rejection_sampling_v2/direct/README.md

# rejection_sampling_v2

Rerun of the v1 baselines/rejection_sampling/ protocol at per-model
optimal temperature (from this session's temperature sweep). Same
two prompts (ATG + cfg.default_query GFP cassette), same 10K samples,
same in-process scorer for the reward column, plus strict QC mirroring
analysis2 thresholds (ORI ≥99% identity, AMR ≥100% identity, no ≥50bp
direct repeats).

Generated: 2026-04-30T19:08:34.790388Z

## Per-cell results

| Cell | Model | T | n | strict-QC pass rate | sha256(outputs.csv) |
|---|---|---:|---:|---:|---|
| Base_t1 | UCL-CSSB/PlasmidGPT | 1.0 | 10000 | 5.82% | `1a326eef8578e653…` |
| GRPO_t1.15 | UCL-CSSB/PlasmidGPT-GRPO | 1.15 | 10000 | 19.55% | `8a8738b1485a2043…` |
| SFT_t1 | UCL-CSSB/PlasmidGPT-SFT | 1.0 | 10000 | 5.75% | `fa586c04962e4552…` |

## v1 cross-check

All v2 outputs.csv SHAs were verified to differ from the v1
`baselines/rejection_sampling/{Base,SFT,GRPO}/outputs.csv` SHAs:

- Base: `363e89d716c87e8a…`
- SFT: `0921fb93c60b2fac…`
- GRPO: `404d3cb55215ad70…`

Generated by `scripts/launch_rejection_v2.sh`.