34 KiB

Raw Blame History

1

现在进入到第6轮，prior-QPI handoff & direct IA exception product contract review

Codex完成了相关的文档准备，我已经全部上传。

按设计，08_CCRA_REVIEW_QUESTIONS_ROUND06.md 中的 R06-Q01 至 R06-Q23 需要后续 Owner / CCRA 判断后，才能考虑 Round 06.1 selector-support planning。

你怎么判断？

According to a document from 2026-06-19：我的判断是，Round 06 可以作为 document-only contract review 关闭，但还不应直接进入 Round 06.1 selector-support patch；可以进入“06.1 planning 前置整理”，前提是先按下面几处修改合同文本。当前 readiness 文档本身也标记为 NOT_READY_FOR_SELECTOR_PATCH，且 Round 06 明确不是 selector patch。

1. 总体判断

我理解这个产品暂时可以定义为：在 QPI 与思想考古学之间建立一个可审计、可拒绝、可回归的结构化交接合同，防止“深度语言”或“用户声称已 QPI”绕过前置路由。

我的结论分三层：

产品方向成立：prior_qpi_result / qpi_context_provided 是必要的。否则系统会把“QPI 已经判断”这种自然语言当成事实，破坏 QPI-before-IA 的默认管线。
直达 IA 例外必须保守：四门槛设计是对的，尤其是“材料充分”和“深度会改变边界”这两个门槛，可以防止思想考古学变成泛哲学化按钮。
现在不该改 selector：Round 06 的范围是定义合同字段、门槛、例子、calibration / regression 候选标准；明确不改 selector、regression、model card、生命周期。

2. R06-Q01 至 R06-Q23 判断表

ID	判断	我的意见
R06-Q01	accept	必须区分 `system_generated_qpi`、`owner_confirmed_qpi`、`user_claimed_qpi`。这是 skip-QPI 的信任门。
R06-Q02	accept	`user_claimed_qpi` 必须失败 skip-QPI。自然语言声称不能成为 handoff proof。
R06-Q03	accept with edit	最小字段方向正确，但建议把 `qpi_version`、`original_input_ref` 从推荐字段提升为强建议字段，不一定阻塞 selector，但阻塞审计闭环。
R06-Q04	accept	`classification_confidence: low` 应阻塞 skip-QPI。低置信度只能触发轻量复核或 clarification。
R06-Q05	accept	`validity_scope.applies_to_current_input === true` 必须是硬门槛。否则 prior QPI 会漂移复用。
R06-Q06	accept	`handoff_allowed_models` 与 `handoff_blocked_models` 都要保留。只用 allow 不够表达显式禁止。
R06-Q07	accept with edit	`qpi_context_provided` 名字和形态可接受，但建议补一个 `context_object_ref` 或 `same_object_claim`，否则“它 / 这个现象”仍可能指代漂移。
R06-Q08	defer / accept-for-now	当前 action taxonomy 够 MVP，但先不要宣称充分。建议 06.1 planning 里测试是否需要 `resume_prior_analysis` 或 `validate_prior_qpi`。
R06-Q09	accept	QPI-before-IA 必须继续是默认。likely-IA 语言不能改默认管线。
R06-Q10	accept with edit	四个 direct-IA gates 方向正确，但“no prior QPI need”要写成可操作判定：当 dominant scarcity 未定、对象边界未定、材料充分性未定时，一律 QPI-first。
R06-Q11	accept	三标签优于二分类：`ia_direct_allowed`、`qpi_first_with_ia_candidate`、`no_call_or_clarify` 能表达中间态。
R06-Q12	accept	`likely_intellectual_archaeology` 必须只是 review label，不能是 selector condition。
R06-Q13	accept	`底层`、`机制`、`结构`、`世界观` 等深词没有 analyzable material 时不能触发 direct IA。
R06-Q14	accept	`R04-BI-026/027/038` 在 direct IA contract 被接受并实现前，应保持 QPI-first。
R06-Q15	accept	valid / invalid structured prior-QPI examples 应在合同接受后才进 calibration。
R06-Q16	accept	`R04-BI-025` 不现在进 regression；等 structured prior-QPI selector support 实现或确认为高风险后再加。
R06-Q17	accept	`R04-BI-027/038` 只有在 direct IA support 带来过召回风险时才进 regression。
R06-Q18	accept	`R04-BI-030` 继续作为 no-call / clarify guardrail candidate。
R06-Q19	accept	`R04-BI-022` 已由现有 regression 表达，不重复加。
R06-Q20	accept	Round 06 应 document-only close，除非 Owner 明确批准 Round 06.1。
R06-Q21	accept	任何 Round 06.1 selector change 都必须沿用 Round 05.1 diff contract。
R06-Q22	accept	所有 non-target behavior change 都需要 Owner / CCRA disposition。
R06-Q23	accept	接受 Round 06 合同不等于生命周期升级；QPI / IA 生命周期保持不变。

