Serve and Promote Learned Routing Policy

lrp serve loads a versioned bundle and answers strategy: external on a loopback URL in the router's network namespace. The router still enforces caller access, PII, budgets, and request-shape eligibility before asking LRP.

Router modes

external_policy.mode values:

• baseline: no policy call. First eligible configured target is served.
• shadow: LRP is called and the recommendation is recorded. First eligible configured target is still served.
• enforce: the LRP recommendation is served. Omitted mode defaults to enforce.

Rollback of learned influence is mode: baseline or restoring the previous group config. That does not restore a former weighted mix by itself.

Shadow first

models:
  workload-staging:
    strategy: external
    external_policy:
      url: http://127.0.0.1:18093/route
      allow_hosts: [127.0.0.1]
      include_request: true
      mode: shadow
      timeout_ms: 750
      on_error: fail_closed
      feedback:
        enabled: true
      headers:
        X-LRP-Auth: ${LRP_POLICY_AUTH_HEADER}

include_request: true is trusted-infrastructure only. Configure group pii_filter first when prompts must not leave the router unredacted.

Feedback posts status, usage, cost, latency, TTFB, and selected target. It does not retrain quality. Quality labels arrive from offline verifiers or judges.

Synthetic demonstrations prove wiring. They do not authorize live target promotion. A public seed reference bundle is also an example for specific targets and data. First-run success is not automatic production readiness. Operator promotion gates and rollback: LEARNED_ROUTING_POLICY.md. Signed load: Operator-signed LRP bundles. External-policy contract: External Routing Policy.