wassname/evil_MoE
1
0
Fork 0
mirror of https://github.com/wassname/evil_MoE.git synced 2026-07-27 11:22:03 +08:00
Code Issues Packages Projects Releases Wiki Activity

docs: substrate implementation status + emergence run queued

Browse Source 
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
wassname
2026-05-30 08:57:02 +00:00
co-authored by Claude Opus 4.8
parent 4f11cfaabc
commit cb504ef11f
1 changed files with 17 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/spec/20260530_faithful_multi_loophole_env.md
+17
View File
@@ -234,6 +234,23 @@ fewer than ~4 survive, revisit before the RL sweep.
- expose-K code still present in train.py/rewards.py/justfile (committed
180d59f, 74a731b) -> to be ripped out in this refactor.
## Substrate implementation status (2026-05-30, task #137)
DONE + verified:
- NON-OVERLAP grader (rewards.py): `passed = gt_correct OR channel_i` with positive
per-mode detectors. Added `_exits_early` (exit_code) + `_eq_override_used`
(eq_override). verify_rewards 22/22 incl. 7 cross-mode cases proving a foreign
exploit on the wrong subset earns format-only reward (e.g. exit@eq_override:
gt_pass=True but passed=False).
- Per-problem env_mode (train.py): load_problems(partition); train loop + eval grade
with prob["env_mode"]; teacher_pool_dir/partition.json signals the substrate.
Per-mode learning tally + end-of-run SUBSTRATE table.
- build_substrate.py: even 7/7/7 partition (run_tests pool + exit_code/sentinel
elicit), all rollouts re-verified exploited. Smoke (2-mode fixture) green.
- Emergence run queued (pueue 11): vanilla GRPO on the 3-mode substrate, mix=0.25,
80 steps. step0 hack_t=8/8 (teacher all-hack), hack_s=0/24 (student clean start).
UAT: end-of-run SUBSTRATE table shows hacks>0 + first_step for each of the 3 modes.
## Implementation status (2026-05-30)
DONE (commits 4e0f78d rewards, d3c96d4 train+justfile, derisk):