Repository navigation
Description
Summary
Update the knowledge-authorship model admission list (.github/knowledge-model-policy.json and docs/knowledge-model-policy.md):
Admit
gpt-6.2-astra(GPT-6.2 Astra)gpt-6.2-sol(GPT-6.2 Sol)
Retire (remove admission)
gpt-6-astra(GPT-6-Astra)gpt-6.1-sol(GPT-6.1 Sol)gpt-6.1-terra(GPT-6.1 Terra)
Claude Opus 5.5 and Claude Fable 5.2 are unchanged.
Background
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, and both are available in the API as gpt-6.2-astra / gpt-6.2-sol (announcement, GPT-6.2 Astra docs, GPT-6.2 Sol docs). GPT-6.2 Sol is positioned as the successor to GPT-6.1 Sol and GPT-6.2 Astra as the new frontier model, so the owner wants knowledge authorship to move to the 6.2 generation and stop admitting the previous GPT entries.
Proposed scope
Follow the same shape as the Terra admission (#298) and the Fable 5.2 admission (#439):
- Add the two exact canonical IDs, no aliases, at
xhighandmaxonly, matching the existing GPT entries. Before merging, the implementing PR must check the canonical IDs and support forxhigh/maxagainst the official model docs and record the check date in the policy doc. If either effort is not supported, flag it here rather than silently changing the effort set. - Remove
gpt-6-astra,gpt-6.1-solandgpt-6.1-terrafrom the allowlist and the owner table, and drop or rewrite their verification paragraphs. - Update
tests/test_knowledge_model_policy.py(andtests/test_release_notes.pyif it pins the model list). Cover:- The new IDs are accepted at
xhigh/max. - Other efforts and inferred aliases (e.g.
gpt-6.2,astra,sol) are rejected. - The retired IDs are now rejected.
- The new IDs are accepted at
Open question for reviewers
Should already-merged knowledge contributions that declare the retired models stay valid? Today, the policy is evaluated on new PRs against trusted main, so existing contributions are not re-evaluated. Please confirm that is the intended behaviour. Also confirm whether a short transition window is needed for in-flight PRs that already declare a retired model (they would need a model change or an exception after this lands).
Out of scope: evaluator logic, exception process, protected paths, and trusted-main policy selection are unchanged. This is an admission-policy change only; it does not qualify knowledge content or run a model.
Activity
Dong1017 commentedon Nov 18, 2026
Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).
Assessment: The scope is clear and consistent with the earlier admission changes (#298 Terra, #439 Fable 5.2). I'd treat it as ready to implement once the two decisions below are recorded. I haven't verified the model IDs or effort support. I can't browse the linked OpenAI docs, and I haven't read the policy files or tests.
What it asks
gpt-6.2-astraandgpt-6.2-solatxhigh/maxonly, with no aliases.gpt-6-astra,gpt-6.1-solandgpt-6.1-terra..github/knowledge-model-policy.json,docs/knowledge-model-policy.md,tests/test_knowledge_model_policy.py, andtests/test_release_notes.pyif it pins the model list.Gaps and risks
xhigh/maxsupport are asserted but not yet confirmed. The implementing PR should record the check date in the policy doc, as the issue says. If either effort isn't supported, the PR should flag it here and not change the effort set quietly.xhighandmax.high,medium,low), not just one.gpt-6.2,astra,sol).gpt-6.2-terraorgpt-6-solthat might match by prefix.Open question: existing merged contributions
Suggested next steps
gpt-6-astra,gpt-6.1-solorgpt-6.1-terra.Dong-bot (authorized delegate) triage; comment only.
david6666666 commentedon Nov 18, 2026
A cutover detail: changing the policy JSON or docs forces a full refresh of open PRs targeting
main, even if their heads don’t change. I checked the current declarations and protected-file diffs: #456 declaresgpt-6.1-soland #480 declaresgpt-6-astra; #457’s Sol mention is reviewer-only. Could we decide how to handle #456 and #480 before cutover, then check their individual declaration statuses after the push? A green sweep job can still leave either PR red. Already-merged contributions are outside that sweep.Dong1017 commentedon Nov 18, 2026
Posted by an agent on behalf of @Dong1017 (incremental review only; the in-flight PR list (#456/#480), the doc-check date and alias/effort test coverage were already raised by Dong-bot and @david6666666 above and are not repeated).
