Plan Council

Review one authored plan or spec with declared cross-family judges. Disagreement is evidence: preserve each judge's independent rubric, validate claims against live local state, and let the conservative non-author-family score govern disputed lanes.

Scope

The council runs 3 visible seats from 3 model families, self-invented rubrics, live-repo read-only validation, measured same-family bias, a lift round, and exact-head review loops.

Output contract

EVERY artifact this protocol emits — ballot, lift table, convener report — leads with its table; environment caveats and preamble go BELOW the first table, never above it.

Route first

Load-bearing rules from those references, so a router that stops here still gets them right:

• Anonymized ranking: no authorship leak (never name a participant engine or say "my proposal"), a numeric Score: per candidate, an ## Inputs read section, and the sentinel as the last line. Resolve every cited symbol/line against the code before hand-off.
• Bulk judging: probe brain_search before fan-out (an enrichment-locked DB → HOLD); stage under a durable eval_results/<campaign>/, never /tmp; reason per prompt (scripts may validate, never produce verdicts); completion = DONE/<worker>.done files; append-only collab sections; flag evidence_degraded and filter it before bulk-apply; red-team gate before any bulk-apply.
• Assumption check: run it before any overnight or multi-agent sprint. Every PHANTOM claim is killed or verified; quantitative claims need a source.

Five laws

1. Seats

• Convene 3 voting seats from 3 model families in visible cmux panes: R1 claude.judgment, R2 codex.implement, and the separately declared R3 Fable 5 seat. Never use background subagents.
• Pin every seat explicitly and verify the effective model from the pane. The Fable seat uses raw claude --dangerously-skip-permissions --model claude-fable-5 because the repoGolem launcher cannot express it.
• Supply no shared rubric. Each judge invents named dimensions and weights totaling 100.
• Each judge scores every lane 1-10, checks the live repo read-only, makes findings falsifiable, gives GO / NO-GO / CONDITIONAL GO per gated unit, names its top three changes, signs its family, and ends with its agreed sentinel.
• Each judge's FIRST output line after the ballot heading must be the scorecard TABLE header; judges write no prose before it.

Use references/seat-briefs.md and workflows/run-council.md.

2. Merge

• The plan author holds no seat and scores nothing.
• Do not average disagreement away. For each disputed lane: merged(lane) = min(non-author-family judge scores).
• Report same-family bias per lane and the round mean. Flag a same-family judge at ≥ +2 above the conservative non-family read.

python3 <plan-council-skill-dir>/council_bias.py <ballots-dir> --author-family <author-family>

3. Lift round

For every lane below 8, ask every judge for THE ONE artifact that would lift it above 8 and an explicit research call. The governing instruction is:

one artifact each, and it must be the thing whose absence is why the score is low — not a wish list.

Run workflows/lift-round.md. Conditions addressed to the author are work: apply them and continue. They are not ratification theater. Irreversible state changes retain their own receipt-first HARD STOP.

4. Exact-head review

Every execution verdict names the exact head SHA. ACCEPT / ITERATE / MUST_FIX without that SHA is void. The ACCEPTED head is the merged head; any later head movement makes the verdict stale and requires re-review.

Every review round declares its axis and method. A stale-only round counts toward a published stop-line only when it explores the same axis as the prior round and uses an adversarial method that attempts to defeat the rules. A confirmatory happy-path round never counts; changing axis resets the counter. Publish axis, method, counter, and stop-line before dispatch. See workflows/exact-head-review.md.

5. Conditions are work

Apply council conditions assigned to the author and go. Do not park approved work behind another ratification step. This does not weaken independent data-safety gates for migrations, backfills, or other irreversible operations.

Validate ballots

python3 <plan-council-skill-dir>/council_lint_cli.py ballot.md \
  --lane "W3 preconditions" --lane "3a source_class" \
  --author-seat <author-voting-seat-or-nonseat-id> --sentinel DONE_COUNCIL_R1
python3 -m pytest <plan-council-skill-dir>/tests/ -q

author-scored enforces the operator-declared author seat; detecting covert authorship would require adding agent identity to the signature contract.

The mechanical contract has 10 gate rule IDs + 3 warning types. Checkability is ballot-scoped: unfalsifiable-finding rejects a ballot when none of its findings has a concrete locator, then warns on individual uncited findings when at least one sibling is cited. The refuter and receipts warnings remain visible but non-gating; the source corpus proves that strong ballots can omit those house-style forms and can include a paper-contract finding beside live verified findings.

Anti-patterns

Receipts

The numeric claims and worked example come from the read-only source trail documented in references/worked-example.md.