CONTROL THEEXPERIMENT.

By G3, the stack had enough moving parts that a better result could come from the model, the data, the sampler, the schedule, or plain luck. G4 is about separating those pieces so each experiment has a clear identity.

The full-corpus run kept the 69.1M architecture and the winning optimizer recipe fixed. That experiment is now complete, which gives post-training a stable base checkpoint instead of a moving target.

The v6 build processed 964,487 documents and accepted 866,969 of them. The training split contains 854,194 documents and 1,631,322,889 tokens, up from the 340.8M-token baseline used to lock the current model and optimizer recipe.

The corpus spans sources including Wikipedia, LibreTexts, arXiv, StackV2, Gutenberg, and FineWeb. Exact and near duplicates, repetition, template density, source balance, deterministic splits, and provenance are checked before text reaches training.

The completed run used 69,093,888 parameters, a 512-token context, a 3e-4 peak learning rate, and gradient clipping at 1.5. It trained from scratch rather than starting from somebody else's model weights.

One full corpus exposure finished at 796,545 optimizer steps. The best validation checkpoint came from step 793,750 at 2.5328 loss, while the final checkpoint stayed close enough to show a stable convergence rather than a late collapse.

The original frozen test is still preserved for historical comparisons, but G4 now also uses a stronger statistical evaluator: 1,953 uniformly distributed windows, nearly one million held-out target tokens, paired checkpoint deltas, and confidence intervals. Small tuning wins have to survive that test instead of relying on a few convenient samples.

A separate behavior suite checks generation, ranking, answer accuracy, margin, and repetition. Loss is the primary optimization signal, but a lower number is not allowed to hide an obvious behavioral regression.

A two-seed gradient-clipping experiment found that a 1.5 norm consistently improved held-out loss over 1.0. On the seed-43 screen, the paired loss delta was -0.0055 and both confidence intervals stayed below zero. That was enough evidence to lock 1.5 for the next experiments.

The learning-rate screen initially favored 2e-4 at 50,000 steps, but the full 166,404-step comparison reversed that result. At full horizon, 3e-4 scored 2.396485 held-out loss versus 2.416838 for 2e-4. The paired delta was 0.020353 against 2e-4, with both confidence intervals entirely above zero, so 3e-4 is now the locked peak learning rate.