end-to-end without a host harness writing on its behalf.

C3 — Cross-Model ✅

Switching models keeps the state usable — credentials do not reset.

Test: same task.origin.json, two different model endpoints (via ANTHROPIC_BASE_URL + --settings override), each fresh session asked "what is the task?" Compare answers.

Pass = both models report the same goal / state / next-steps without drift.

Evidence (2026-08-08, truthful):

Model A: deepseek-v4-flash (official DeepSeek endpoint)
Model B: deepseek-v4-pro (Xiapan Cloud endpoint api.u-claw.org.cn, a genuinely different model — returns reasoning_content/thinking, proving it's a distinct inference engine)

Both auto-loaded the same task.origin.json via SessionStart hook and reported identical goal, identical current state (including the C5 bug-catch detail), identical first next-step — zero drift across a genuine model swap.

Honesty trail: an earlier draft claimed ✅ with a same-endpoint rename — corrected to ⚠️, then re-verified with a real second endpoint (Xiapan Cloud deepseek-v4-pro). Final status ✅ is backed by two genuinely different models.

C4 — Portable actions ✅

The same action can be executed by different drivers (API / CLI / GUI / device).

Test: node examples/portable-action/document-save.js --driver cli and --driver api; compare output hashes.

Pass = identical sha256 from both drivers.

Evidence (2026-08-08): both drivers produced byte-identical document.save output (21e01778…, 41 bytes). See examples/portable-action/.

C5 — Verifiable results ✅

Outcomes are confirmed by observing real state, not exit codes.

Test: node conformance/tools/verify-state.mjs <task.origin.json> — checks artifacts exist on disk, verified facts carry sources, and state is self-consistent.

Pass = verdict ✅ VERIFIED from real observation.

Evidence (2026-08-08): verified task1 + task2. On first run it caught a real bug — task2 declared actionable-notes.md in artifacts but the file didn't exist on disk; verdict was ❌ NOT VERIFIED. Rebuilding the file flipped it to ✅ VERIFIED (9 checks pass). The validator's job: architecture proves itself by observing reality, not by trusting claims.

C6 — No auto-permanent learning ✅

Experience must pass candidate → verified. One success is not a permanent truth.

Test: inspect the learnings[] lifecycle: new items are candidate, only promoted after confirmation.

Evidence (2026-08-08): task.origin files carry candidate items; promotion is a manual/review step, not automatic.

C7 — Auditable ✅

All critical actions, promotions, and policy changes are recorded.

Test: confirm a write action leaves an audit trail (actor, action, target, timestamp).

Evidence (2026-08-08): Southbridge MCP logs every southbridge_write (done / denied) to audit.log. End-to-end: Codex's southbridge_write of demo/southbridge-e2e.md is recorded (done, 25 bytes); a path-traversal attempt (evil/../shadow.txt) is recorded as denied. Every write leaves an auditable trail.

Summary

#	Check	Status
C1	Cross-Session	✅
C2	Cross-Harness	✅
C3	Cross-Model	✅ 真实第二端点
C4	Portable actions	✅
C5	Verifiable results	✅
C6	No auto-permanent learning	✅
C7	Auditable	✅
C8	Cross-Session Retention	✅ 100%（机制）
C9	Continuation Cost	⚠️ 降级为机制自测，原结论已撤回
C10	State vs Transcript	✅ 97.5% vs 60.0%（对最强对照 rag，区间不重叠）

10 项：9 ✅ / 1 ⚠️。 两条诚实链条：

C3：初标 ✅（误：同端点改名）→ 纠正为 ⚠️（只有一个端点）→ 用虾盘云 deepseek-v4-pro 真实第二端点复测 → 最终 ✅（真跨模型，零漂移）。
C9：初标 ✅「传统 transcript 无法无追问续作」→ 复核发现该结论由对照组的生成方式 保证恒真（判分器测的是自己写下的话）→ 降级为机制自测，"更有效"的证据改由 C10 用真实语料 + 真实模型重新取得。真实答案比原结论更锋利：传统 transcript 能续， 但会自信地续错任务。

首测证据必须诚实；证据补足后可以如实标绿，证据被推翻时必须如实降级。

C8 — Cross-Session Retention (ShadowWork Bench) ✅

The credential-survival claim, quantified. Traditional harness = 0%; 2Origin = ~100%.

Test: bash conformance/run-tests.sh → C8 runs 2origin-harness/bench/shadowwork-bench.mjs.

Pass = retention ≥ 80%.

Evidence (2026-08-08): 50 facts across 20 simulated session closes/reopens → 100.0% retention. New session auto-loads all credentials via benjing bundle. Traditional harness (no 本境) = 0%.

Honest scope: this measures the mechanism (the 本境 preserves credentials across sessions), not the semantic quality of what's preserved (that's gated by promotion + auto-forgetting).