3. 我会要求 Codex 先改的地方

第一处：把 Q03 的“最低字段”分成 runtime minimum 与 audit recommended。 现在最低字段能支持 selector 判断，但不够支持长期追溯。prior_qpi_result 的目的不是自然语言声明，而是可检查、可判断适用范围的产品交接对象；文档已明确它不是用户自然语言 claim。建议拆成：

runtime_required:
- source
- classification
- classification_confidence
- dominant_scarcity
- evidence_gap
- misclassification_risk
- next_model_candidates
- validity_scope.applies_to_current_input

audit_recommended:
- qpi_version
- original_input_ref
- recommended_next_step
- scarcity_profile
- validity_scope.scope_note

第二处：给 Q07 增加“同一对象判定”字段。 qpi_context_provided 的核心问题是：prior QPI 是否强到足以适用于“当前这个输入”。文档已经把这个问题写得很清楚。但真实用户输入里常出现“它”“这个现象”“还是那个问题”，所以 wrapper 里最好增加：

"context_continuity": {
  "same_object": true,
  "same_frame": true,
  "continuity_note": "Same object and same analytical frame as prior QPI."
}

这不一定进入 06.1 selector 第一版，但要进入合同。

第三处：把 Q10 的 no-prior-QPI-need 写成反向规则。 现在四门槛是对的，但 Codex 实现时容易把它做成关键词判断。建议明确：

Direct IA must fail when:
- dominant scarcity is not yet determined;
- material sufficiency is unclear;
- current input may be path/resource/consensus/order scarcity;
- user asks for deep analysis but object reference is unresolved;
- depth would only add explanation flavor, not change judgment, path, verification, or action boundary.

这与 direct IA policy 的四门槛和 anti-overreach rules 是一致的。

4. Calibration / regression 的判断

我不建议 Round 06 现在新增任何 regression。理由是文档自己已经给出 regression 标准：只有合同边界被接受、案例是高风险 guardrail、预期行为稳定、输入固定、未来 selector change 可能破坏边界时，才升级成 regression；普通探索例子应留在 calibration。

我的选择是：

Calibration candidate:
- structured_prior_qpi_handoff_valid
- structured_prior_qpi_handoff_invalid
- natural_language_prior_qpi_claim
- direct_ia_exception_candidate
- direct_ia_material_insufficient
- qpi_first_with_ia_candidate
- no_call_or_clarify_low_context
- depth_limited_qpi_override_ia_rejected

Regression now:
- none

Regression later:
- R04-BI-025 only after structured handoff support exists or becomes high-risk
- valid/invalid structured fixture only after field contract accepted
- R04-BI-027 / R04-BI-038 only if direct IA support creates over-selection risk
- R04-BI-030 only if 06.1 risks reopening low-context behavior
- R04-BI-022 no duplicate

5. 是否授权 Round 06.1？

我的判断是：

Round 06 close: YES, document-only close.
Round 06.1 selector-support patch: NOT YET.
Round 06.1 selector-support planning: YES, but only as planning, not patch.

06.1 planning 可以做三件事：

把 accepted / deferred / edited decisions 写成 owner-decision.md 或等价 review decision 文件；
准备 fixture 设计，不写入 selector / regression；
输出 06.1 patch charter，列明 changed files、target behavior、non-target diff protocol、lifecycle guard。

但不能做：

- 修改 selector_rules.json
- 修改 selector_calibration_inputs.json
- 新增 regression
- 改 model card / model spec
- 生命周期升级
- 引入 LLM selector / RAG / 第三模型

这也符合 readiness check 中 “Round 06.1 only after approval” 与 “Round 06 stops at product-contract review” 的边界。

2

当前边界是：Round 06 close；Round 06.1 planning 已完成；Round 06.1 selector-support patch 未授权。 文档已上传，你检查一下，看是否继续

