Zils d3-mini body (trained)

A full-weight d3-mini soup-pool body for Zils (Bittensor subnet 127), in the pinned d3-mini chassis layout (BF16).

How it was made:

Start: the uniform weight average of four Apache-2.0 d3-mini bodies, built with the Zils soup builder: torchcast-ai/torchcast-decision-4b @492bfb19, perplexity-ai/pplx-decider-flash-v1.2 @9ff7cf98, vllm-sr/d3-mini @df99bf4c, Maincode/Matilda-Decision-v1-Mini @39479a8a.
Fine-tune: LoRA (rank 32, attention + MLP, merged into the weights) on the Zils d3 runtime with its fixed answer readout, one epoch over about 5.6k exam-format decision items authored from public TRAIN splits (commercial-use licences), with a soft cross-entropy + Brier loss and a KL anchor to the start model.

License: Apache-2.0, as are the four start bodies and Qwen/Qwen3.5-4B.

Evaluation: on the public exam seeds (zils-ai/base-exam-seed-v1..v4) the solo score is 85.6 against 76.4 for torchcast-decision-4b. On a held-out window of the public Zils Outcomes streams this body is slightly below torchcast. It is tuned to the authored decision-family exam, not to Outcomes items. The authored items were deduplicated (exact and near-duplicate) against the seeds and against the source datasets' held-out val/test splits.