@metaharness/avo

Governed autonomous variation for real repositories and evaluator suites.

The AVO package exports an autonomous variation API and versioned evidence schemas.

The model chooses repeated inspect, search, hypothesis, edit, execute, evaluate, repair/revert, branch, memory, and commit actions. MetaHarness retains immutable authority over capabilities, budgets, protected invariants, promotion, quarantine, rollback, and signed receipts.

import { GovernedVariationOperator } from '@metaharness/avo';

const result = await new GovernedVariationOperator({
  // Inject a model/host agent, bounded repository adapter, evaluator suite,
  // capability policy, approval gate, RVF memory, supervisor, signer,
  // checkpoint store, budgets, and protected invariants.
}).run();

Production memory uses the optional agenticow peer over RVF. Install it when using RvfGovernedMemory:

npm install @metaharness/avo agenticow

The initially evolvable surfaces are retrieval policy, model routing, context policy, test policy, and repair strategy. Security policy and capability expansion are not evolvable surfaces.

DarwinVariationAdapter is the versioned, structural bridge to Darwin's opt-in EvolutionConfig.variationOperator. When present it replaces the one-call CodeGenerator child path; the outer Darwin loop still evaluates, archives, promotes, budgets, and rolls back the returned candidate.

Proofs

npm run build
npm test
npm run benchmark:swebench

The test suite includes a real RVF run killed and restored at actions 100 and 200, requiring identical signed receipts, state hash, evaluator evidence, lineage, and winner. The default benchmark data is a synthetic mechanism fixture and cannot authorize an AVO-class product claim. Supply preregistered observations with SWE_BENCH_RESULTS=/path/results.json for the 100-task ship gate.

npm run claims:avo verifies the governed release surfaces. Performance and frontier claims fail closed unless an exact-tag-SHA evidence bundle binds the exact npm tarball, preregistered predicate, measured costs, complete lineage roots, and two official grader receipts from distinct pinned organizations. The trust policy digest must come from the protected AVO_CLAIM_TRUST_POLICY_HASH release variable; a key embedded only in the bundle is never trusted.

See ADR-251 and ADR-276.