minecraft-run-report-author

name

minecraft-run-report-author

description

Write, review, validate, or prepare Minecraft runtime run reports, HTML/static reports, model-comparison summaries, Qwen/OpenAI/Gemini writeups, visual evidence reports, public artifact summaries, or Korean/English "보고서" requests from this repo's social-cycle, action-skill, provider, visual, or embodied co-actor legibility artifacts including transition-row, public-history, experiment-declaration, prediction, and score-report bundles. Use before drafting report prose or publishing experiment conclusions.

Minecraft Run Report Author

Use this skill to keep reports evidence-grounded. The goal is not pretty prose; it is preventing repeated overclaims, missing quota context, broken artifact refs, and screenshot-only conclusions.

Quick Start

1. Identify the exact artifact bundle. Do not use "latest" blindly. Supported primary inputs include social-cycle-run-report/v1, legibility-session/v1, transition-row/v1 batches, visual evidence bundles, and curated public report directories.

2. If the run used a live provider, locate the preflight artifact or run provider-quota-preflight before rerunning anything.

3. For social-cycle-run-report/v1, run the runtime-review summarizer first:

   bun run .agents/skills/minecraft-agent-runtime-review/scripts/summarize-social-cycle-report.ts <report.json>
   

4. Run the readiness check before final report prose. For legibility sessions, pass session.json; for social-cycle reports, pass report.json. Use --publishable for public/HTML/model-comparison reports where missing provider preflight or scoring/preflight refs should fail instead of warn:

   bun run .agents/skills/minecraft-run-report-author/scripts/report-readiness-check.ts <report-or-session.json> --publishable
   

5. If the report will compare models or become a public/HTML artifact, also run the focused audit/review commands that apply to the artifact:

   cd probe
   bun run probe:social-cycle-audit -- <report.json> --audit-report <audit.json>
   bun run probe:social-cycle-review -- <report.json> --markdown <review.md>
   bun run probe:social-cycle-benchmark-metrics -- --report <report.json> --out <metrics.json> --html <metrics.html>
   bun run probe:social-cycle-benchmark-score -- --metrics <metrics.json> --out <score.json> --html <score.html>
   

Required Report Shape

Start from evidence, then write prose. Include:

• Recording verdict: whether artifacts are complete enough to review.
• Experiment verdict: what the run actually supports.
• Claim table: material claim, social claim, visual claim, provider/cost claim, and exact source artifact/ref for each.
• Run identity: command, scenario, seed, actor, provider/model, cycle count, visual profile, and date.
• What happened: outcome distribution, verified mutations, blockers, observation loops, action concentration, retry constraints, and social signals.
• What was recorded: report JSON, actor workspace, provider snapshots, evidence refs, visual captures, preflight artifacts, usage ledger refs.
• What is not proven: missing refs, weak screenshots, provider fallback, stale/local-only artifacts, budget caveats, or unsupported research claims.
• Next experiment: one narrow follow-up run or implementation fix.

For legibility-session/v1 or transition-row/v1 reports, also include:

• Row/scoring verdict: row count, valid/scorable rows, label-lock status, public-history leakage checks, score-report support, and which active-plan gate the bundle does or does not satisfy.
• Row quality: archived rows, non-excluded rows, scorable rows by layer, partial/excluded row reasons, and contract failures.
• Legibility provenance: seed/reset records, active actor count, condition, scenario family coverage, provider/cost guard, declaration refs, and environment blockers.
• Separation statement: observed_delta is runtime evidence, not actor expected_outcome, provider rationale, report prose, or predictor output.

Evidence Rules

Prefer evidence in this order:

1. Runtime evidence artifacts with verifier-backed world, inventory, container, position, chat, helper-event, or tool-result mutation.
2. transition-row/v1, legibility-session/v1, public-history, score-report, seed/reset, natural-spawn validation, and visual audit artifacts when relevant.
3. Provider input/output snapshots for what the model saw and emitted.
4. Review-summary markdown as an index, not as the source of truth.
5. Human-visible notes from the user.

Do not count provider prose, memory notes, wait, repeated observation, or screenshots alone as physical/social success.

Forbidden Claims

Do not claim these unless artifacts explicitly prove them:

• leaderboard result;
• research conclusion;
• model superiority;
• prediction-prior lift;
• durable society/sociality;
• block identity from pixels;
• progress from animation or movement only;
• provider budget safety without preflight and usage evidence.
• legibility/prediction lift from a bundle with missing label locks, predicted_delta inside rows, failed public-history leakage checks, missing support counts, or score rows joined before labels were locked.

When in doubt, say what the run recorded and what remains unproven.

Reference Files

File	Read when	Contents
references/legibility-session-reporting.md	Reporting legibility-session/v1, transition-row/v1, public-history, prediction, or score-report bundles	Required queries, claim boundaries, and common mistakes

Related Skills

• Use provider-quota-preflight before provider-backed reruns or comparisons.
• Use minecraft-agent-runtime-review when the report depends on behavior-loop diagnosis, action-skill matrix results, or visual artifact interpretation.
• Use minecraft-research-value-harness before turning a report into a research direction or paper claim.