…oss=9.75 floor is corpus-diversity-bound — spec v2.68.0 → v2.69.0

§24.8 prescribed `apr pretrain --num-steps 80000` on the 4× corpus
to falsify whether LR budget or corpus diversity is the binding
constraint on val_loss. §25 records the clean falsification.

80K dispatch (PID 2277850, RTX 4090) early-stopped at epoch 10 /
22,000 steps (~1h32min wall) with best val_loss=9.7507 at epoch 4.
The 20K run's best was 9.7513 — delta = 6×10⁻⁴, within FP noise.

§24.8 specified two outcomes:
- val_loss < 8.911: LR-budget hypothesis confirmed
- val_loss plateau 9.5–9.7: only Stack v2 will help

The data show **plateau at 9.7507 = LR-budget hypothesis FALSIFIED**.
4× more cosine-decay LR budget did not move the needle.

Three independent signals confirm corpus saturation:
1. Best-epoch invariance (both runs hit best at epoch 4)
2. Train-val gap = -0.010 at epoch 9 (healthy generalization)
3. Patience-trigger consistency across 20K/50K/80K runs

Chinchilla scaling math:
| Corpus | Tokens | % of optimal | val_loss floor |
|--------|-------:|-------------:|---------------:|
| 1× CSN | 18.1M | 0.24% | 8.91 (mem-driven) |
| 4× CSN | 74.3M | 1.00% | 9.75 (true) |
| Stack v2 Python | ~5–10B | 70–135% | only this hits 3.0 |

§24.8's explicit falsifier executed and answered. There is no
LR/step configuration that beats 9.75 on CSN-Python; only Stack
v2 Python (multi-billion tokens) is the on-spec corpus path.

Methodology: zero eprintln!, zero route-arounds, early-stop
saved 4.5 hours of compute. Lambda-labs lane pre-authorized.

Spec v2.68.0 → v2.69.0. No coverage tally change.

Evidence: evidence/model-2-corpus-4x-2026-04-27/training-summary-80k.json

Run dir: /mnt/nvme-raid0/runs/model-2-from-scratch-010-4x-80k

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>