核心判断：不继续进入 selector-support patch。当前应停在 Round 06.1 planning complete 的边界，只允许做状态同步、Owner decision 记录、审批清单整理，不允许 Codex 开始实现。

1. 文档总体一致

你给出的当前边界，和上传文档基本一致：

Round 06 仍被定义为 contract review / document-only，明确不得修改 selector、calibration、regression、model card、model spec、lifecycle。
prior_qpi_result 已按上轮意见拆成 runtime required 与 audit recommended，这说明 Q03 的关键修改已落地。
qpi_context_provided 已加入 context_continuity.same_object / same_frame，Q07 的修改也已落地。
Direct IA 的 reverse fail rules 已加入，Q10 的“不要让深词变成直达 IA 关键词路由”也落地了。
06.1 planning 文件明确写着：planning complete，但 patch not authorized，Codex should stop here。

所以，合同文本修订已经完成，Round 06.1 planning 也已经足够完整。

2. 但有一个状态文字小不一致

有些文档状态仍写成：

Status: contract_review_draft

以及 09_ROUND06_GPT_REVIEW_DISPOSITION.md 里仍说：

document_only_close_ready

而你现在给出的边界是：

Round 06 close
Round 06.1 planning completed
Round 06.1 selector-support patch not authorized

这不是实质矛盾，但属于状态标记滞后。建议后续只做一次轻量状态同步，把状态统一成：

Round 06: document_only_closed
Round 06.1 planning: planning_complete
Round 06.1 patch: not_authorized

这可以避免 Codex 或 Local CCRA 后面误读为“还在等待 Round 06 close decision”。

3. 是否继续：我的判断

不建议继续让 Codex 做任何 patch。原因很明确：07_SELECTOR_SUPPORT_READINESS_CHECK.md 仍标记为 NOT_READY_FOR_SELECTOR_PATCH，并且只允许未来在 Owner / CCRA 接受所有相关 gates 后才考虑 patch。

11_ROUND06_1_SELECTOR_SUPPORT_PATCH_PLANNING.md 也列出了 patch 前置门槛 P01-P08：不仅要 Round 06 close，还要批准 exact implementation targets、fixture IDs、changed-file allowlist、calibration decision、regression decision、diff protocol、non-target disposition protocol。当前 planning 已把这些列出来，但没有显示它们已经被逐项授权。

所以当前结论是：

Round 06 close: 可以视为完成
Round 06.1 planning: 可以视为完成
Round 06.1 selector-support patch: 不继续
Codex: stop

4. 唯一可以继续的安全动作

可以继续的只有三类，不属于 patch：

状态同步：把文档状态从 contract_review_draft / close_ready 统一到 document_only_closed / planning_complete / patch_not_authorized。
Owner decision 记录：新增或更新一个简短 decision 文件，明确当前边界。
Patch authorization checklist：准备一页 P01-P08 审批清单，供你未来决定是否授权 06.1 patch。

不应继续做：

- selector/selector_rules.json 修改
- selector_calibration_inputs.json 修改
- regression 新增
- model card / model spec 修改
- lifecycle promotion
- audit script 创建
- before / after baseline 捕获

5. 给 Codex 的直接指令建议

可以给 Codex 这一句：

Round 06 is closed as document-only. Round 06.1 planning is complete. Round 06.1 selector-support patch is not authorized. Do not modify selector, calibration, regression, model cards, model specs, lifecycle fields, or create patch audit scripts. Only perform documentation status synchronization and owner-decision recording if explicitly requested.

我的判断：现在不继续实施；只做状态归档或等待下一次明确授权。

According to a document from 2026-06-19: 我作为 Web 版 CCRA 的正式判断是：Round 06 可以关闭；Round 06.1 planning 可以关闭；但 Round 06.1 selector-support patch 现在仍不授权。

Web CCRA 判断

Round 06 document-only close: PASS / authorize close
Round 06.1 planning: PASS / authorize planning close
Round 06.1 selector-support patch: HOLD / not authorized
Next allowed action: authorization packet / owner-decision record only

理由很简单：现在的文档已经把产品契约、planning prep、fixture draft、patch charter、allowlist、diff protocol 都铺好了，但它仍然没有满足“可以动 selector”的全部前置条件。07_SELECTOR_SUPPORT_READINESS_CHECK.md 仍明确标记为 NOT_READY_FOR_SELECTOR_PATCH，并列出 G01-G09 仍是 draft 或待 Owner / CCRA 接受状态。

