跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
视频与音频

lofn-qa

Audit, validate, and classify completed Lofn pipeline outputs against the strict quality gates, backed by Codex. Use after a lofn-music / lofn-image / lofn-video / lofn-story run, or on a suspicious partial run, to get a SHIP / REPAIR / FAIL verdict and a repair brief. Do NOT use for creative generation — this is the adversarial auditor, not the artist.

LocalSymmetry · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add LocalSymmetry/lofn --skill lofn-qa

技能说明


name: lofn-qa description: Audit, validate, and classify completed Lofn pipeline outputs against the strict quality gates, backed by Codex. Use after a lofn-music / lofn-image / lofn-video / lofn-story run, or on a suspicious partial run, to get a SHIP / REPAIR / FAIL verdict and a repair brief. Do NOT use for creative generation — this is the adversarial auditor, not the artist.

Lofn QA — Codex-backed adversarial gate

The competitive auditor. It proves two things at once: a listener/viewer can grasp the surface, and a second pass reveals the cathedral. It fails both extremes — impressive obscurity AND competent blandness. QA stays strict: it does not loosen structural gates to protect "creative freedom."

Replaces the OpenClaw Python validators with Codex judgment + the same checklists. Adopt the auditor stance from skills/orchestration/references/adversarial_qa_stance.md first.

Procedure

  1. Run in a fresh, clean-context judgment subagent. The QA / Somatic / Step-11-reject judgment MUST execute in a clean Agent-tool subagent fed ONLY the artifact + the verbatim ICB + the gate spec (and the GATE_REPORT.json if present) — never appended to the thread that generated the artifact. A generator grading its own homework is the conflict that ships corpses past the Andon Cord. Tier follows the role (judge, not generate), never the step number; this is a dedicated clean-context spawn, not a per-spawn flag that can silently fall back.
  2. Identify the run directory + modality.
  3. Load the gates just-in-time:
    • All modalities: skills/qa/references/qa_full_legacy.md (the full tuned procedure — authoritative).
    • Music: skills/qa/references/suno_15_point_qa.md + skills/qa/references/eligibility_7_properties.md. Classify ACCESSIBLE vs AMBITIOUS, then run the 16-point gate.
    • Image: vault/VISION_QA_DEPTH_AUDIT.md (Visual Somatic Gate + 7-element density checklist).
    • Video/Animation: vault/DIRECTOR_QA_DEPTH_AUDIT.md (Cinematic Somatic Gate + 5-element shot checklist).
    • Thresholds: numeric bands are read from vault/gates.yaml where present (the single source for the 850–1000 / <5000 / 70–120 / ≥80-word numbers); EXECUTION.md §4 is authoritative if the two disagree.
  4. Read the deterministic evidence first. If scripts/validate_step.py wrote a GATE_REPORT.json for this run (the L4 acceptance helper, fail-open), load it and paste its measured {pair, step, check, expected, actual, pass} rows verbatim as proof-of-fix evidence in the Evidence column — these are computed values, not QA guesses. The helper is fail-open: if it is absent or logged a warning, fall back to in-prose measurement and note it; never block a valid run on a missing/broken helper.
  5. Verify ICB integrity (cheap proof, not a soul substitute). Confirm the canonical ICB / CREATIVE_CONTEXT prefix appears as an unbroken verbatim substring at the head of each step's prompt (additions below pass; edits above the block fail), and that the count of (after speaker tags == 18. A voice count ≠ 18 or a broken prefix is caught as REPAIR — THREAD LOSS. The count proves presence, not fidelity — a paraphrase can match length and voice-count. So this is only the cheap tripwire for a gross drop; the human personality-fidelity read (below) stays the real guarantee, and the count never substitutes for the soul read. Never summarize, trim, or "optimize" the ICB to make a count easier. 4.5. Blind golden+decoy comparative judging (music finals — the judge's own calibration). For each finalist package, the coordinator (not the judge) assembles a blind set of three unlabeled packages, shuffled: (a) the candidate, (b) one Golden Song payload — from skills/music/references/suno_format_example_{triple_arch,blue_screen,five_wrong_colors}.md or skills/qa/references/golden_song_examples.md (the two judge-side libraries; this is the ONE context where golden payloads belong, generators never see either), (c) a decoy — a known-mediocre package (a SELECTED_FOR_TOP_SIX: No also-ran from an archived run works well). The clean-context judge must rank the three and justify the ranking. Readings:
    • Candidate ranked below the decoy → REPAIR (it lost to known filler).
    • Judge ranks the decoy above the golden → the judge is broken — halt QA, audit the judge context/model tier, re-run; never ship on a broken judge's verdicts.
    • Candidate ranked above the golden → fine, but the justification must name what specifically beats it (guards against reflexive candidate-flattery). This is a checklist the judge cannot pass by agreeing with everything — the mechanism that keeps QA's "no" real.
  6. Run the gate without weakening any check. Apply the Codex-native self-check gates in .agents/skills/lofn/EXECUTION.md §4 for structural/pipeline integrity.
  7. Write QA_REPORT.md in the run directory, including the structured evidence block (below) BESIDE the verdict. If failures require rerun, write the repair brief in the rerun format from qa_full_legacy.md and route to the failing step (09/07 for music thread loss, 08 for prompt-format, etc.). REDIRECT — mandatory when the gate is stuck (EXECUTION.md §7.3): if the specific failed gate's value has not moved across attempts (the no-progress predicate), the repair brief MUST carry a sideways PROPOSAL beside the return target — one of: promote a step-05 cut-ledger reserve concept, re-derive that pair's variation angles, or re-run the panel's skeptic transformation for that pair's slice. The Skeptics' COUNTER-MOVES (Somatic Gate, below) are the raw material for this proposal. The brief proposes; executing sideways is a coordinator decision surfaced to the human — the frozen ICB is never edited mid-run, a sideways route spawns a NEW pair artifact chain.
  8. On SHIP (every shipped/selected piece): append ONE curated failure-ledger entry to `vault/CO
阅读完整来源 ↗