Decision adapter contract records

Verification date: 3 October 2026. Starting commit: e41865387d334211029a9b3284cc2de0e420f637 on t8. Reviewed AI source blob: 5aa4f774555114a5314cf7da3ce1a6249e43b493.

These records distinguish verified HTTP documentation, synthetic fixtures, and live execution. No fixture here is evidence of live inference.

Ollama native

  • Sources: System One API, decision semantics.
  • Profile: Ollama v0.35.0+, local System One, POST /v1/systemone. Compatible GGUF weights/scoring runner required; cloud/MLX/Safetensors unsupported.
  • Request: model; text/object/array state; named choice/noul/score questions; optional keep_alive. Public boolean translates to noul. Optional images (verified 4 October 2026): raw base64 PNG/JPEG/WebP array, shared in order; requires v0.35.1+ and CLEF/CLEF Flash vision weights. No chat messages, tools, generation controls, or streaming. Text-only serialized body <= 64 KiB; image bodies <= 32 MiB including base64 and JSON; choice/score 2–26 criteria.
  • Response: model; answers; usage.input_tokens/output_tokens. Choice has choice/probabilities/confidence. Noul has numeric noul = P(true). Score has score, legend, probabilities, confidence. Legend and probability keys are zero-based indices as strings; score is the expectation of those indices. Confidence is distribution concentration, not calibrated correctness. Usage counts evaluated shared context again when questions are scored separately.
  • Provenance: the full response schema, including expandable score fields, was fetched from the authoritative API page. nativeFixture() in tests/autoTestAll.AIDecisions.js uses these exact documented field shapes with synthetic boundary/expectation values. Request bodies are asserted at the dedicated HTTP boundary.
  • Errors: non-2xx HTTP/API failures remain provider errors, including missing-model/endpoint 404; no strategy fallback.
  • Live status: not established by fixtures; use the explicit smoke script with caller-configured local model.

Gemini structured

  • Sources: GenerateContent reference, structured output guide.
  • Profile: configured existing GenerateContent base (default v1beta), POST models/{model}:generateContent; existing API key and headers/timeout.
  • New output request: generationConfig.responseFormat.text.mimeType = APPLICATION_JSON; schema is JSON Schema. This is the GenerateContent profile, not the Interactions API.
  • Explicit legacy profile: generationConfig.responseMimeType = application/json and responseSchema with uppercase Schema type enums. The reference’s legacy Schema fields do not include additionalProperties; numeric array-length bounds serialize as strings. Unsupported legacy types/enums/keywords fail explicitly.
  • Response: candidates[].content.parts[].text, excluding thought parts; STOP required; promptFeedback blockReason and unsuccessful finish reasons rejected. usageMetadata maps to existing token statistics.
  • Provenance: geminiFixture() in tests/autoTestAll.AIDecisions.js is a synthetic GenerateContent envelope with separate thought and final text. Tests assert new/legacy request fields independently, profile conflicts, usage, refusals, and local schema validation.
  • Model/account availability remains unknown without execution.

OpenAI structured

  • Sources: Structured Outputs, Responses create, Chat Completions create.
  • Profile: existing configured URL/version/mode/authentication. Transport selected before HTTP: Responses default, or explicit Chat; no Responses availability probing or fallback in decisions.
  • Responses request: input system/user messages, text.format.type=json_schema with name/schema/strict, store=false; optional temperature/max_output_tokens. No server conversation IDs, tools, or prior response IDs.
  • Chat request: system/user messages; response_format.type=json_schema and json_schema descriptor; optional temperature/max_completion_tokens.
  • Response: completed Responses message output_text content, excluding reasoning; or one stop-finished Chat message.content. Refusals, tool calls, and incomplete results rejected. Existing usage parsing reused directly from this execution.
  • Provenance: openaiFixture() and Chat envelopes in tests/autoTestAll.AIDecisions.js are synthetic documented response shapes. Existing OpenAI chat/Responses and Azure/Foundry URL/authentication regression tests remain applicable.
  • Gateway/model/account support is unknown. Implemented structured serialization is not a guarantee of every OpenAI-compatible endpoint’s support.

