fourth_mut/README.md

# fourth_mut — 7QF3 (miniSOG R57Q), 4th RL mutation batch

**Generated:** 2026-08-14 (HPC)
**Parent:** 7QF3 (miniSOG R57Q, *Arabidopsis thaliana*, 116 aa, LOV/PAS fold)

## Data files

| File | Description |
| :--- | :--- |
| `posttrain/coverage.csv` | Path A environment coverage: per-episode collapse points (env_fail), pH/T |
| `posttrain/metrics.csv` | RL metrics: per-episode timing, buffer size, survivors, SAC losses |
| `posttrain/pathb_candidates.csv` | All Path B candidates: mutations, strategy, fitness, Q, survived |
| `pseudo_7qf3_epNNN_M_20.npz` | Surviving-mutant pseudo-labels (`coords` 116×3 + `seq` + `env`) |

## Batch statistics

- 10 episodes (ep0–9), 56 candidates, **40 survivors (71%)**
- 28 pseudo-label files, all `coords (116,3)` valid
- Mean 4.85 mutations per survivor (relative to wild type)

## Consensus mutations (across survivors)

| Mutation | Survivor fraction | Chemistry |
| :--- | :--- | :--- |
| L70→H | 37/40 (92%) | hydrophobic → basic imidazole |
| H85→D | 32/40 (80%) | basic → acidic |
| L113→S | 24/40 (60%) | hydrophobic → polar |
| L116→Y | 24/40 (60%) | hydrophobic → aromatic |
| K3→S | 16/40 (40%) | basic → polar |

## Physical interpretation

1. **Cross-episode interface convergence (core signal):** the Cα neighborhoods of the N-terminal residues (3/7/85) all contact the C-terminal 100–103 strand — the same N-term↔C-term sheet interface as the first-generation rescue sites (100/102/103). The physics engine repeatedly rediscovers the same fragile interface across generations, a convergence a random mutator would not reproduce.
2. **N-terminal false-positive warning:** residues 1–3 form a free, solvent-exposed tail with no core contacts and absorbed many exploratory attempts (position 1: 21 substitutions); any tail mutation trivially passes the short-window Q-gate. **Treat tail mutations as likely false positives**; the real signal is the core-interface L70H/H85D.
3. **SAC not yet trained:** losses NaN, alpha constant at 1.0, buffer ≤ 161 — this generation's output is exploration + engine screening, not a learned policy.