Reweight synthetic.jsonl corpus by per-item graduation lift

filter rejected CORPUS reversible: simple 7h proposed 31 May 2026

What is the proposed change?

For each item in synthetic.jsonl, compute graduation_lift = (filter_kept_rate when this item was retrieved) / (baseline filter_kept_rate). Persist lift to a sidecar synthetic.weights.json. Modify search.js retrieval to sample proportional to lift (clipped to [0.1, 3.0]) instead of uniform. Items with <5 retrievals keep uniform weight until they have signal. Recompute weights weekly via a cron.

Target files

hypothesis_engine/corpora/synthetic.jsonl hypothesis_engine/moves/search.js

Expected effect

Filter-kept rate of genesis output rises 10-25% within 30 cycles, because dead synthetic items stop being retrieved at uniform rate. Top-quartile synthetic items get retrieved 2-3x more often.

Falsifier — what would prove this wrong?

Run 30 cycles with new weights, 30 with uniform (A/B by cycle parity). If filter-kept rate delta <5% with p>0.1, reweighting is noise and revert. If filter-kept rises but argument-pass rate falls (overfitting to filter), revert.

Evidence that triggered the proposal

D — brain/S162_GEMINI_35_FLASH_FILTER_SWAP_DROPPED.md (empirical-falsifier discipline)
E — move cost rollup: search retrieves from synthetic at uniform weight
T — hypothesis_engine/corpora/synthetic.jsonl (modified per git status)

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
2026-05-31 04:06	meta_filter_score
2026-05-31 04:03	meta_genesis