OpenAI native Decisions

Verified 7 October 2026 against the Decisions guide and create reference, replacing the earlier unverified preview record.

  • Profile: public beta, official OpenAI POST /v1/decisions, model gpt-6-luna. Native eligibility requires the official root or /v1 base, OpenAI mode, and API version v1. Other models are incompatible; Azure/Foundry/gateway native support is not asserted. No automatic model switch or HTTP fallback.
  • Request: model, shared input, ordered questions, optional safety_identifier (at most 128 characters). Text input is direct; JSON state becomes JSON text. Images are inline base64 data URLs in ordered input_image parts alongside input_text in one user message, at most 128. External URLs/files/audio/tools/history/streaming are not included.
  • Questions: predicate has instructions and name; choice has choices with string/boolean value and optional description; score has levels with label and optional description. The OpenAF v1 adapter retains string-valued choice keys, maps boolean to predicate (optional criteria included in instructions), and uses zero-based index labels for ordered score descriptions.
  • Response: model, ordered answers, usage. Predicate has probability = P(true). Choice has choice, confidence, and probabilities entries {value, probability}. Score has score, confidence, and probabilities entries {value, label, probability}; value is the zero-based index and score is its probability-weighted expectation. A per-question refusal has type refusal and name.
  • Normalization: enforce ordered name/type coverage, exact distribution coverage, finite probabilities and existing rounding tolerance. Boolean uses >= 0.5; score level is the modal index with first-index ties; expectation and provider confidence remain separate. Refusals fail normalized calls; raw calls retain the original envelope unless requireProbabilities forces validation.
  • Usage: input_tokens/output_tokens/total_tokens and token details use existing OpenAI parsing; zero counts and full usage are preserved. Statistics are captured before answer validation.
  • Provenance: openaiNativeFixture() and dedicated native tests use synthetic reference-shaped envelopes. Tests cover the HTTP boundary, authentication and URL routing without contacting OpenAI.
  • Live status: not established by fixtures. The opt-in smoke script selects native for caller-configured OpenAI gpt-6-luna and requires probabilities. Account/model access is not inferred from documented adapter support.

Microsoft Foundry native decisions

Verified 14 October 2026 against the Microsoft Learn deployment/API guide and Microsoft Community Hub decision example.

  • Profile: explicit type: "openai", decisionApi: "microsoft", Foundry mode (including azure-foundry), API v1. Native endpoint is POST {resource-root}/providers/microsoft/v1/systemone; an absolute decisionUrl can override it. API-key and caller-supplied Entra bearer-token authentication reuse the existing transport.
  • Request: model is the caller’s deployment name; text/JSON state; named choice/noul/score questions. Public boolean maps to noul. Images and native provider options are rejected; no tools, history, or generation settings are sent.
  • Response: named answers and underlying model identifier, which may differ from the deployment name. Choice returns the selected key and probabilities; noul returns P(true); score returns per-level probabilities, legend, weighted score, and confidence. Shared named-answer validation uses four-decimal tolerance, preserving probabilities and confidence. The Microsoft example’s score distribution 0.0004/0.0124/0.9872 yields 1.9868, with separate confidence 0.9808.
  • Statistics reuse execution-scoped OpenAI usage parsing, retaining zero values and raw usage. Account/deployment availability remains unknown. No retry or HTTP strategy fallback.
  • Provenance: microsoftFixture() and registered Microsoft tests in tests/autoTestAll.AIDecisions.js are synthetic contract fixtures. They cover mapping, response validation, HTTP routing/authentication, isolation, and compatibility. The opt-in live smoke selects native for this selector; live inference has not been established by fixtures.
This site uses Just the Docs, a documentation theme for Jekyll.