Runtime Policies

A Delegation of Authority document tells a human what's permitted. A Runtime Policy tells a deterministic evaluator the same thing, in a form it can execute against an Intent in sub-millisecond time, with no ambiguity left for the evaluator to interpret. For the product-level explanation of why this exists, see Runtime Policies on the product side.

What compilation produces

{
  "policy_key": "vendor-payment-approval-v4",
  "version": 4,
  "status": "active",
  "scope": { "action": "vendor_payment", "principal": "Finance Manager" },
  "conditions": [
    { "field": "amount", "operator": "<=", "value": 10000 },
    { "field": "currency", "operator": "==", "value": "USD" }
  ],
  "effect": "allow",
  "constraints": { "requires_dual_approval_above": 5000, "evidence_required": true }
}

This is the artifact Runtime Authority actually evaluates against -- structured conditions and a fixed effect (allow, deny, or require_human_review), not a paragraph a model has to interpret at decision time.

Runtime evaluation

Every field in conditions is checked against the Intent's actual content. All conditions must hold for the policy to match; the first matching policy's effect becomes the Decision, combined with whatever the Authority Graph separately determines about delegated authority.

Deterministic decisions

The same compiled policy, evaluated against the same Intent, always produces the same effect. There is no model call, no sampling, and no version drift inside evaluation itself -- if a decision needs to be explained months later, re-running the same policy version against the same Intent reproduces it exactly.

Policy versioning

Publishing a new version doesn't edit the old one in place. Every version is retained, and every Decision's evidence records the exact version that was active when it was evaluated -- so a policy change next quarter never silently rewrites what a decision made this quarter meant.

Policy deployment

A policy moves through draft → review → compiled → active → retired. Only an active policy is evaluated against live Intents; a draft can be edited freely because nothing depends on it yet, and a retired policy stops being evaluated going forward without deleting the history of decisions made while it was active.

Why this differs from the governance document

The governance document remains the source of truth for what your organization has actually decided; the compiled policy is PayReality's operationalization of it, not a replacement. If the two ever disagree, that's a signal the compiled policy needs to be corrected to match governance, not the other way around -- Runtime Policies attest to enforcement of your authority, they don't originate it.