我接受什么

我接受以下内容作为 Web CCRA 判断：

接受 Round 06 document-only close。 Round 06 的范围本来就是产品合同字段、source trust、例子、calibration / regression 候选、readiness gate，不应修改 selector、calibration、regression、model card、model spec 或 lifecycle。
接受 Round 06.1 planning 已完成。 11_ROUND06_1_SELECTOR_SUPPORT_PATCH_PLANNING.md 已经把未来 patch 的目标、约束、前置条件、allowlist、disallowlist、target matrix 和 diff contract 写出来了；它的状态也明确是 planning_only_not_patch。
接受 Round 06 的三处关键编辑已经落地。 Q03 字段拆分、Q07 context continuity、Q10 reverse fail rules 都已记录为 Codex-applied edits。

我不接受什么

我不接受现在直接授权 selector-support patch。

原因是 11_ROUND06_1_SELECTOR_SUPPORT_PATCH_PLANNING.md 自己写得很清楚：patch execution 必须满足 P01-P08，包括 exact implementation scope、exact fixture IDs and expected behavior、changed-file allowlist、calibration decision、regression decision、diff protocol、non-target disposition protocol；任何 gate 缺失，Codex 必须停在 planning。

现在缺的不是“文档够不够多”，而是授权对象还不够精确：

缺 exact approved target IDs
缺 exact fixture IDs and expected behavior
缺 calibration JSON 是否新增的明确授权
缺 regression 是否新增的明确授权
缺 implementation changed-file allowlist 的 Owner 授权记录
缺 non-target behavior disposition 的执行口径确认

尤其是 direct IA 的 material-sufficient fixture 仍要求 “new exact text required”，这说明至少一个核心目标行为还没有固定成可测试输入。

你现在应该授权什么

我建议你现在只授权一个很窄的动作：

Authorize Codex to create / update Owner decision record for:
1. Round 06 document-only close;
2. Round 06.1 planning complete;
3. Round 06.1 selector-support patch remains not authorized;
4. Future patch requires separate approval of P01-P08;
5. No selector, calibration, regression, model card, model spec, lifecycle edits are allowed.

也可以允许 Codex 做一次状态同步，把文档状态从 contract_review_draft / close_ready / planning_only_not_patch 同步成：

Round 06: document_only_closed
Round 06.1 planning: planning_complete
Round 06.1 patch: not_authorized

给 Codex 的授权文本

你可以直接发：

Owner authorization based on Web CCRA judgment:

Round 06 is authorized to close as document-only.
Round 06.1 planning is accepted as complete.
Round 06.1 selector-support patch is not authorized.

Codex may only create or update documentation-level owner-decision / status-synchronization records reflecting this boundary.

Codex must not modify selector/selector_rules.json, selector/selector_calibration_inputs.json, regression files, model cards, model specs, lifecycle status fields, or create patch audit scripts.

Any future Round 06.1 implementation requires separate Owner authorization covering exact target IDs, exact fixture IDs and expected behavior, changed-file allowlist, calibration decision, regression decision, deterministic before/after diff protocol, and non-target behavior disposition protocol.

我的 Web CCRA 结论：现在可以授权“关闭与归档”，不可以授权“实现”。

3

Owner decision 记录：新增或更新一个简短 decision 文件，明确当前边界。

Patch authorization checklist：准备一页 P01-P08 审批清单，供你未来决定是否授权 06.1 patch。

你帮我准备这两个文件的文本，我参考一下。

下面两个文本可以直接给 Codex 落成 Markdown 文件。依据是：Round 06 文档允许 planning prep 但不允许 patch；06.1 planning 明确是 planning_only_not_patch；readiness 仍是 NOT_READY_FOR_SELECTOR_PATCH。

