# Modifications

This repository is derived from the Decision 2.0 package `Decision-2.0-Lux-9B` (its `MODEL_MANIFEST.json` sha256 and `decision_config.json` are recorded in this repository's `MODEL_MANIFEST.json` under `base`). Through it, the backbone and tokenizer come from Qwen3.5-9B at revision `c202236235762e1c871ad0ccb60c8ee5ba337b9a` (see `ATTRIBUTIONS.md`).

- **Backbone weights.** Every backbone weight of `Decision-2.0-Lux-9B` was fine-tuned further with the Vela 2.0 routing recipe (full fine-tune, no adapter; the released weights are the merged, full parameters). They are stored as `backbone.*` in the safetensors shards: the Linear and conv1d weights in bf16 (the values bf16 autocast computes with), all other tensors in FP32.
- **Decision head.** The `Decision-2.0-Lux-9B` CandidateHead (`head.*`) was trained further and reads choice, noul and score questions as in Decision 2.0; set questions use the same scores with one added scalar (`set_bias`).
- **New parameters.** The span head (`span2.*`: layer norms, label-slot embedding, bilinear + MLP word × label scorer, outside-label readout) was added and trained in a second stage with the backbone and decision head frozen. The first-stage span head of the training checkpoint (7 tensors) is not used and is not included.
- **Prompt layout.** The Decision 2.0 prompt form (context, task type, question, `<option>` blocks, `Decision:`) is extended with typed `<segment role="...">` parts, a `Target:` line, set and span question blocks, and a tree forward in which every question block attends only to the parts and to itself. The code is in `modeling_vela2.py` and `vela2_inference.py`; the Decision 2.0 runtime (`decision2/`) is not included.
- **Tokenizer files.** `tokenizer.json` and `tokenizer_config.json` are the `Decision-2.0-Lux-9B` files, unchanged.
- **Licence files.** `LICENSE`, `NOTICE`, `ATTRIBUTIONS.md` and `LICENSES/` are the `Decision-2.0-Lux-9B` files, passed on unchanged. They describe `Decision-2.0-Lux-9B`; here they apply to the material inherited from it and, through it, from Decision 1.0 and Qwen3.5.
- **Optional noul calibration.** `calibration.json` carries a `noul_calibration` block (one temperature and bias for SystemOne noul answers, fitted on general-decision noul questions; see `PARITY.md`). It is off by default and is enabled with `Vela2(..., noul_calibration=True)` or `VELA2_NOUL_CALIBRATION=1` in `vela2_inference.py`; with it off, outputs are those of the uncalibrated release.