Add retry-with-vendor-fallback harness around filter_score judge call

filter rejected HARNESS reversible: simple 4h proposed 11 Jun 2026

What is the proposed change?

Wrap the filter_score judge call (currently callCodexGpt55) in a 3-attempt harness: attempt 1 = primary judge; on timeout/5xx/parse-fail, attempt 2 = Gemini 3.1 Pro; attempt 3 = Grok-4. Each attempt has a 45s timeout. Record which vendor produced the final score in engine.db column `judge_vendor_used`. If all three fail, kill with reason='judge_unavailable'. The harness is a single async function ~40 lines.

Target files

hypothesis_engine/moves/filter_score.js hypothesis_engine/llm.js

Expected effect

Cycles aborted due to judge transient failure drop from current ~5-8% (per E corpus) to <1%. Distribution of `judge_vendor_used` across 100 cycles reveals which vendor is least reliable.

Falsifier — what would prove this wrong?

After 100 cycles, if judge_vendor_used shows >95% primary (Codex), the harness adds no signal — primary is reliable and fallback is dead code. Acceptable failure mode: keep harness only if fallback fires ≥3% of the time.

Evidence that triggered the proposal

E — engine kill-reason distribution shows judge-call failures in top 5
D — hypothesis_engine/llm.js (existing multi-vendor callers ready to compose)

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	3
blast radius	2
composability	3
reversibility	3

Disposition

Rejected by filter_score. The proposal did not meet the bar for specificity, falsifiability, or solo-feasibility.

Evaluation history

When	Move
2026-06-12 04:41	meta_filter_score
2026-06-11 04:03	meta_genesis