文件 1：`12_OWNER_DECISION_ROUND06_CLOSE_AND_R061_PLANNING_COMPLETE.md`

# Owner Decision: Round 06 Close And Round 06.1 Planning Complete

Date: 2026-06-19

Round: `06` / `06.1-planning`

Status: `owner_decision_record`

## 1. Decision Summary

Owner accepts the Web CCRA judgment that:

```text
Round 06 document-only close: authorized
Round 06.1 planning: accepted as complete
Round 06.1 selector-support patch: not authorized

This decision records closure and planning status only.

It does not authorize selector implementation, calibration JSON changes, regression additions, model card changes, model spec changes, lifecycle promotion, audit script creation, or any Round 06.1 patch execution.

2. Accepted Boundary

The accepted current boundary is:

Round 06: document_only_closed
Round 06.1 planning: planning_complete
Round 06.1 selector-support patch: not_authorized

Round 06 is accepted as a document-only product contract review.

Round 06.1 planning is accepted as planning-only material that prepares possible future implementation scope, fixture groups, allowlist, disallowlist, baseline capture requirements, diff protocol, and closeout requirements.

Round 06.1 selector-support patch remains blocked until separately authorized by Owner / CCRA.

3. Accepted Contract Edits

Owner acknowledges that the following Round 06 contract edits have been applied:

prior_qpi_result fields are split into:
- runtime_required
- audit_recommended
qpi_context_provided includes context-continuity fields:
- context_continuity.same_object
- context_continuity.same_frame
- context_continuity.continuity_note
Direct IA exception policy includes reverse fail rules for cases where:
- dominant scarcity is not yet determined;
- material sufficiency is unclear;
- current input may still require QPI;
- object reference is unresolved;
- depth only adds explanation flavor without changing judgment, path, verification method, or action boundary.

4. Explicitly Authorized Actions

Codex is authorized only to perform documentation-level closure actions:

Record this Owner decision.
Synchronize status labels in Round 06 / Round 06.1 planning documents if needed.
Prepare a future patch authorization checklist.
Preserve existing Round 06 and Round 06.1 planning materials as review artifacts.

5. Explicitly Unauthorized Actions

Codex must not perform any of the following actions under this decision:

modify selector/selector_rules.json;
modify selector/selector_calibration_inputs.json;
modify tests/qpi.regression.json;
modify tests/intellectual_archaeology.regression.json;
modify tests/regression_cases.json;
modify model cards;
modify model specs;
change lifecycle status fields;
create or run Round 06.1 patch audit scripts;
capture before / after selector baselines as patch evidence;
add calibration rows;
add regression cases;
introduce LLM selector;
introduce RAG, vector database, frontend, backend, service layer, account system, or a third model.

6. Lifecycle Boundary

This decision does not promote any model lifecycle status.

QPI remains:

status: draft
stability_level: B
regression_status: pending

Intellectual Archaeology remains:

status: draft
stability_level: B
regression_status: pending

Round 06 contract acceptance and Round 06.1 planning completion are not lifecycle upgrade evidence.

7. Future Patch Requirement

A future Round 06.1 selector-support patch requires a separate Owner / CCRA authorization record.

That future authorization must approve, at minimum:

exact implementation target IDs;
exact fixture IDs and expected behavior;
changed-file allowlist;
calibration decision;
regression decision;
deterministic before / after diff protocol;
non-target behavior disposition protocol;
lifecycle non-promotion statement.

Until that separate authorization exists, Codex must stop at documentation-level closure.

8. Final Owner Decision

Round 06 document-only close: ACCEPTED
Round 06.1 planning complete: ACCEPTED
Round 06.1 selector-support patch: NOT AUTHORIZED
Next action: stop after documentation closure and authorization checklist preparation


## 文件 2：`13_ROUND06_1_PATCH_AUTHORIZATION_CHECKLIST.md`

```markdown
# Round 06.1 Selector-Support Patch Authorization Checklist

Date: 2026-06-19

Round: `06.1`

Status: `authorization_checklist_not_approved`

## 1. Purpose

This checklist defines what Owner / CCRA must explicitly approve before any Round 06.1 selector-support patch may begin.

This file is not patch authorization.

It is a future decision aid.