C9 — Continuation Cost (simulated self-test) ⚠️ 已降级

原标题 "Continuation Efficiency (2Origin vs traditional)"，原结论 "传统 transcript 无法无追问续作"。该结论已撤回——它是构造出来的，不是观察到的。

Test: bash conformance/run-tests.sh → C9 runs 2origin-harness/bench/compare-bench.mjs.

为什么降级： compare-bench.mjs 的对照组 transcript 由脚本自己的 genTranscript 生成，而它从没把 goal / next_steps[0] 原文写进 transcript，于是 tradCanResume 恒为 false，而 success 的判定条件正是 !tradCanResume。判分器测的是自己写下的话，不是世界。

实测复核：

transcript 是否含 goal 原文 : false
transcript 是否含 step 原文 : false
→ compare-bench 的 success 条件恒为 true

现在的身份： 机制自测（bundle 侧确实含 goal/next-steps，这部分成立）。 "更有效"的真实证据改由 C10 承担。

诚实链条：C9 初标 ✅「传统无法续作」→ 用真实语料 + 真实模型复测（C10）→ 发现传统 transcript 能无追问续作，只是续错任务。原结论方向对、机制说错了，故撤回重述。这是 C3 那条"初标✅ → 纠正 → 真实复测"的同一条纪律。

C10 — State vs Transcript (ShadowWork Bench spec 0.3) ✅

"状态优于对话流"的真实对照实验：真实语料 + 真实模型 + 真值取自磁盘。

Test: bash conformance/run-tests.sh → C10 跑 shadowwork-bench-live.mjs --dry-run （免费、离线：验证对照组是真实会话记录、两臂 payload 等长）。真打模型需 SHADOWWORK_LIVE=1 bash conformance/run-tests.sh——一次 474 调用、约 600 万输入 token，要花钱，所以做成显式开关：会悄悄花钱的 conformance 项本身就是缺陷。

Pass = 对照组必须是真实会话记录（不是脚本生成的），且两臂 payload 等长（预算对齐）。

Evidence (2026-08-09, spec 0.3): 9 份真实 task.origin.json（131 条已验证事实）+ 2 份真实 Claude Code 会话记录（3.0MB + 2.5MB，按字节钉死）。deepseek-v4-flash，temperature 0， 79 题/臂 × 6 臂 = 474 次调用，0 错误 0 截断。六个臂只差"开场喂什么"，除 10x 外同预算 9,635 字符：

arm	输入token	四选一（瞎猜25%）	事实均衡准确率（瞎猜50%）	反问
2origin（本境 bundle）	5,191	100.0% [79.6,100]	97.5% [87.1,99.6]	否
summary（模型摘要 /compact）	5,611	60.0% [35.7,80.2]	45.0% [30.7,60.2]	否
rag（按题 BM25 检索）	4,586	46.7% [24.8,69.9]	60.0% [44.6,73.7]	否
transcript（尾部截断）	4,083	46.7% [24.8,69.9]	47.5% [32.9,62.5]	否
transcript-10x	40,397	53.3% [30.1,75.2]	35.0% [22.1,50.5]	否
none（空上下文）	231	33.3% [15.2,58.3]	47.5% [32.9,62.5]	是

四条结论：

补上最强对照后差距仍显著：事实召回 2origin 97.5% vs 最强对照 rag 60.0%， Wilson 区间不重叠。v0.2 只跟"尾部截断"比是挑软柿子，0.3 把 summary / rag 补上了。
摘要（/compact）把"哪条事实已验证"压没了：45.0%，低于瞎猜线，也低于什么都不喂。它保住了"在干什么"的梗概，丢掉了续作最需要的那一半。
给对话流 10 倍预算反而更差：35.0%，显著低于瞎猜。更多历史 = 更多过时的、被推翻的、属于别的任务的断言。比"信息不足"更糟，因为它有自信。
唯一说"我不知道"的仍是空上下文那一臂，五个有上下文的臂全都没反问，直接答。

地板线自查通过：空上下文臂在所有 n≥8 的指标上贴着瞎猜线 → 题目没泄题。（v0.2 时它曾在状态字段题上拿 100%——因为目标任务字段最长，"挑最长的"就能蒙对；已改为干扰项按长度就近挑 + 正解位置轮流，并加了自动泄题检查。）

Honest scope: 单模型 / 单任务 / 单次，未重采样；2origin 的 100% 是开卷（事实字面在 bundle 里），本项测的是**「状态能不能无损送达」，对照臂的低分才是信息量。rag 是词法检索不是向量检索。 一条如实作废的指标：为突破开卷天花板新加的"两跳题"分不开任何一臂（六臂挤在 13.4pp 内，空上下文臂还比 2origin 高 0.5pp）——它测的是主题相似度而非状态归属，不用于任何结论。完整方法、置信区间、以及这轮抓出的四个判分器自身的 bug** 见 2origin-harness/bench/RESULTS-v3.md。