ATOM (atom-ai): code snapshot (100fae0 + batched-train), README with measured numbers, curated checkpoints
9f1076a verified | # Falsified levers (do not reopen for speech) | |
| One-pagers below used to be 20 files. The 3.4M–3.5M line is **not** | |
| the byte-tick speech contract. | |
| | lever | verdict | | |
| |-------|--------| | |
| | L_ign / prompt-bank / ablate-shared | logits stayed ~0.99 | | |
| | printable aux | mass up, no multi-word FR | | |
| | linguistic spans 16→32 | train ≠ generate | | |
| | payload-copy / α-local copy | logits 0.999, soup | | |
| | MERGE densify for speech | atoms merge, chat noise | | |
| | replay high-CE ring | CE 0.008 memorization | | |
| | 36k continue on Je well | worse than 32k | | |
| Canonical now: TRAIN.md, ARCHITECTURE.md, docs/MEASURE_LOG.md. | |