Skip to the content.

Default-cache readiness alternatives — 2026-08-10

Status: INCOMPLETE. Issue #6178 tracks real integration/policy-engine runs and independent resource/cost witnesses.

Capability boundary

internal/vcachescore.DefaultReadiness, consumed by fak cachevalue status, decides whether cache behavior is ready for default-on posture. It requires independently proven cold-path correctness, the versioned fak.cache.default_usefulness.v1 verdict, explicit provider/kernel/context/external/forecast evidence planes with correct provenance, and no unsupported active-cache path.

This contract covers only that readiness gate. Aggregate usefulness scoring, economics, agentic activation, concentration/index planning, and artifact construction remain distinct benchmark debt in this multi-capability package.

Same-workload oracle

Every arm evaluates five ordered reports:

  1. fully witnessed planes and correct cold path — ready;
  2. provider-only evidence — blocked, default-usefulness reason;
  3. kernel evidence mislabeled OBSERVEDblocked, provenance reason;
  4. unsupported external active-cache path — blocked, unsupported-capability reason;
  5. otherwise-ready report with cold-path correctness removed — blocked, cold-path reason.

Correctness requires all five verdicts and reason classes. A scalar score that happens to separate some fixtures is not sufficient.

Arms

Arm Class Local status Boundary
fak native default-cache readiness gate native available real DefaultReadiness over committed report schema
usefulness-score threshold only tuned no-policy baseline available, incorrect committed readiness score threshold, no provenance/cold-path policy
fak + Prometheus first-class integration unavailable real metric export, rules, and read-back
fak + OpenTelemetry first-class integration unavailable real collector/exporter policy and read-back
OPA/Rego external unavailable pinned OPA and real Rego policy
Prometheus rules external unavailable pinned Prometheus rule evaluation
Datadog monitors external unavailable real ingestion and monitor evaluation
LangSmith evaluations external unavailable real trace/evaluation boundary

Unavailable product arms retain Available=false and zero measurements. An in-process imitation would not witness the named product.

Completion metrics

All arms must report true-ready, true-blocked, false-ready, false-blocked, reason mismatches, latency and throughput, CPU, peak RSS, input/network/storage bytes, setup/operator time, service/license charges, and total cost. Reproduction requires pinned versions/configuration, raw decisions, and independent read-back.

TestCompareDefaultReadinessLocalKeepsPolicyAlternativesExplicit locks the inventory, local oracle, baseline failure, and unavailable-arm zeros. BenchmarkDefaultReadinessFiveCases exercises all five real native decisions per iteration. Local timing is not a cross-product claim; no alternative is ranked until #6178 has complete real-boundary witnesses.