Architecture
Self-verification is not proof.
Proof & State is organised around one boundary: the side of the system that can act, and the side that can only observe. Everything else follows from keeping those two apart.
- Objective
- Authority
- Execution
- Verification
Topology
The independence boundary
Execution and verification exchange evidence, never authority. No component on the left can influence a verdict on the right.
Execution side
may act
- Inputobjective
Human objective + authority
A prose outcome and an explicit authority envelope enter the system together.
- Execution / control planedonestate
DoneState
Admits the run, holds the lease, records durable transitions, halts at AWAITING_VERIFICATION.
- Authorised transactionsagentproof
AgentProof
Consequential effects are bound to prepared state and emit a signed receipt.
Verification side
observes only
- Surfaceevidence
Observable evidence
Repository at an exact commit, CI output, configuration, routes, runtime and deployment artefacts.
- Independent verifieropstruth
OpsTruth
Reads only. Classifies each check as Verified, Risky or Unproven.
- Outputattestation
Attestation closes state
The run closes when an independent attestation says observed state matches the claim.
Self-verification is not proof
Flow
How one run moves through the system
Five steps, each producing an artefact that survives the session that created it.
01
Objective and authority enter together
A human states an outcome and, in the same act, declares the authority envelope it may be pursued under: permitted scope, explicit denials, budget ceiling and expiry. Admission is a check against that envelope, not a formality.
02
DoneState executes durably
The run becomes a durable object. Leases prevent concurrent execution, idempotency keys prevent repeated side effects, and every state transition is recorded before the next begins.
03
Consequential effects go through AgentProof
Anything with real-world consequence is prepared, authorised against exact state, executed exactly once and signed. The receipt is the durable artefact, not the log line describing it.
04
OpsTruth observes from outside
The verifier reads repository, stack, test, build, CI, secret, config, route, runtime and deployment evidence at an exact commit. It holds no write authority over any of it.
05
Verification closes the state
An attestation from the verifier is what moves a run from AWAITING_VERIFICATION to VERIFIED. Nothing inside the execution path can perform that transition.
Artefacts
What the architecture leaves behind
Three durable artefacts per run: a state history, a signed receipt for each consequential action, and a classified verification report.
ADMITTED
Outcome accepted inside authority envelope
EXECUTING
Lease held, budget consuming
VALIDATING
Local checks and build evidence gathered
AWAITING_VERIFICATION
Halted — independent attestation required
VERIFIED
Attestation received, state closed
- action
- merge_pull_request
- prepared state
- sha256:9f2c41ab6e0d7c5183ba0e77c4d21f9a
- authority
- envelope/repo-write@exp-2026-08-29T09:00Z
- proposer
- agent:coding-harness
- executor
- agentproof:executor-01
- signer
- key:ed25519:AP-3f8c
- verifier
- opstruth:report-2f19
- outcome
- committed_once
Illustrative specimen. Field values show the shape of a receipt, not a real transaction.
| Check | Surface | State |
|---|---|---|
Build reproduces at commit Artefact matches source tree | build | verified |
Test suite executed Exit status recorded in CI evidence | tests | verified |
No credentials in tracked files Scan clean at this commit | secrets | verified |
Declared routes respond One route returns an error status | runtime | risky |
Deployed artefact matches commit No deployment evidence available | deployment | unproven |
Specimens are illustrative shapes, not records of real runs.
Go deeper
Two models worth reading in full
The state model governs what a run may do next. The authority model governs what it may do at all.
- Definition — Proof & State
- Proof & State is AI work accountability infrastructure: DoneState executes under declared authority, AgentProof authorises consequential actions and signs receipts, and OpsTruth verifies the result read-only from outside the execution path.
Give agents authority. Keep the proof.
The architecture is only as credible as the source behind it. Read both.