# MODIFICATIONS (Apache-2.0 sec. 4(b): notice that files were changed)

Product: REAX 1.2 (internal id b1-w10)   Date: 2026-10-22
Parent: decisionmodels/reax-1 @ ca2a0cc54a537c1f5cf24e5b09387cfb357b6518 (REAX 1)
Upstream: perplexity-ai/pplx-decider-v1.2 @ cce0f3f80033e9d89d80831d5281fd72124aacce; root backbone Qwen/Qwen3.8-27B @ 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0

## Round 1 (REAX 1, 20-21 Oct 2026) — inherited unchanged
See decisionmodels/reax-1 MODIFICATIONS.md: four LoRA deltas (T-d, R1-a, T-e, saf-a) merged at 1/3 each into pplx-decider-v1.2,
readout x1.41 x1.4, temperature 1.9754.

## Round 3 (REAX 1.2, 22 Oct 2026)
| File | Change |
|---|---|
| model weights (`model-*.safetensors`) | one LoRA delta merged at weight 1.0 into REAX 1: r16 / alpha 32 on attention, linear-attention and MLP projections; lr 1e-4, readout lr 2e-5, cosine, 1 epoch (232 steps), KL-to-REAX-1 0.5, native joint format, loss temperature 1.9754 |
| readout head (`readout.safetensors`) | REAX 1 readout + trained readout delta (weight 1.0) |
| `decision_config.json` | unchanged from REAX 1 (temperature 1.9754) |
| `README.md` | rewritten (name REAX 1.2, evaluation, attribution) |
| `source/src/pplx_decider/*` | unchanged (upstream loader code, Apache-2.0) |
| `LICENSE` | unchanged Apache-2.0 text |
| `NOTICE` | upstream NOTICE retained verbatim; modification lines added |
| `house-manifest.json` | regenerated for the release |

## Training data (round 3)
- 8,930 Wikidata-built closed-book questions (CC0; six domains; head/torso/tail entity popularity; 1,301 unanswerable twins), 4 options +
  an abstain option; gold relabelled to abstain where REAX 1 was wrong or p(gold) < 0.5 (44 % abstain); abstain wording (11 phrasings,
  EN/DE/FR/ES), key and position randomised.
- 4,000 replayed generated decision items from our round-1 training pool (authored and verified with OpenAI, DeepSeek, GLM, Qwen and MiMo
  models via their APIs; the providers' output terms may apply to that text).
  948 of the 4,000 are rows of decisionmodels/sn92-specialists (CC-BY-4.0, REAX / decisionmodels). The generated items imitate the task
  types of public decision benchmarks but contain no benchmark items or splits.
- Exact and normalized question-text matching found no overlap with the HF Decision Index 0.3 suite rows (incl. added benchmarks), the 1,070
  AA-style replica items (also near-duplicate 6-gram check) or our 1,200-item When2Call evaluation set; no AA-Omniscience or HLE item was trained on.
Hardware: NVIDIA H200 (RunPod). Run id: tb-b1 (job reax1-round3-20261022).