Two test-side details the implementing PR will hit, checked at main
c37832c3:attribution()defaults tomodel="gpt-6.1-sol"(tests/test_knowledge_model_policy.py:65), anddeclared()/pr()build on it, so it is the "compliant author" for most of the file (~39 call sites). After retirement, success-asserting tests go red, which is fine, but failure-asserting tests that use the default as the good author stay green for the wrong reason. Example:test_every_material_author_must_qualify(L124) asserts failure for[attribution(), attribution("unapproved-model")], and that would pass even if the second author were never checked. Suggest moving the default togpt-6.2-solfirst, then doing the retirement.test_suspension_overrides_canonical_and_alias(L113-121) indexespolicy["models"]["gpt-6.1-terra"], which raisesKeyErroronce the entry is deleted.test_effort_is_exact_set_membership(L100) ongpt-6.1-sol/gpt-6.1-terraturns vacuous, because an unknown model already errors before effort is checked (tools/check_knowledge_model_policy.py:205-208). Both should switch to the new IDs.Optional: the evaluator already rejects
"suspended": truewith the same verdict as an absent ID (L205, "unapproved or suspended"). So retiring by suspension, rather than deletion, would keep each entry'ssourceURL and verification history in the JSON at no behavioral cost. Either works; removal matches the issue text.Not in scope but adjacent:
tools/skill_eval_pr_cases.json:4pins the Skill Eval executor togpt-6.1-sol. That is not authorship admission, so this policy change doesn't require touching it. Noting it only so nobody greps it up later and treats it as a miss.Checked: issue body, both comments,
.github/knowledge-model-policy.json,tools/check_knowledge_model_policy.pymodel_errors,tests/test_knowledge_model_policy.py, and a repo-wide grep for the retired IDs. Not run: tests. Not checked (unverified): the OpenAI docs for the new IDs and effort support.zengchuang-hw commentedon Nov 18, 2026
独立范围审阅,基于
main@c37832c3c7a6c457f533683b7b3386f2c3a0c004的实际政策 JSON、checker、workflow 与两份测试源码、隔离的现行文档规范,以及官方资料。只使用隔离出的标题目标,不读取其他人的评论或作者转述。以下是验收建议,不是批准政策变更或宣告完成。1. 将“仓库准入退役”与“厂商产品下线”分开。 OpenAI 发布公告及官方 Astra、Sol页面可支持两个新 canonical ID 的存在;这些资料不能代替仓库选择淘汰旧模型的决策依据。当前政策仍允许三个旧 GPT ID,尚无两个新 ID。建议在验收中明确生效版本、旧 ID 是删除还是保留并 suspended、以及在途贡献如何处理,不要仅因新模型发布就宣称旧型号已被供应商停用。
2. 需要正反两面的精确身份矩阵,而非只加两条成功用例。 checker 的 model_errors按 canonical/显式 aliases 匹配;缺失或 suspended 的型号失败,effort 必须同时属于政策值和
{xhigh,max}。建议覆盖新 ID × 两档 effort、三个旧 ID 的拒绝、未声明别名/大小写/空白/非法 effort,以及任一 material author 不合规的混合声明。若保留 suspended 条目,还应验证 alias 不会绕过。官方reasoning 文档说明支持值取决于具体模型;通用枚举不等价于对每个型号/宿主组合的执行证据。本次没有做付费 API 或实际模型调用。特别注意 test_knowledge_model_policy.py:65–121:默认 attribution 仍是旧 Sol,成功矩阵和 suspension fixture 也引用旧型号。只改 JSON 会让很多非模型主题测试继承一个不再合规的默认作者,或在删除条目后访问不存在的 key。应迁移通用 fixture,并独立保留旧型号负例,不能靠删除历史覆盖来消绿。
3. 验收必须包含已存在 PR 的重新评估,而不仅是新 PR。 workflow读取 trusted main;routing:1221–1226在政策输入变化时走 full sweep。因此“曾经成功、仍包含旧模型 material author 的 open PR”是必须覆盖的切换场景。现行文档规范也要求按目标分支政策而非 PR 自带替代政策评估。政策更新不能把真实旧作者身份改写成新模型来通过检查;若需要例外或过渡,须明确其独立规则和证据。当前
assess对 policy_paths 只做分类,不以旧 policy-change 评论自动授予例外(257–275 行),也不能把技术门禁成功误当作维护者已接受政策提案。4. 不要将无关模型透传测试误改成准入测试。 test_release_notes.py:507–515验证用户显式给出的模型 ID 原样进入请求,且未擅自添加 reasoning。它并不验证 Knowledge 作者准入。如果替换示例型号,应保留这两个断言;仓库准入退役本身不意味着通用客户端应改写或禁止该 ID。
**证据与限制:**完成了上述源码静态检查、固定 main 身份核验及官方资料核查;未执行仓库测试、lint、真实模型调用或任何实验,原因是可用执行环境的隔离可信性未确认。尚未获得可安全分离的详细正文 AC;已核对的现行文档要求前瞻适用,不要求重建未知历史身份或提供私密会话日志;隔离视图不能补足本提案的详细生效/过渡 AC。故不能给出整体 PASS。这里没有修改仓库政策、模型设置、权限或恢复被停止的实验。
Dong1017 commentedon Nov 18, 2026
Posted by an agent on behalf of @Dong1017 (RFC owner decision record).
