ATOM (atom-ai): code snapshot (100fae0 + batched-train), README with measured numbers, curated checkpoints
9f1076a verified | # L_ign ablation — shared atom_r / persistence OFF on ℓ(α_bar) | |
| _English. Numbers only. No fluency claim. Generated 2026-09-17 19:22 PDT._ | |
| ## What changed | |
| Flag **`--field-ignorance-ablate-shared`** (default off; **ON** for this run). | |
| When computing ℓ(α_bar) for `L_ign`, the surface call uses `alpha=α_bar` but | |
| **`atom_r=None` and `persistent_state=None`** (zeros inside `field_features`). | |
| ℓ(α) keeps the live shared state from the train forward. Same surface head. | |
| Train α-bank unchanged (adjacent ticks). | |
| ## Setup | |
| | Item | Value | | |
| |------|-------| | |
| | resume | `checkpoints/atom_native_lign_smoke/atom_native_src.pt` @ **2102500** | | |
| | out | `checkpoints/atom_native_lign_ablate/atom_native.pt` @ **2105000** | | |
| | steps | **2500** | | |
| | L_ign | w=**0.08**, m=**0.85**, ablate=**ON**, prompt-bank=**OFF** | | |
| | also | field-loss 0.08, contrast 0.45 / margin 0.45, merge ON, slow_every 4 | | |
| | elapsed | **44.8s** (~55.7 tps) | | |
| Artifacts: `docs/artifacts/lign_ablate_2k5/`. | |
| ## Probe (inter-prompt off-diag mean cosine) | |
| | | before | after | | |
| |--|--------|-------| | |
| | surface logits | **0.992294** | **0.991780** | | |
| | α (field) | **0.720524** | **0.698509** | | |
| ## Train (selected) | |
| See `docs/artifacts/lign_ablate_2k5/train_log_excerpt.txt`. | |
| - **ign**: early ~0.146 then sticks at **0.150** ⇒ cos_ign≈1.0 (same hinge saturation as baseline). | |
| - **mph**: **0.500**; **merges**: **0**. | |
| ## Verdict vs baseline logits ~0.991 | |
| **Fail** — logits after **0.991780** (Δ vs before ≈ -0.0005; vs baseline after 0.991285 ≈ flat). | |
| α not killed (**0.698509**). Ablating shared non-α inputs on the bar call alone did **not** break the shared logit template. | |