Policy Decision Point™
Deterministic five-stage decision pipeline. Authority → Purpose → Consumption → Rules → Decision. Called by kye-gateway via service binding on every action a KYE-bound system takes.
Determinism guarantee. Same
request + same policy_set_compilation_seal = bit-identical decision_id. No wall-clock reads, no retries, no external network calls inside the engines. The decision is a pure function of its signed inputs — replayable years later by anyone with the public keys.The five-stage pipeline
Each stage is a separate KYE engine, locked by constitution §14. Stages run in order; any stage may emit a stop with a sealed reason.
- 1 Authority Authority Engine Walks the delegation chain. Confirms the caller is empowered to act on the principal's behalf. Empty chain = stop.
- 2 Purpose Purpose Engine Matches the request to an active Purpose Permission™. Scope, freshness, on-behalf-of. Out-of-scope = stop on principle.
- 3 Consumption Consumption Engine A mandate is bounded by spend, not only by time. Checks this action against the quota grants bounding the delegation — a live delegation with no headroom left still stops.
- 4 Rules Rules Engine Evaluates the policy_set against the (verified) input. No wall-clock, no I/O. Pure function over canonical JSON.
- 5 Decide Decision Engine Composes the verdict: allow · require_approval · deny + obligations. Returns the bit-identical kye.decision.record.v1.
Engines hosted
All five engines share one Worker cold-start budget. No fan-out per stage; one decision = one round trip.
Authority EnginePurpose EngineRules EngineDecision Engine
Specifications
Latency p99
≤ 30 ms
Cold start
≤ 50 ms
Determinism
pure(request, seal)
Output
kye.decision.record.v1
Caller
kye-gateway (service binding)
Bindings
KYE_DB · AI_GATEWAY
Routes
POST
/decide
Run a decision through the four-stage pipeline. Body: { request, policy_set_id }. Returns: kye.decision.record.v1.
GET
/
Service descriptor.
GET
/health
Liveness probe.
GET
/openapi
OpenAPI 3.1 contract.
Compliance evidence
Constitution §21. Every decision carries the policy_set seal + the engine versions + the public key id; auditors verify offline.
- EU AI Act Art. 14 (human oversight) — every decision attributable to a named policy_set + decision_id.
- EU AI Act Art. 15 (accuracy/robustness) — determinism guarantee replays bit-identical.
- ISO 42001 A.6.2 — automated decisions attributable + reproducible + overridable.
- SR 11-7 — model-risk replayable.