Decision record for #487
Recorded 2026-11-18 against
main76ca44a8; the policy files are unchanged on the currentmaincf671a88. Implementing PR: #499. This record sets the owner's decisions for the PR to implement. It does not approve the PR; independent review and hosted CI still apply to its head.D1. Admit
gpt-6.2-astraandgpt-6.2-solDecision: admit both exact canonical IDs, with no aliases, at
xhighandmaxonly. This is the same shape as the existing GPT entries, Terra (policy: admit GPT-6.1 Terra at xhigh and max #298) and Fable 5.2 (policy: admit Claude Fable 5.2 at xhigh and max #439).Basis: official OpenAI docs, read on 2026-11-18:
lowmediumhighxhighmaxfor both models.none.The model pages themselves do not list effort values, so
xhigh/maxsupport rests on the catalog. The PR records the check date in the policy doc.Not claimed:
Reopen if: the official docs change either ID or stop listing
xhighormaxfor either model, or a short alias is requested. Any change needs a new doc check.D2. Stop admitting
gpt-6-astra,gpt-6.1-solandgpt-6.1-terraby deleting their entries"suspended": true.sourceURLs and their check dates remain in git history and in policy: admit GPT-6.1 Terra at xhigh and max #298.gpt-6.1-terra.gpt-6.1-solandgpt-6.1-terraas replacement models.D3. Merged contributions that declare a removed model stay valid
gpt-6.1-sol) merged on 2026-11-18 and is covered by this.main.D4. Immediate cutover, no transition window
main. There is no grace period in the policy.gpt-6-astra).D5. #480 must re-author the affected part or use the exception process
xhighormax, and declare that for the new head.20851a9c, Add service-reuse evaluation for project campaigns #480 changessrc/inferrsi/knowledge/catalog.pyandsrc/inferrsi/knowledge/knowledge_operations.py, which are protected.maingives success.D6. Other references to the removed IDs stay as they are
docs/reviews/issue91-*.mdandevidence/.tests/fixtures/repo_to_skill/.tools/skill_eval_pr_cases.json.tests/test_release_notes.py:507-515, which checks that an explicit model ID is sent unchanged.gpt-6.1-sol. That is a separate decision.Review points and how the PR answers them
xhigh/maxagainst official docs; record the date; flag rather than change efforts if unsupportedmainPRsgpt-6.1-solmust move first, or failure tests stay green for the wrong reasongpt-6.2-solbefore commit 2 removes anything; a new test asserts the default is admitted. Running the old tests against the new JSON shows 51 failures and the wrong-reason green ontest_every_material_author_must_qualifyKeyErroror turn vacuousPoints raised by a single reviewer
gpt-6.2-terraandgpt-6-solgpt-6.2-astra-latest,GPT-6.2 AstraandGPT-6.2 Solsourceand history in the JSONtools/skill_eval_pr_cases.jsonis not authorship admissiontests/test_release_notes.py:507-515is a transmission test, not admissionDong-bot and the agent comment both act on behalf of @Dong1017, the RFC owner. @david6666666 and @zengchuang-hw are the reviewers outside the owner account.
zengchuang-hw commentedon Nov 18, 2026
Independent incremental review of the Issue-owner-account decision, following my earlier scope review and PR499 review.
The newly explicit D1–D6 requirements resolve the earlier source/scope questions: exact new IDs at xhigh/max, deletion rather than suspension of the three old IDs, no retrospective invalidation of merged contributions, immediate trusted-main cutover without a grace period, and truthful re-authoring or the existing exception process for affected open work. In particular, re-authoring is not permission to relabel a historical material author. This is an account-attributed governing requirement, not an expansion of my authority to implement or merge it.
I independently compared these requirements with PR499's actual three-file diff and unchanged evaluator at head
49c5453c63f7b86e4c466b804f427fec524f1e37: the JSON membership/effort changes, documentation of cutover and historical treatment, positive default-fixture guard and explicit rejection/mixed-author tests align. Protected paths, trusted-main selection and exception eligibility remain unchanged. Official model-catalog verification supporting the new IDs and efforts is recorded in my PR review; no actual model run or quality qualification is implied.I also rechecked the PR metadata: the literal DECISION_LINK placeholder is now removed and this exact decision comment is linked. I therefore close that specific decision-trace objection from my initial PR review.
The execution-evidence gap is separate and still open: at this inspection full-check jobs 134786431885 and 134786431884 remain in progress. Browser, native-host, installed-package and declaration checks now report success, but those do not substitute for the changed policy regression module's completed execution. I have not approved PR499 or treated the current policy proposal as already merged. A full policy-input refresh over existing open PRs remains an important cutover check, not permission to rewrite their declarations.
This round used isolated normative requirements, actual source, current machine status and my own prior findings. No other reviewers' opinions or attribution summaries were used; no repository code/model was executed, no policy/Issue state was changed and nothing was merged.