Validation gate

The Stage-1 deterministic evaluator exists in source. The complete validation gate remains a governed design for placing a server-owned decision boundary between an agent’s proposed action and an external effect, and it is not publicly callable through the frozen Northsun API today.

Intended flow

proposed action
  + versioned policy bundle
  + evidence-backed facts
  -> Stage 1 deterministic rules
  -> Stage 2 isolated qualitative grader when required
  -> allowed | denied | needs_human

Deterministic rules are intended for invariants that can be replayed without a model. Qualitative rules may use an isolated role-assigned grader. The grader does not own the external decision, and acting-agent code must not receive grader keys, prompts, configuration, or raw output.

The internal design distinguishes grader bands from external verdicts. Those names, mappings, retry behavior, HTTP statuses, and payload fields remain non-public until the runtime and interface graduate.

A real gate needs red and green proof

A passing example does not prove a gate. Graduation requires at least:

1. A prohibited fixture that is observed reaching denied or the designed equivalent.
2. A neighboring allowed fixture that is observed passing.
3. Evidence showing which bundle version, facts, and rules produced the result.
4. Rerun behavior on the exact stored inputs.
5. Failure-mode tests for timeout, missing evidence, unavailable grader, and tenant mismatch.
6. Proof that a denied high-impact action did not reach the external effect.

Internal engineering latency budgets are not public performance claims. Public latency language remains limited to “designed for sub-300ms voice recall,” which applies to the caller-memory path rather than a future validation API.

What the gate does not touch

The gate is an action checkpoint, not a retrieval ranker. It does not order, score, filter, or re-rank caller-memory results: recall_and_enrich returns stored caller memory newest call first, up to five, independently of any validation work described here. Reading this page as evidence that recall output is graded, ranked, or quality-scored would be wrong in both directions.

Current callable boundary

Do not send the illustrative gate payloads from older drafts. No validation endpoint belongs to the frozen three-route REST contract. Use API reference for current calls and Policy bundles for the design that would govern future actions.