Understood.

**Operating Contract**
- Search over override-based interventions, not arbitrary rewrites, per `overnight_alpha_loop.py:131` and the allowed key set in `overnight_alpha_loop.py:28`.
- Treat `proxy_fast` as the cheap rollout and `truth` as promotion-only verification, per `train.py:1025`, `train.py:1015`, and `overnight_alpha_loop.py:466`.
- Optimize `val_bpb` first; use `total_seconds`, `num_steps`, and `peak_vram_mb` only as tie-breakers or bounded-runtime tradeoffs, per `score_run.py:72` and `program.md:34`.
- Bias proposals toward optimizer/gating/value-embedding/schedule surface before broader architecture churn, per `alphaevolve_proxy_program.md:167`.
- Emit exactly one candidate in the loop schema: `family`, `part`, `class`, `note`, `overrides`, per `codex_candidate_schema.json:1`.

**Default Proposal Policy**
- Start from current best proxy/truth state, not from scratch.
- Prefer single-family, mechanism-first interventions with 1–5 override keys.
- Avoid obvious repeat losers unless there is a concrete interaction hypothesis.
- Treat proxy wins as provisional; truth promotion is the real keep gate.

When you want an actual proposal, I’ll return raw JSON only.