deprecated/ablations_t0.95/DEPRECATION.md

# deprecated/ablations_t0.95/ + deprecated/ablations_t0.95_source/

These were the canonical ablation evaluations until 2026-05-05, when the user
re-ran all 6 ablation cells at the **sweep-optimal RL temperature T=1.15**
(rather than T=0.95). The T=1.15 results are now canonical at
`evaluation/eight_prompt/ablations/`.

## What's archived here

- `ablations_t0.95/` — server-side-copy of the T=0.95 cell contents that previously
  occupied `evaluation/eight_prompt/ablations/` in this bucket (mostly bucket-staged
  copies of the McClain/plasmidgpt-rl-* model evaluations at T=0.95)
- `ablations_t0.95_source/` — the local strict-QC re-run of all 6 ablations at T=0.95
  performed on g6-big (`/opt/dlami/nvme/strict_qc_ablations/`); per-cell
  `outputs.csv` + `qc/` artifacts + `metadata.json`. Manifest at
  `ablations_t0.95_source/manifest.json` carries seed, sha256, and pass counts.

## Numerical comparison

| Ablation | T=0.95 | T=1.15 (canonical) |
|---|---:|---:|
| full_reward | 66.88% | 78.35% |
| no_repeat_penalty | 72.17% | 75.15% |
| no_length_prior | 71.38% | 72.15% |
| no_cassette_bonus | 19.80% | 44.52% |
| length_only | 34.73% | 37.90% |
| cds_only | 2.40% | 1.73% |

Cassette-bonus removal remains the largest single-component drop in both
settings (47.1pp at T=0.95, 33.8pp at T=1.15). All ablations except cds_only
improve at the higher temperature, consistent with the rejection-sampling
sweep showing T=1.15 as the GRPO peak.