Declare
Schedule, objective, constraints, budget, coordination, memory, workspace.
System score / five durable movements
State before spectacle.
BATON’s architecture is a compact organizational kernel. It records coordination decisions before visualizing them, and recovery is a normal state transition rather than an exceptional UI path.
I / Runtime topology
Five layers. One record.
Each layer has one narrow job, so an operator can follow an objective from scheduled intent to durable evidence without reverse-engineering a framework’s internal graph.
Pin
Version the exact instructions and active advisories read by the run.
Coordinate
Lease a scope with an opaque owner token or join its FIFO queue.
Recover
Commit step, workspace ref, harness state, and handoff version.
Explain
Append the transition and render it into the shared organizational score.
II / Operational contracts
The boundaries are the interface.
The prototype makes its consistency points explicit. That keeps claims and recovery testable with no external service and makes omissions visible.
Claims
Compare-and-delete release.
Only the opaque owner token can release a live claim. Expiry or valid release grants the next queued contender in deterministic order.
Advisories
Data, never hidden commands.
Topic-scoped advisories carry publisher, severity, TTL, and provenance. Agents receive them only at a step boundary and record their reaction.
Recovery
Resume from the durable edge.
Runtime.resume restores
the latest checkpoint and its exact handoff. Work after that edge must be
re-executed, never imagined complete.
Memory
Abstention is a valid verdict.
Invalid or missing comparison evidence yields UNDETERMINED.
Positive registered deltas promote; non-positive evidence archives.
Control plane
One API over the same runtime.
The local HTTP service performs real state transitions, returns structured errors, exposes evidence bundles, and serves the score same-origin when requested. It does not simulate an agent loop.
III / MVP limits
What this score does not claim.
- Single nodeSQLite coordination is process-local; there is no distributed consensus or cross-host lease service.
- Harness scopeClaims cover BATON-dispatched work, not arbitrary operating-system writes.
- Recorded agentsThe demo is deterministic and ships no model SDK, provider account, or quality claim.
- Fixture evalThe promoted lesson proves gate behavior for registered scores, not compounding from run 1 to run 20.
- UI evidenceScenario controls are client-side projections. Fixture mode is the only measured replay; installed live mode reads real events but does not claim the caller’s work succeeded.
See the state move