← all meta proposals

Add genesis JSON output validator harness with single retry

council rejected HARNESS reversible: simple 6h proposed 30 May 2026
What is the proposed change?
Wrap the genesis LLM call (Sonnet 4.6 OAuth) in a harness that JSON.parse + schema-validates the proposal array against the meta_engine genesis schema (title length, required fields, change_type enum, solo_time_estimate ≤16, m_a_self_score keys present, reversibility enum). On parse or schema failure, the harness retries once with the validator error appended to the prompt as 'Your previous output failed validation: <err>. Return corrected JSON.' If the second attempt fails, the cycle records status='genesis_malformed' with the raw output for audit, rather than crashing downstream filter_score.
Target files
hypothesis_engine/moves/genesis.js hypothesis_engine/lib/validate_proposal.js
Expected effect
Genesis-induced cycle failures (silent outage class — see fix in commit 0f2d20d for Bedrock Opus 4.6) drop to zero over 60 days. Downstream filter_score.js never receives a malformed proposal in production. Retry rate sits between 5-15% (signal that the validator is firing) but final success rate ≥ 98%.
Falsifier — what would prove this wrong?
If the second-retry success rate is <70% over 30 days, the validator is too strict or the model is consistently misreading the schema — degrade to a soft-warn (log but pass through) and review the schema for ambiguity. If retry rate is 0% over 30 days, the validator never fires and provides no value; remove it.
Evidence that triggered the proposal
  • D — commit 0f2d20d: fix meta_engine genesis 5-cycle silent outage
  • T — SerpAPI search: 'structured output JSON' + 'agentic workflow failure modes' (validator + retry is consensus pattern)
  • E — move_cost: meta_genesis $0 (OAuth) — retry cost is negligible

Proposer self-score

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

AxisScore
specificity3
falsifier3
solo feasible3
blast radius2
composability3
reversibility3
Disposition
Rejected at the council verdict. The two-judge council did not find the case strong enough to advance to Commander review.

Evaluation history

WhenMove
2026-05-30 04:11meta_council_verdict
2026-05-30 04:07meta_argument
2026-05-30 04:06meta_filter_score
2026-05-30 04:03meta_genesis