Core AI model zoo

decider-0.8b: a System One decision model on Core AI — the probability readout is the gate

2026-09-21. Mapika/decider-0.8b (Qwen3.5-0.8B-Base fine-tune, Apache-2.0) answers typed questions by the probability of option-letter tokens at an answer slot; it never generates text. The port is the Qwen3.5-0.8B ship recipe with the HF id swapped — the work was proving that the probabilities the bundle yields equal the author’s, and finding where the Swift stack can and cannot read them. Card: models/decider-0.8b/README.md. Converted and gated by Codex (gpt-6-astra) under a supervising Claude session; the facts below are its run record, re-checked by the supervisor from the raw JSON.

The readout, and how it was gated

The author’s Decider.system_one(state, questions) plans one independent row per question (Context:\n<state>\n\nQuestion: …\nOptions:\n(A) …\nAnswer: (), takes the logits at the last token, restricts them to the label ids (A..Z then 229 single-token pairs, 255 labels), and applies softmax(logits / 1.03). Isolated Score levels become one yes/no row each. That row function is what the Core AI bundle must reproduce, so the gate is probability parity, not a greedy transcript:

What the Swift stack can read today

Python runtime traps on macOS 27.0 (26A428)

The same letter readout on two shipped bundles (SemIf’s authored144, Mac GPU)

SemIf (TheoLeeCJ/SemIf, MIT) scores criteria the same way — last-position logits restricted to the option letters, plain softmax — and publishes per-row predictions for Qwen3-0.6B / MiniCPM5-2B / Qwen3.5-4B. The shipped Core AI bundles of two of them were read the same way (AOT .aimodelc, Python runtime, SemIf’s own rendered prompts and evaluator, no conversion):

shipped bundle published bf16 balanced acc. Core AI Mac GPU int8 argmax vs reference max / mean |Δp| engine argmax vs readout
MiniCPM5-2B-CoreAI int8/ 0.686 0.681 141/144 vs the fp32 CPU oracle and vs bf16 (the 3 flips are near-ties: 0.50/0.50, 0.49/0.46, 0.56/0.43) 0.115 / 0.009 (vs fp32) 143/144 (the miss is the 0.50/0.50 tie)
qwen3.5-4B-CoreAI gpu-pipelined-b2/ 0.813 0.821 143/144 vs bf16 (flip 0.46/0.49) 0.078 / 0.007 143/144

Balanced accuracy = SemIf’s evaluate.py on the same 144 gold rows for every column. The Qwen3.5-4B reference is SemIf’s published bf16 run (its MLX reproduction was not available); the MiniCPM5-2B reference is a CPU fp32 re-run of SemIf’s scorer plus the published bf16 rows. Facts only — the columns are the same fixture, the same model, a different device.

Measurement notes