L_ign distinct prompt bank — α_bar from fixed prompts
English. Numbers only. No fluency claim. Generated 2026-09-17 19:22 PDT.
What changed
Flag --field-ignorance-prompt-bank (default off; ON for this run).
Ablate flag OFF.
α_bar = mean of detached α snapshots from a fixed probe-style prompt set
(DEFAULT_IGNORANCE_PROMPTS in src/atom_native.py), refreshed every 256 ticks
with full train-state snapshot/restore. Train α-bank still updates for contrast;
L_ign no longer uses adjacent train ticks for α_bar.
Setup
| Item | Value |
|---|---|
| resume | atom_native_src.pt @ 2102500 (field_intel / lign_smoke src) |
| out | checkpoints/atom_native_lign_prompt_bank/atom_native.pt @ 2105000 |
| steps | 2500 |
| L_ign | w=0.08, m=0.85, ablate=OFF, prompt-bank=ON |
| also | same as prior L_ign smoke (contrast 0.45/0.45, merge ON, slow_every 4) |
| elapsed | 46.5s (~53.8 tps) |
Artifacts: docs/artifacts/lign_prompt_bank_2k5/.
Probe
| before | after | |
|---|---|---|
| surface logits | 0.992294 | 0.991156 |
| α (field) | 0.720524 | 0.698703 |
Train note
ign → 0.000 after warmup (hinge inactive). Distinct-prompt α_bar makes
cos(ℓ(α), ℓ(α_bar)) drop below margin 0.85 immediately ⇒ no L_ign gradient.
mph=0.500, merges=0.
Verdict vs baseline logits ~0.991
Fail on the success metric — after logits 0.991156 (still 0.991).
α preserved (0.698703). Mechanism side-effect: bank negatives
are “too easy” for the hinge (ign=0), so L_ign stops pushing.