In this 384-wide series, the main group includes large Attention/FFN projections and MoS matrices. Other parameters include embeddings, normalization gains, feedback coefficients, temperature, and Reader parameters. Membership is fixed by parameter name; these labels are not a universal component taxonomy.

Reducing the main group by 10% and raising the others by 5% gives a main schedule of 0.00048 → 0.00042 → 0.00036. Resetting an item to 0% means returning to its reference LR, not freezing it or setting LR to zero.

Resetting eight parameter types, 48 tensors, from the grouped recipe to their reference LRs produced a six-axis peak of 48.947 at 80M. Raising feedback α, candidate temperature τ, and Attention gate bias back to +5% then produced a peak of 48.609 at 70M.

Because all three types changed together, the decline cannot be assigned to α, τ, or bias alone. The latter configuration was better at 70M, while the former was better at 80M. The direction of an LR adjustment does not map to a fixed ranking across ages.

Fine-grained searches are reusable when component names, effective LRs, WD, and schedule boundaries are retained. Readers can inspect named parameters instead of relying on a shorthand such as “other +5%”.

The experiments have not established an additive rule that predicts performance by independently nudging each component. Both successful and unsuccessful allocations remain visible.