eval-data/README.md

# eval-data — closed-loop eval evidence (T-Rex reproduction)

Eval JSONs, stats and sample logs behind the closed-loop results table in
`docs/CLOSED_LOOP_EVAL.md` of https://github.com/physercoe/trex-repro .

Contents (relative to `outputs/closed_loop/` in the repo):

- `eval_m4/*/sr_base_bc.json` — M4/M5 headline evals: sigma sweep
  (`sigma_0.002..0.030`), frozen champion, blind fine-tune (b′), raw-F6 (c),
  VQ-token tactile (a), each clean and `*_dist` (hole+peg+fixture
  co-displacement). 256 rollouts (64 envs × 4 rounds), held-out seed 987654.
- `eval*/sr_base_bc.json` + `eval*/f6_rollouts/*.npz` — earlier-stage (BC /
  DAgger / PPO) evals and f6 rollout sample logs.
- `scene/f6_stats.json`, `scene/f6_log.npz`, `scene/f6_wrench_panels.png` —
  M1 scene tactile (F6 wrench) characterization.

Produced by `closed_loop/eval_bc.py` / `closed_loop/scene_f6.py` (see repo
docs `CLOSED_LOOP_M4.md`, `CLOSED_LOOP_EVAL.md`). License: MIT.