## 2. Current Authorization State

```text
Round 06 document-only close: accepted
Round 06.1 planning: complete
Round 06.1 selector-support patch: not authorized

Codex must not start implementation until this checklist is completed and approved by Owner / CCRA.

3. Patch Authorization Gates

gate_id	gate	required approval evidence	current state	decision
P01	Round 06 close	Owner / CCRA confirms Round 06 is closed as document-only.	satisfied	accept
P02	Implementation scope	Owner / CCRA approves exact Round 06.1 implementation targets.	missing	defer
P03	Fixture set	Owner / CCRA approves exact fixture IDs and expected behavior.	missing	defer
P04	Changed-file allowlist	Owner / CCRA approves files Codex may edit.	missing	defer
P05	Calibration decision	Owner / CCRA approves calibration JSON additions, or confirms none.	missing	defer
P06	Regression decision	Owner / CCRA approves regression additions, or confirms none.	missing	defer
P07	Diff protocol	Owner / CCRA approves deterministic before / after behavior diff.	draft exists	defer
P08	Non-target disposition	Owner / CCRA confirms every non-target behavior change requires accept / reject / defer.	draft exists	defer

4. Minimum Approval Required Before Patch

Patch execution may begin only when all gates below are marked accept:

P01 accept
P02 accept
P03 accept
P04 accept
P05 accept
P06 accept
P07 accept
P08 accept

If any gate remains defer, reject, or missing, Codex must stop at planning.

5. Candidate Implementation Targets

Owner / CCRA must approve exact target IDs before implementation.

target_id	fixture_group	proposed future behavior	current decision
T01	`structured_prior_qpi_handoff_valid`	May consider IA without re-running QPI only when trusted source, runtime fields, validity scope, context continuity, and `enter_ia` all pass.	defer
T02	`structured_prior_qpi_handoff_missing_runtime_field`	Must not skip QPI.	defer
T03	`structured_prior_qpi_handoff_user_claimed`	Must not skip QPI; user-claimed QPI remains lightweight QPI review.	defer
T04	`structured_prior_qpi_handoff_low_confidence`	Must not skip QPI when confidence is `low`.	defer
T05	`structured_prior_qpi_handoff_scope_false`	Must not skip QPI when validity scope is false or unclear.	defer
T06	`structured_prior_qpi_handoff_continuity_false`	Must not skip QPI when same-object or same-frame continuity fails.	defer
T07	`direct_ia_exception_material_sufficient`	May route to `ia_direct_allowed` only with exact object, sufficient material, no prior-QPI need, and boundary-changing depth.	defer
T08	`direct_ia_exception_material_insufficient`	Must not become IA-only; should fall back to `qpi_first_with_ia_candidate` or `no_call_or_clarify`.	defer
T09	`natural_language_prior_qpi_claim`	Must not become IA-only.	defer
T10	`low_context_prior_reference`	Should remain no-call / clarify.	defer

6. Candidate Fixture Approval

Owner / CCRA must approve exact fixture IDs and expected behavior before any calibration or selector changes.

fixture_id	source	exact input fixed?	calibration authorized?	regression authorized?	decision
`R06-PS-001`	new structured object	yes / draft	no	no	defer
`R06-PS-002`	new structured object	yes / draft	no	no	defer
`R06-PS-003`	`R04-BI-025` variant	pending exact text	no	no	defer
`R06-PS-004`	new structured object	yes / draft	no	no	defer
`R06-PS-005`	new structured object	yes / draft	no	no	defer
`R06-PS-006`	new structured object	yes / draft	no	no	defer
`R06-IA-001`	new exact text required	no	no	no	defer
`R06-IA-002`	`R04-BI-026`	yes	no	no	defer
`R06-IA-003`	`R04-BI-027`	yes	no	no	defer
`R06-IA-004`	`R04-BI-038`	yes	no	no	defer
`R06-LC-001`	`R04-BI-030`	yes	no	no	defer

7. Proposed Changed-File Allowlist

If a future patch is authorized, Owner / CCRA must explicitly approve the allowlist before Codex edits any file.

Proposed allowlist:

selector/selector_rules.json
selector/selector_calibration_inputs.json
scripts/run_round06_1_selector_patch_audit.py
tests/test_round06_1_selector_patch_audit.py
selector/README.md
scripts/README.md
docs/DECISIONS.md
reports/round06_1_selector_patch_audit/*
reports/round06_1_post_patch/*
ccra_review_bundle/round-06.1_*/*

Current decision:

allowlist: not approved

8. Explicit Disallowlist

These files must not be edited in Round 06.1 unless Owner / CCRA separately authorize the exact change:

