Product documentation
The Execution Path
Syndicate Code provides structural enforcement by separating the non-authoritative AI runtime from the authoritative control plane.
Proposal
The runtime submits a structured JSON request containing the proposed action, arguments, and context. The control plane records this intent before any evaluation occurs.
Canonicalization
Arguments are normalized and hashed. This creates a stable identifier for the action that is used for policy matching and approval binding. No fuzzy matching during enforcement.
Policy Evaluation
The control plane evaluates the canonical action against the active policy bundle. It returns a decision: ALLOW, DENY, or CHECKPOINT. The decision is recorded in the local ledger.
Checkpoint
If required by policy, execution suspends. A human operator reviews the exact arguments and context. Approval binds the operator's identity and decision to the action hash.
Permit
Upon successful evaluation or approval, the control plane issues a short-lived, single-use execution permit. The permit is cryptographically linked to the approved action.
Evidence Generation
The action executes. The outcome (stdout/stderr/exit code) is captured and recorded in the append-only audit chain, linked to the proposal and permit.
Authoritative Control
In this model, the AI tool never executes code directly. It submits a proposal to the control plane, which manages the lifecycle of that action. If the control plane process is terminated, execution stops. If the policy database is unavailable, execution is denied. This "fail-closed" behavior is the core security guarantee of the system.