Implement B.v2: NBJ 5-Q describability shadow gate before first argument move

council rejected GATE reversible: simple 8h proposed 19 May 2026

What is the proposed change?

Add describabilityShadowCheck(hypothesisId) inside argument.js, gated on argument_moves_count===0 (fires only before the first steelman/red_team_kill dispatch per hypothesis, not on every subsequent argument call). Single Gemini Flash call with the NBJ 5 questions (Q1: inputs nameable without vague phrases; Q2: outputs nameable in pre-AI terms; Q3: exception classes named upfront; Q4: accept/reject criteria for one canonical case against external ground truth; Q5: describable without 'AI' as the primitive verb). Returns {passes:[Q1..Q5], composite_score:0-5, kill_recommendation:bool}. Persist to meta_engine/data/shadow_gate_verdicts/describability-YYYY-MM-DD.jsonl and as a moves table row (move_type='shadow_describability_check'). Argument proceeds regardless. Fail-open: if Gemini Flash unavailable, log null verdict and proceed. Shadow output MUST NOT be injected into the context of subsequent scoring moves. Add /commander/<token>/shadow_verdicts route to digest_server rendering a sortable table (date, hypothesis_id, Q1-Q5 pass/fail, composite, kill_recommendation, hypothesis's actual council outcome). Graduation to hard-kill requires: Wilson false-kill upper bound ≤2.5% from accumulated shadow data AND ≥5 ROBUST candidates evaluated by live council during shadow period AND Commander explicit approval.

Target files

hypothesis_engine/moves/argument.js meta_engine/data/shadow_gate_verdicts/ digest_server/routes/commander.js

Expected effect

Codex retrospective on 43 S157-scored candidates: gate kills hyp-2026-05-06-847f7e (S157 score 0/5, structurally fragile on all 5 Q dimensions) and does not kill any of the three 5/5 ROBUST candidates (ec4507, 24a849, 3656a0). Spearman rank correlation between gate composite_score and S157 manual scores on the 43 candidates: ρ ≥ 0.6.

Falsifier — what would prove this wrong?

Run Codex empirical on 43 S157 candidates. Gate is mis-designed if: (a) 847f7e does NOT receive kill_recommendation=true; (b) any 5/5 ROBUST candidate receives kill_recommendation=true; (c) Spearman ρ < 0.6. Any single failure blocks graduation to hard-kill permanently and triggers a prompt revision cycle.

Evidence that triggered the proposal

Corpus D: brain/S157_NBJ_DESCRIBABILITY_TEST.md — 43-candidate manual sweep: Q3 (exception classes named) is systematic engine blindspot, 1/43 candidates named exception queue explicitly; 847f7e scores 0/5 while ec4507 scores 5/5
Corpus D: brain/S158_SHADOW_P4_SPEC.md + brain/META_ENGINE_S158_ROUND2_SYNTHESIS.md — B.v2 COMMIT_WITH_REVISION with statistical graduation criterion (Wilson bound + ≥5 ROBUST candidates), fail-open, isolation requirement, Commander explicit approval gate
Corpus E: Kill reason distribution — council_verdict_unanimous_kill n=2, fatal_objection_both_confirm n=2; these late-stage kills cost ~$0.18/run; pre-debate describability check costs ~$0.001/call, asymmetry justifies early shadow

Proposer self-score

The proposer scored its own draft on these axes (0-3 each) before submitting.

Axis	Score
specificity	3
falsifier	3
solo feasible	2
blast radius	2
composability	3
reversibility	3

Disposition

Rejected at the council verdict. The two-judge council did not find the case strong enough to advance to Commander review.

Evaluation history

When	Move
2026-05-23 04:36	meta_council_verdict
2026-05-23 04:22	meta_argument
2026-05-19 13:12	red_team_kill
2026-05-19 11:30	steelman
2026-05-19 10:08	meta_filter_score
2026-05-19 10:04	meta_genesis