models/qpi.model.json
models/intellectual_archaeology.model.json
cards/qpi.md
cards/intellectual_archaeology.md
tests/qpi.regression.json
tests/intellectual_archaeology.regression.json
tests/regression_cases.json
sources/source_articles.json
sources/source_excerpts.json
lifecycle status fields

Current decision:

disallowlist: active

9. Calibration Decision

Current decision:

calibration JSON additions: not authorized

Before patch, Owner / CCRA must choose one:

[ ] no calibration additions in Round 06.1
[ ] add only approved structured prior-QPI fixtures
[ ] add only approved direct-IA exception fixtures
[ ] add approved structured prior-QPI and direct-IA fixtures

Exact fixture IDs must be listed here before implementation:

approved_calibration_fixture_ids:
- TBD

10. Regression Decision

Current decision:

regression additions: not authorized

Before patch, Owner / CCRA must choose one:

[ ] no regression additions in Round 06.1
[ ] add exact high-risk guardrail cases only

Exact regression case IDs must be listed here before implementation:

approved_regression_case_ids:
- TBD

11. Required Diff Protocol

A future patch must capture deterministic before / after behavior for:

approved target fixtures;
existing selector calibration inputs;
existing selector regression cases;
Round 04 blind pool cases if included in the approved baseline scope;
QPI and Intellectual Archaeology regression suites.

Each diff row must include:

case_id
source_collection
targeted
before.selected_model_ids
before.rejected_model_ids
before.no_call
before.scores
before.reasons
before.penalties
after.selected_model_ids
after.rejected_model_ids
after.no_call
after.scores
after.reasons
after.penalties
changed_fields
owner_disposition
ccra_disposition
follow_up_action

If selector behavior changes and no before / after diff exists, Round 06.1 cannot close.

12. Non-Target Behavior Disposition

Every non-target behavior change must receive Owner / CCRA disposition:

accept / reject / defer

If any non-target behavior change lacks disposition, Round 06.1 cannot close.

13. Lifecycle Guard

Round 06.1 must not promote model lifecycle status.

Required closeout statement:

QPI lifecycle status unchanged.
Intellectual Archaeology lifecycle status unchanged.
Round 06.1 selector support does not constitute model stability upgrade evidence.

14. Final Authorization Block

Patch execution is currently blocked.

To authorize patch execution later, Owner / CCRA must fill this block:

Owner authorization date: TBD

P01 Round 06 close: accept / reject / defer
P02 Implementation scope: accept / reject / defer
P03 Fixture set: accept / reject / defer
P04 Changed-file allowlist: accept / reject / defer
P05 Calibration decision: accept / reject / defer
P06 Regression decision: accept / reject / defer
P07 Diff protocol: accept / reject / defer
P08 Non-target disposition: accept / reject / defer

Round 06.1 selector-support patch:
[ ] authorized
[ ] not authorized

Owner notes:
TBD

15. Current Stop Condition

Round 06.1 planning is complete.
Round 06.1 selector-support patch is not authorized.
Codex must stop unless Owner / CCRA explicitly approve all patch gates.


我的建议：第一个文件可以现在落库；第二个文件也可以落库，但它的状态必须保持 `authorization_checklist_not_approved`，避免被误读成实现授权。

34 KiB Raw Blame History Unescape Escape

1

1. 总体判断

2. R06-Q01 至 R06-Q23 判断表

3. 我会要求 Codex 先改的地方

4. Calibration / regression 的判断

5. 是否授权 Round 06.1？

2

1. 文档总体一致

2. 但有一个状态文字小不一致

3. 是否继续：我的判断

4. 唯一可以继续的安全动作

5. 给 Codex 的直接指令建议

Web CCRA 判断

我接受什么

我不接受什么

你现在应该授权什么

给 Codex 的授权文本

3

文件 1：12_OWNER_DECISION_ROUND06_CLOSE_AND_R061_PLANNING_COMPLETE.md

2. Accepted Boundary

3. Accepted Contract Edits

4. Explicitly Authorized Actions

5. Explicitly Unauthorized Actions

6. Lifecycle Boundary

7. Future Patch Requirement

8. Final Owner Decision

3. Patch Authorization Gates

4. Minimum Approval Required Before Patch

5. Candidate Implementation Targets

6. Candidate Fixture Approval

7. Proposed Changed-File Allowlist

8. Explicit Disallowlist

9. Calibration Decision

10. Regression Decision

11. Required Diff Protocol

12. Non-Target Behavior Disposition

13. Lifecycle Guard

14. Final Authorization Block

15. Current Stop Condition

34 KiB

Raw Blame History

文件 1：`12_OWNER_DECISION_ROUND06_CLOSE_AND_R061_PLANNING_COMPLETE.md`