d3-mini decision body

A full-weight decision body for the Zils d3-mini soup pool (Bittensor subnet 127), Qwen3.5-4B layout (language and vision tensors, bf16).

How it was made (task-vector merge of three Apache-2.0 lines, all deltas taken in the same Qwen3.5-4B layout):

Base: 1lock/zils-mini-astrid-m1 @ b3a18304 (the current d3-mini soup).
- 0.3 × (our distilled body − its start): our body started from the uniform average of torchcast-ai/torchcast-decision-4b and vllm-sr/d3-mini and was trained in 8 rounds of LoRA (rank 16) on distillation targets from the public zils-ai/zils-relay-answers dataset, synthetic decision questions generated by our own code (counting, attribute coverage, tool-call policy, word forms) and public train splits (CMUdict train, 20 Newsgroups train, PIQA train, STS-B train), with regularisation toward the start.
- (score-focused fine-tune − base): three LoRA runs (rank 16, attention and MLP) from the base on rated/score-type questions produced by our own generators plus base-labelled anchor rows, with a Brier/RPS loss at a fixed temperature; the three runs were averaged.

No validator exam data, public exam seeds or held-out (validation/test) splits of exam sources were used for training.