balance-report-critic

Status: v0.1 — minimum viable. Full skill (REPORT_CHECKLIST integration) lands in Phase 1.

When to invoke

Trigger words: "the report shows", "win rate is", "X is better than Y", "above baseline", "trained config", "balance is", "this strategy". Trigger context: about to print a table of percentages, render a chart, or finalize reports/vX.YY/*.md.

The Gate (HARD)

You may NOT present a balance/strategy claim until ALL of the following are explicit in the report:

1. Sample size n — exact integer per cell, not "many" or "enough".
2. Confidence interval — ±X percentage points at the stated n (Wilson, Clopper-Pearson, or bootstrap — name the method).
3. Lobby composition — smart-only, mixed, or which difficulties. Per global CLAUDE.md: balance claims are smart-only lobbies.
4. Rules version — vXpYY of the engine + xlsx the simulation ran on.
5. Baseline comparison — what is "above" or "better than"? Name the baseline run and its n + CI.
6. Cell coverage — which player counts, which founder abilities, which configs were swept. Missing cells = unknown, not "matches baseline".

Minimum viable checklist (v0.1)

1. Read the draft report. For each numerical claim, check if items 1–6 above are stated within 2 sentences of the claim.
2. If anything is missing, edit the report to add it before showing the user — do not present and ask the user to fix later.
3. If a CI was not computed, compute it before presenting (Wilson formula for binomial proportions: standard).
4. If lobby was mixed or baseline is unknown, flag explicitly in the report with "NEEDS CONFIRMATION" — do not paper over.

Anti-patterns

• "Win rate is 52%" without n. ← blocked
• "Above baseline" without naming the baseline. ← blocked
• Presenting CMA-ES training WRs as final numbers. Per global CLAUDE.md, these are inflated; require n=1000 verification before adoption.
• Rounding away the CI ("52% (±0)") when n is too small to compute one.

Cross-references

• After the report passes critique, invoke superpowers:verification-before-completion to confirm the report file is written, committed, and matches the printed claims.
• If a checklist exists (REPORT_CHECKLIST.md), invoke bg-superpowers:checklist-enforcer instead — it subsumes this gate.

TODO (Phase 1 full implementation)

• Auto-detect the report file from the conversation and run lint against it.
• Compute Wilson CI inline when raw counts are available.
• Pull baseline from a project-local config (bg-superpowers.config.json).
• Integrate with takeaway's existing REPORT_CHECKLIST.md as the canonical version.