llm-vision-convergence-loop
Use when an LLM judge (vision or text) is the acceptance gate for subjective quality (render realism, design polish, doc quality) — run a fresh-context, schema-validated, two-consecutive-approvals-at-
SKILL.mdproject-trajectory/skills/llm-vision-convergence-loop/SKILL.md
GitHubWhen to use. Any gate where the pass/fail judgment is an LLM's. Why: judge scores carry a measurable noise floor (we measured ±0.29 mean and MAJOR/MINOR severity flips on byte-identical images); a single APPROVE is a coin flip near the bar.
Procedure.
- Pin a rubric (hash it); build an evidence pack; compute a content hash over everything that produced the evidence. Run cheap mechanical probes first — don't spend a judge round on what a histogram can reject.
- Spawn a fresh-context reviewer per round from ONE canonical prompt template, instantiated verbatim (only verdict filename + date substituted). Never freehand the prompt — a freehand round produced a schema-invalid verdict and was voided.
- Validate the verdict mechanically (rubric hash, content hash, every dimension with evidence, mandatory defect sweep); a schema hole is a rejection, never a silent pass.
- Converge on two consecutive APPROVEs at one content hash. Triage findings: real → fix; contradicted-by-evidence → decide-and-log the contradiction; lottery re-draw on multiply-approved content → decide-and-log citing the approval history. Every dismissal gets a written rationale.
- Done when: the recorder (not you) prints converged, the noise-floor caveat is disclosed to the human owner, and every probe that guarded a fixed defect class is pinned so it can't silently regress.
Knowledge: FIELD-KNOWLEDGE-NOTHOMEWRECKER.md §F1.