Falsifiable guarantees
Due process for AI agents
Six product claims, each tied to the module that enforces it and the machine-generated release artifact that tested it.
The five personas are AI agents powered by GPT-5.6; portraits generated with ChatGPT Images; any resemblance to real persons is coincidental.
This is not a feature checklist. A guarantee appears here only when the checked-in registry resolves to one PASS claim in the generated claim map. Missing or malformed evidence fails the build-time join instead of leaving an unsupported row on the page.
| Guarantee | Enforcing module | Generated claim evidence |
|---|---|---|
| Bailiff has one toolThe Bailiff can request only execute_warrant(id); it cannot compose shell, test, or mutation commands. | backend/src/crosspatch/mcp/broker_server.py | product.specialist-contractPASS
|
| Refusal becomes ABSTAINRefusal, incomplete output, timeout, network failure, invalid schema, missing evidence, SDK failure, guardrail stop, and unknown verdict create no authority. | backend/src/crosspatch/orchestration/failures.py | product.fail-closed-abstainPASS
|
| Evidence stays untrustedModels receive sanitized UNTRUSTED_EVIDENCE envelopes; raw evidence bytes remain outside model context. | backend/src/crosspatch/evidence/views.py | security.evidence-boundaryPASS
|
| Review escalation is boundedA model-driven review escalation follows a recorded REMAND, materially changes the semantic output, and stops after two attempts per seat. Test-failure recovery and explicit human revision remain separate recorded triggers. | backend/src/crosspatch/orchestration/escalation.py | product.effort-escalationPASS
|
| Authority is single-use and hash-boundApproval binds one warrant to its incident, evidence head, base and patch hashes, paths, plans, runner and environment digests, expiry, and one consumed nonce. | backend/src/crosspatch/broker/store.py | runtime.warrant-boundaryPASS
|
| Publication is the read boundaryThe judge surface reads only explicitly published, sanitized projections and exposes no approval, mutation, shell, test-run, secret, or raw-evidence capability. | backend/src/crosspatch/runtime/readers.py | runtime.mcp-zonesPASS
|