Reweight synthetic.jsonl corpus contribution by graduation-yield

filter rejected CORPUS reversible: medium 5h proposed 11 Jun 2026

What is the proposed change?

Add a per-row `weight` field to synthetic.jsonl (default 1.0). Compute weights as a one-shot offline job: for each synthetic seed, count how many cycles it appeared in that produced a graduated candidate vs. killed candidate. Weight = graduated/(graduated+killed+1). Genesis sampler reads weights and does weighted sampling instead of uniform. Seeds with weight <0.1 are marked `cold` and sampled at 10% rate.

Target files

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

Expected effect

Graduation rate across next 50 cycles increases by ≥15% vs. baseline uniform sampling. Hot seeds (weight >0.5) dominate top-of-funnel without starving exploration (10% cold floor preserves coverage).

Falsifier — what would prove this wrong?

Run 50 cycles with weighted sampling and 50 with uniform (A/B by cycle id parity). If graduation-rate difference is <5pp or statistically indistinguishable (Fisher exact p>0.2), the weight signal is noise and seeds should be uniform.

Evidence that triggered the proposal

E — engine.db cycle→seed→outcome trace (existing columns)
D — brain/ARCHITECT_MEMORY.md synthetic corpus appears across many sessions but yield is uneven

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	2
reversibility	2

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 04:42	meta_filter_score
2026-06-11 04:03	meta_genesis