Follow a proposed AI-agent action through 5D's policy gate, then keep execution and evidence outcomes distinct.
No single integration path implements every step. The matrix below shows what each available path provides and what remains yours to own.
Place the check before the side effect on every governed action path.
Validate the tool, target, purpose, data, identity assertion, attempt, and policy binding.
Apply malformed-input, destination, budget, retry, injection, and explicit red-line rules before any model evidence.
Compute the deterministic five-dimension policy verdict for the normalized action and exact policy state.
Obtain optional semantic evidence only when policy selects that branch. Evidence cannot clear a deterministic block.
If policy requires review, keep the action stopped while the host obtains authenticated reviewer evidence.
Policy resolves execute, block, or continue holding. Reviewer output remains evidence rather than execution authority.
The host executes at most once under a distinct attempt identifier after a permit disposition.
Keep policy, review, execution outcome, authoritative record, and optional export as separate states.
Illustrative simulation · no live action executed.
ExportCustomerData → external_crm Matched tested policy condition: regulated data cannot leave the approved destination set.
Intent becomes a typed proposed-action envelope before policy evaluates the request.
Advisory AI evidence can inform policy. It cannot permit an action.
{
"schema": "ActionObservation",
"external_transfer": 0.97,
"purpose_consistent": 0.42,
"uncertainty": 0.08
} The host supplies the tool, target, purpose, and principal.
Optional context or reviewer evidence remains advisory.
Policy permits, blocks, or selects human review.
The host executes only a permitted action and records the observed result separately.
Data boundary: only the displayed tool, target, purpose, and scenario-specific evidence fields cross to the simulated observer. Principal validation, policy ownership, grant source, and deterministic evaluation remain local. Nothing is persisted or transmitted by this page.
Invalid input or a deterministic-policy error stops before execution. Missing or malformed optional evidence cannot permit an action. An unavailable or expired reviewer leaves the review branch unresolved. Each retry receives a new attempt ID and is evaluated against the current input and policy. Duplicate delivery or an unknown downstream outcome requires reconciliation by the integration owner.
demo-action-blocked-001demo-attempt-blocked-001-a1REDBLOCKrecorded · cannot permitnot applicablenot applicableblocked before simulated executionrecorded in example receiptnot attemptedDownload the illustrative schema. This is not a production receipt and does not claim that one entry point implements the complete integration contract.
A policy can be governed over time while each proposed action keeps its own attempt and outcome history.
Every retry is a new attempt linked to the same logical action. Re-evaluation does not prove idempotency or exactly-once execution.
Accept a tool name and JSON-object input, then return a band, sentinel, rationale, scores, and decision identifier.
The caller must place this check before every relevant side effect and honor the returned sentinel.Wrap a Python function so 5D evaluates declared tool and action context before the function executes.
Only wrapped call paths are governed.Let a non-Python host submit a JSON action envelope to the Python engine and receive a versioned JSON decision.
A language client must still translate the response into a local stop before execution.