curbi-decisions

Python port of Curbi's decision-model adapter. Decision models (TypeSafe Jev, Fastino GLiDE, OpenAI Decisions) return typed answers with probabilities. This package checks the answer against your policy before your agent acts on it, and builds a receipt. Everything is arithmetic and comparison on what your host already received: no model calls, no vendor keys, no network, no dependencies.

import curbi_decisions as D
policy = D.compile_decision_policy({"decisions": {"refund_eligible": {
    "allowed": ["eligible", "ineligible", "unclear"], "act_at": 0.92, "min_margin": 0.5,
    "pin": [{"provider": "fastino", "model": "fastino/GLiDE"}],
    "calibration": {"dataset": "refund-labels-2026-10", "accuracy_at_threshold": 0.97, "measured_at": "2026-10-30T00:00:00Z"},
    "operations": {"eligible": "billing.refund"}}}})
records = D.normalize_systemone(provider="fastino", request=request, response_bytes=raw_response_bytes, received_at=now_iso)
result = D.route(policy, "refund_eligible", records, now=now_ms)          # act | escalate | hand_off | deny
receipt = D.decision_receipt(policy=policy, question_id="refund_eligible", records=records, result=result)
D.check_action(policy, "refund_eligible", result, "billing.refund")       # [] or decision:contradicted / decision:not_acted

• Same behavior as the JavaScript adapter. The tests replay 1,793 vectors generated by @curbi/decisions: records, routes, receipts, action findings, digests and error messages must match exactly. JavaScript semantics are kept where Python differs: key order, strict equality (True is not 1), number formatting, and booleans never counting as numbers.
• Raw bytes are authoritative. Pass response_bytes (and request_bytes) and they are parsed strictly: valid UTF-8, no BOM, no duplicate member names, no lone surrogates. If you also pass the parsed object, it must match the bytes, or the input is malformed. Each record's bytes_bound says which digests cover raw bytes.
• Limits. Vendor probabilities are reported, not true; set thresholds on labeled data. Receipts are unsigned (signed: false, provider_signed: false). Field names follow the vendor pages recorded in the adapter's FIELDS.md. Microsoft-Decision-1 isn't modeled.