Skip to content

docs: trajectory-distillation design (weakness → executable held-out tasks) - #48

Merged
Niraya666 merged 2 commits into
mainfrom
codex/trajectory-distillation-design
Jul 16, 2026
Merged

Niraya666 merged 2 commits into
mainfrom
codex/trajectory-distillation-design

Conversation

@Niraya666

Copy link
Copy Markdown
Owner

新设计文档 docs/trajectory-distillation.md:把已录制轨迹 + 挖出的弱点蒸馏成可执行的 EvalTaskSpec 候选(fixture + prompt + argv verifier),经确定性 fail-before 校验与人工 accept 后进入 held-out suite——裁判的任务集随观察到的弱点生长,不再停在手写任务上。

补上的正是 §5.6 那半环

weakness-mining-design.md §5.6 的 evaluator surface 原标"半闭环"。本设计把它变成与其他 surface 同构的 proposal-first 闭环:loop 产出裁判候选,人是唯一的 promote 通道(git commit),冻结面纪律原样保留。已交叉更新 §5.6 表格与 §9 的"仍未做"行(有效性判据:设计已就绪、实现未动)。

设计要点

  • 零新格式:产物就是现有 EvalTaskSpec JSON + fixture,load_eval_task/两个 gate 原样消费。
  • fail-before 硬门:裸 fixture 上 verifier 必须失败,挡"空过"的裁判。
  • lineage 排除:从 session S 蒸馏的任务不得验证 S 催生的 edit——不许拿作业题当期末考。
  • 安全命门:candidate 住 gitignored 的 .agent-harness/,suite 住 git 管控的 tests/e2e/,okc distill 无任何写 suite 的代码路径。
  • 诚实边界:不是所有机制都能蒸(metrics-only 的显式 skip);fixture 是机制级近似,不是案例复现;风险 Add run policy summary compact and skills #6(suite 漂向历史弱点)只有纪律缓解。

纯文档;分阶段落地 P0–P3 已列,未动代码。

… tasks)

New design doc: distill recorded trajectories + mined weaknesses into
EvalTaskSpec candidates (fixture + prompt + argv verifier), deterministically
verified (fail-before) and human-accepted into the held-out suite -- the judge's
task set grows with observed weaknesses instead of staying hand-written.
Cross-link weakness-mining-design.md: the evaluator surface row (was
'semi-closed') now points here; the effectiveness-signal residual is marked
design-ready.
…op line number

The lineage-exclusion filter reads source_sessions at holdout-suite assembly,
but provenance.json stays in the gitignored candidate dir and does not survive
promote. Carry source_sessions on task.json itself (optional EvalTaskSpec field,
empty for hand-written tasks so they are never excluded). Also: correct the
held-out gate entry points (okc propose --llm --validate-held-out / okc edits
--revalidate), and drop a rot-prone mining.py line number.
@Niraya666
Niraya666 merged commit d973599 into main Jul 16, 2026
2 checks passed
@Niraya666
Niraya666 deleted the codex/trajectory-distillation-design branch July 16, 2026 10:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant