本文へ移動
cccskills
無料GitHub で公開

main-judge

Use when adding, changing or reviewing Anabasis's Main Judge: the battery review, its recorded disagreements and the advice they feed. Keeps the Judge outside verification, evidence-bound, separately reported, and unable to alter pass, acceptance, claimability or adoption.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md5.5 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

The Main Judge

The Main Judge advises. It may not change pass, acceptance, claimability or adoption. A disagreement with the verifier is a reason to audit the verifier, never to change a score.

ComponentInputOutputAuthority
Main Judgeaccepted battery cases with a boolean verifier verdictJudgeEvidencemay flag verifier doubt; cannot change truth

resolveSlots chooses the review slot's kind and model; reviewSlot in src/review/review-session.ts resolves it through the shared pi provider layer and openReviewSession opens one pi host session on the caller's tools and framing. The Judge battery review is the only caller on that slot for verdict sessions (judgeSessionFor); the diagnosis reader and Epoch Reviewer open their own sessions on the same slot through the reader. The kind set is closed. The slot was called judge (--judge-backend, HARNESS_JUDGE_BACKEND) before 2026-08-19 and those spellings no longer resolve; on a tree that predates the rename, read judgeBackendFor in src/run/judge-session.ts instead.

No control census

The operator removed the control census on 2026-09-14: in Sol run truss-sol-20260913T192126453Z-0123e5 it was invalid in 3 of 3 rounds, so its standing decided nothing useful. A new review sends only battery cases to the Judge and writes no census sessions, judge/control-census-sample.json, judge/bait-corpus.json or judge/review-standing.json. Evidence records judge: "unvalidated" over zero controls, and a record with a control census behind it is refused (operator decision 2026-09-22). Do not reintroduce a validity rule or a bait corpus.

JudgeDecision is closed in source: non-result, incomplete-census and advisory-comparison, which a complete review records; an undecided recorded between 2026-09-29 and 2026-09-30 counts as an answer in each. Claims recorded before 2026-09-29 may also carry no-battery-verdicts. A disagreement means audit the verifier. Never rescore.

Fresh sessions

Keep these frozen judgeDeAnchoring rows at the judge input boundary:

  • fallbackConfigured: false
  • commitAllTasksInPhaseZero: true

Every subject uses a fresh session. Never add a fallback model.

Judge framing

Keep all review instructions in the one model-visible prompt file. State the request shape: a battery case has its bound public task, the submitted artifact and the domain contract. The verdict is pass or fail: a fail rests on a breach the shown material itself states, and a pass claims only what the Judge could read, leaving a run and any multi-step figure to the verifier. The undecided verdict added on 2026-09-29 was the answer to 656 of the 657 cases answered under it, and is gone. Do not say what would make an artifact valid.

Disagreements are advice

Every disagreement is recorded without a materiality threshold, and none of them changes a score. judgeCaseKind in src/review/judge.ts names how each answered case stands against the verifier (agree, veto, unconfirmed fail, disputed pass), and every reader counts by it; an undecided on an older record is no answer to compare. judgeExit in src/analyse/judge-reviews.ts has two kinds: none when the Judge contradicted the verifier on no reviewed verified case, and advisory otherwise, with a count per kind. It records no finding and routes to no owner (judge-reviews/v15). Disagreements in both directions enter the rebuild advice packet as advisory rows named by family, and the exit reason is its judge line; only families and counts cross to authoring. A review recorded under another schema is refused rather than read, except that the whole-run digest still reads v13 and v14, whose census, contested rows and exit it reads as the current ones.

A recorded claim also names the case ids that disagreed on that battery. Those ids sit beside the claim; they change no score, readiness, adoption or statement.

Check a proposed judge dimension

  1. If schema, validator, contract, or host verifier can decide it, do not add it.
  2. Check whether it repeats a verifier check or proxy.
  3. Check that it is useful and visible without hidden truth.
  4. Check that disagreement can be audited.

Inputs, checks, and report

Give the Judge the exact operator request, the bound public task, the submitted artifact, the public artifact schema, the projected Builder tool contract, declared runtime facts and an explicit output schema (JudgePublicDomain in src/review/judge-contract.ts). It receives no solve trace, no Built system prompt and no verifier output. The tool contract is load-bearing: run 69's distrust-verifier hold was manufactured because the binding conventions lived only in agent/tools.ts text the Judge never saw. Never give hidden expectations, control literals, verifier verdicts in an independent pass, verifier remedies, or approval instructions.

  • Validate the exact output schema and pin model and prompt identity.
  • Check that cited evidence was visible to the judge.
  • Keep transport non-results separate from low scores.
  • Treat every disagreement as advice for the verifier's owner, never as a verdict.
  • Confirm the sanitizer ran.

Report purpose, input visibility, raw assessment, disagreements and their advisory exit, and that the Judge owns neither truth, acceptance nor adoption.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Use when deciding which Anabasis stage owns a change, when carrying one bounded build/measure slice from the user's prompt to evidence, or when deciding what may enter a run: the exact one-line prompt, context files, public catalogues, research and solve-side public data. Maps the product loop, the input contract, the authoring sessions, the adoption gates, the four owners, and validate-or-reopen behaviour. Not a long-running campaign controller.

日本語の概要は準備中です。原文の説明を表示しています。

s-smits/anabasis622026年10月13日 更新

Use after an Anabasis run, comparison, or system change and before claiming improvement. Attributes movement to one owner, reports identities and censored denominators, separates deterministic proof from model judgement, keeps measurement validity apart from observed exploitation, and states what remains unrun or provisional.

日本語の概要は準備中です。原文の説明を表示しています。

s-smits/anabasis622026年10月13日 更新

Use for 1-7 independent read-only sessions investigating one bounded Anabasis failure, diff, or design question before a fix. Defines non-leading prompts, evidence packets, symmetric fault hypotheses, self-falsification, source adjudication, and concise synthesis. Do not use for whole-run coverage at any session count; use whole-run-investigation instead.

日本語の概要は準備中です。原文の説明を表示しています。

s-smits/anabasis622026年10月13日 更新

Get told the moment a live Builder session is blocked (five refused submits in a row, a repeated findings set, twelve checks without acceptance, two environment previews in a row, two hours since a submit or rehearsal last returned) and fix the owner iteratively: read, fix on the owning PR, then keep the run or kill it and relaunch on the upgraded system, and rearm. Use while a paid run is live and the operator wants blocking caught and repaired, not only reported.

日本語の概要は準備中です。原文の説明を表示しています。

s-smits/anabasis622026年10月13日 更新

Shared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable or AI-navigable, or when another skill needs the deep-module vocabulary.

日本語の概要は準備中です。原文の説明を表示しています。

s-smits/anabasis622026年10月13日 更新

Launch, start, monitor, and collect independent Codex subagents (gpt-6-luna at high, xhigh or max; gpt-6.1-sol for small review batches) for bounded parallel work, from Codex or from Claude Code, including requests supplied as a Markdown file of session prompts. This skill owns Codex subagent transport even when another investigation or review skill defines the questions. Use whenever the user asks for Luna or Sol agents, a swarm, many sessions, a concurrency test, a particular reasoning effort, a Codex subagent from Claude Code, or later collection of reports. Distinguish launch-only requests from requests to wait, collect, or synthesise. Route high and xhigh through the direct launcher; for 16 or more sessions, also invoke the direct launcher instead of native spawn_agent.

日本語の概要は準備中です。原文の説明を表示しています。

s-smits/anabasis622026年10月13日 更新

s-smits のスキルをすべて見る

このスキルの問題を報告する