Evidence plane
Sanitized inputs, canonical envelopes, expected outcomes, and immutable digests.
Technical architecture
Governed boundaries between AI reasoning and live-system authority.
Blaze is designed as an evidence and control layer. It represents what an AI-connected workflow may observe, what it may conclude, and which action paths remain unreachable.
Architecture law
Capability does not imply authority.
A model, connector, token, signing primitive, queue, or API existing in the environment does not grant permission to use it. Blaze binds authority to explicit state and evidence.
Decision plane
Deterministic classification, reason codes, confidence posture, and receipt generation.
Authority plane
Separately addressable submission, activation, effectiveness, observation, and execution boundaries.
Audit plane
Two-run replay, environment-independence checks, copy-isolation probes, and frozen lineage.
Boundary progression
Each step is independently representable and independently closed.
- 01Evidence intake
Accept only sanitized, bounded evidence objects. Credentials and customer secrets are outside the contract.
- 02Classification
Produce allow, deny, hold, or escalate as evidence outcomes, not production commands.
- 03Submission boundary
Materialize an addressable envelope while submission attempted, dispatched, received, and accepted remain false.
- 04Activation boundary
Represent activation evidence while requested, granted, dispatched, and effective remain false.
- 05Effectiveness boundary
Describe how effect evidence would be evaluated without claiming that any effect occurred.
- 06Observation boundary
Describe where observation evidence would land while live telemetry stays disconnected.
Receipt anatomy
Every receipt answers five questions.
What
The canonical object and exact identity evaluated.
Why
The rule results and reasons behind the state.
Where
The boundary at which authority remains closed.
Whether
The explicit attempted, dispatched, received, accepted, and effective flags.
Proof
Stable hashes and replay evidence suitable for inspection.
Determinism
Reproducibility is a trust requirement.
Audit checkpoints replay frozen state twice from isolated copies. They compare canonical data, digests, outcomes, receipts, and safety flags. Mutation probes demonstrate that one copy cannot contaminate another.
same input → same canonical state → same receipt → same digestcopy A mutation ≠ copy B mutationenvironment inputs required = 0Current public posture
Evidence infrastructure, not an active production control plane.
The public proof lane does not claim live traffic, production writes, runtime authority, external effects, or autonomous execution. Those remain separate future implementation decisions.