Conformance is layered so teams can adopt deterministic recovery incrementally.
A conformant machine:
- Uses a
RECOVERY.yamlfile that validates against the normative schema. - Declares an initial state, accepted events, transitions, and a final state.
- References guards and actions by registered identifier.
- Contains no scripts, executable expressions, prompts, or credentials.
- Never permits retry directly from unknown state after a possible side effect.
A conformant tool integration additionally:
- accepts or returns operation, correlation, and idempotency identifiers;
- reports whether a side effect was possible and confirmed;
- provides an authoritative verification mechanism;
- emits deduplicated recovery events with evidence references; and
- declares whether compensation is supported.
A tool MAY use a fast classification model to propose a candidate recovery event or an advisory urgency signal, provided the classification is recorded as evidence and the controller still validates the event under Level 3. A classifier is never a guard or a transition authority (see specification §11).
A conformant controller additionally:
- validates and pins the machine version and provenance;
- stores current state and evidence in a durable Action Ledger;
- resolves guards and actions through administrator-controlled registries;
- selects transitions deterministically and fails closed;
- records a transition before invoking its actions;
- handles action results as new events; and
- enforces enterprise policy when it is stricter than the machine.
An enterprise-conformant deployment additionally:
- can separately stop execution and revoke authority;
- can isolate an affected capability, destination, or delegated workflow;
- supports risk-driven human escalation and authenticated decisions;
- enforces a resume gate after material incidents;
- correlates recovery events with security and business telemetry; and
- exercises recovery through failure injection and response drills.
A multi-agent-conformant deployment additionally:
- preserves origin and cascade lineage across agents and protocols;
- identifies and stops descendant tasks and queued work;
- revokes delegated credentials throughout the affected chain;
- reconstructs completed effects for every participating agent; and
- prevents a child or peer agent from silently resuming a contained incident.
- timeout before a side effect;
- timeout after a side effect;
- duplicate event delivery;
- duplicate business operation detection;
- verification unavailable or conflicting;
- partial multi-step completion;
- guard resolution failure;
- rejected invalid transition;
- failed compensation;
- containment action failure;
- authorization conflict;
- descendant-task cancellation; and
- resume-gate rejection.