A runtime verification and coordination fabric in which every coordinated action is coupled with structured evidence of compliance.
v1.0 · 28 April 2026ACE is a runtime verification and coordination fabric for agentic systems, in which every coordinated action is coupled with structured evidence of compliance with a governing specification.
It preserves a continuous, verifiable link between agent behaviour and human intent.
ACE grants agents operational space to pursue defined objectives. This agency is delegated and conditional.
Agents operate as Stewards whose authority remains valid only while they maintain an Evidence Heartbeat — a periodic or event-driven attestation of alignment.
The heartbeat is not only a liveness signal. Each attestation must include a semantic alignment signal: a structured assertion that the agent's recent outputs fall within the expected distribution for its current task context, based on the defined success predicate (P1) and measurement criteria (P2).
A heartbeat with an anomalous alignment signal triggers the same isolate-and-signal response as a missed heartbeat, with severity scaled proportionally. A heartbeat from a high-risk agent without a semantic alignment component is treated as a degraded heartbeat, not a valid one.
Failure triggers an isolate-and-signal response: actions are constrained or quarantined, incomplete operations are retracted where feasible, and deviations are surfaced for review. Responses must be proportional to the severity and reversibility of the deviation.
ACE acts as an arbitration layer governing multi-agent interaction. It ensures that agent actions remain within the defined enclosure, conflicts between agents are resolved through defined precedence and arbitration rules, and combined behaviours do not violate system-level constraints.
Evidence is a first-class requirement, equal in importance to action.
An action is not considered complete or committed within the coordination fabric until its evidence packet is recorded.
"Proof" is operationalised as verifiable signals derived from defined evaluation, constraint checks, and traceability mechanisms — not solely agent-generated assertions.
Evidence requirements scale with risk: high-risk actions require stronger verification (multi-agent validation, distributional checks); low-risk actions use lighter telemetry.
Each significant coordination event produces an evidence artifact containing:
| Feature | Traditional Orchestration | ACE |
|---|---|---|
| Logic | Deterministic flows | Goal-directed within constraints |
| Failure Mode | Execution error | Specification deviation |
| Visibility | Logs (post-hoc) | Evidence (continuous) |
| Trust Model | Implicit | Evidence-based |
Evidence generation carries cost. This cost constrains uncontrolled system expansion, prevents overproduction of low-value actions, and enforces alignment between effort and risk.
Evidence cost must be incorporated into success evaluation to prevent optimisation toward low-evidence strategies that distort true objectives.
ACE operationalises the governing system: the Philosophy defines responsibility, the Principles define structural requirements, and ACE enforces and verifies them at runtime. It provides the substrate for audit, monitoring, anomaly detection, and adaptive control.
We coordinate through evidence, not assumption.
We act with structured verification of alignment, not inferred intent.
In ACE, legitimacy is supported by evidence and continuously re-examined, not presumed and never revisited.
Evidence informs action, but does not confer final certainty.
All evidence remains subject to interpretation, challenge, and revision within the governance process.