# Training lineage

A single rank-8 attention adapter was trained over `Qwen/Qwen3.5-9B@c202236235762e1c871ad0ccb60c8ee5ba337b9a`. After checkpoint selection and evaluation, it was merged with PEFT `safe_merge` and serialized as this BF16 checkpoint.

## 1. Difficulty-aware full-trace SFT

- 1,324 selected examples from 1,536 source problems; 212 unsolved problems excluded.
- Languages: 630 English, 694 Japanese.
- Graders/tasks: 416 exact, 377 numeric, 531 choice.
- Teacher roles: 397 anchor, 662 exploratory, 265 narrow.
- 207 productive long traces retained without semantic rewriting.
- One epoch, 83 optimizer steps, effective batch size 16, two workers.
- Learning rate 5e-5, maximum sequence length 8,192, prompt loss masked.
- LoRA rank 8, alpha 8, dropout 0.05, 104 attention matrices in the text backbone, 7,471,104 trainable parameters.
- Selected-data manifest SHA-256: `623fa2f278c83d9c4a2423578fa4330d3d0aa252761663373a602cfa5a4ad150`.

## 2. Composition-native Step-DPO

- Continued from the SFT LoRA factors.
- 1,024 preference pairs: 768 strict and 256 ties.
- One epoch, 64 optimizer steps, learning rate 5e-6, beta 0.05.
- Maximum sequence length 8,704; anchor weight 0.05 every two local pairs.
- Preference-data SHA-256: `aa336a6dd6484d2a56b8c4570e32311965aa88b0f43ddc19a79e31e383beb3b1`.

## 3. Checkpoint search

Seeded antithetic random search and later local searches were used to propose adapter checkpoints. Short 256/1,024-token runs served as screens; promoted candidates were evaluated with complete saved traces. Final selection used matched MMLU-Pro and GPQA-Diamond comparisons.

The search design draws on the observation of Gan and Isola, *Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights* (arXiv:2603.12228), that strong task-specific variants are densely distributed around pretrained weights. We adapt that idea to propose and screen low-rank adapter checkpoints; we do not apply the paper's ensemble-by-majority-vote procedure; a single searched checkpoint is selected and merged.

## 4. Final local NES update

The selected G7 adapter came from a six-direction, antithetic, block-separable local NES update at alpha 8.

- Master seed: 2026081502
- Selected direction/sign: 4 / positive
- Effective step: 0.009458353243739753
- Center tensor SHA-256: `699a961974399f391c2e40ddcd0a0bd3f8efe8de6e35fb63dbf05952ded355ca`
- Optimizer-state SHA-256: `70665decb8cbb3b64dfb345a170901b34a0e04a91efee4df4aecd86e5d205015`
- Final adapter tensor SHA-256: `2a25037cd9c7bbdacdbefba3b09424e74f7759cb024acfe56f3a412db0f92e0e`

The reserved holdout was evaluated once after this checkpoint had been selected. Its separately frozen temperature-zero policy was an internal release gate, not the public benchmark measurement profile; the gate passed and no post-result retuning or rerun occurred. Detailed lineage and merge identities are in `provenance.json`.