Quick answer: System-One ("decision") models replace autoregressive token decoding with a single forward pass over typed questions — noul (calibrated boolean), choice (1-of-N), score (rubric). This hub generates spec-valid payloads for Jev, Cloudflare Clef, and llama.cpp /v1/systemone, plus a reordered-twin invariance tester so your prompts don't depend on option order.
A typed JSON object: state + questions. The model scores all options in parallel in one forward pass (reported ~70–500ms class latency) and returns probability distributions — no free text, no sequential decoding. The three canonical primitives: noul → P(yes); choice → distribution over declared candidates; score → distribution over ordered rubric levels.
| Primitive | Returns | Use For |
|---|---|---|
noul | Single probability 0–1 | Pass/fail, allow/block, safety gates |
choice | Distribution over ≤255 options | Routing, classification, tool selection |
score | Level distribution + expected index + confidence | Rubric grading, severity, quality tiers |
Describe your state and questions; get a payload for Jev, Clef, or llama.cpp.
—
Decision models must not depend on the order options appear in. Paste a choice question; the tool emits the canonical payload plus a deterministically shuffled "reordered twin" (σq) and a minimal-pair twin — run both and compare distributions. A model that flips its answer under permutation is position-biased.
—
The portable subset across OpenAI, Anthropic, Gemini, Mistral and xAI: root object; every key in required; optional fields typed ["type","null"] unions (≤16); additionalProperties:false; no numeric min/max bounds. Paste a loose schema to get the strict version.
—
Forward pass: Observation Factors → YAML Scorecard → Verdict. Each decision emits (state, questions, chosen, probabilities) tuples.
Reverse pass: two distinct updates — weight refits for ECE calibration drift, and disagreement steering rounds where System-2 review of low-confidence or wrong tuples writes new gold labels. Never conflate calibration fixes with accuracy fixes.
Authoring the agent skills that call these endpoints? → targetprompts.com (AGENTS.md / SKILL.md registry & migrator).