JLens Evolve v3

Evidence-centric evolution platform for frozen Agent programs. Strategies submit neutral ExecutionPlan values; the single Runtime emits immutable Receipts; Observers emit Evidence; Claim Engine classifies native outcomes; only Governance may approve promotion. Candidates and capabilities enter registries inactive.

python3 -m pytest -q
PYTHONPATH=src python3 -m evolve --help

The product entry point can execute a fresh three-task, six-arm feedback campaign:

PYTHONPATH=src /path/to/legacy/.venv/bin/python -m evolve fresh-feedback-e2e \
  --config /absolute/path/to/FRESH-FEEDBACK-CONFIG.json \
  --output /absolute/path/to/run
PYTHONPATH=src python3 -m evolve verify-manifest \
  --manifest /absolute/path/to/run/EVIDENCE-MANIFEST.json \
  --root /absolute/path/to/run

For continuous feedback-only Skill/Harness evolution from a frozen Model and a budgeted Teacher, use:

PYTHONPATH=src python3 -m evolve autonomous-evolve \
  --config /absolute/path/to/AUTONOMOUS-EVOLUTION-CONFIG.json \
  --output /absolute/path/to/evolution-run \
  --worktree-root /absolute/path/to/clean/committed/source

See docs/AUTONOMOUS-EVOLUTION.md and the example config. The loop automatically selects feedback tasks, runs the current best Harness as baseline, compiles the Teacher Candidate, performs real local model generation and official matched native evaluation, feeds authoritative Claims into the next round, and exports an inactive BEST-HARNESS.json.

For autonomous Skill evolution, that command is the authoritative live loop. It validates three clean, revision-pinned feedback checkouts; freezes the local Qwen and official evaluator identities; compiles a byte-frozen DeepSeek request/response into an immutable CandidateChangeSet, Skill, zero-argument Operator and Router; and dispatches baseline/taught generation and native evaluation only via ExecutionRuntime. Baseline never reads the compiled candidate, while taught must consume and hash it or fail closed. An E2 claim binds a frozen MatchedCounterfactualPair: both model receipts, external traces, native outcomes, the matched execution identity, and the taught Candidate revision/bundle. Native metadata alone remains E1. E3 additionally requires a MechanismPrediction receipt frozen before model dispatch and independently signed trusted-jlens-v1 observations whose artifact bytes, prediction receipt, model subject and receipt order reverify against a process-local trust root; an observer_id, renamed receipt, self-reported hash or model-provided internal_trace is never sufficient. Before E3 projection, every E2 pair is rebuilt from the Receipt Store and its literal receipt kinds/artifacts. Missing trusted evidence or receipt replay remains E2.

Every round starts with a hash-bound Qwen/native/Teacher health preflight. A failed evaluator is recorded as infrastructure feedback with no gain Claim; it does not update BEST, and the configured consecutive-infrastructure threshold stops the loop before an unattended batch can spin indefinitely. Resume rebuilds state and next-round feedback from the last verified round rather than trusting mutable summaries.

Every legal outcome enters Candidate Registry. Governance projects neutral as no_change, regression as rejected, and evaluator infrastructure failure as blocked; only regression creates a Rejected Registry record. Only an approved E3 decision with explicit human approval can create an inactive Capability revision. Approved decisions are signed by the configured process-local Governance authority; the decision log re-verifies that signature before the Capability Registry will project it. Teacher cost authorization is an append-only sequence/hash chain with an atomic head anchor, restart recovery and replay deduplication. budget_integrity_status=validated is projected only by replaying that ledger; report metadata cannot self-assert budget validity. The command does not open holdout, retrain model weights, auto-activate Skills or mutate legacy data.

Fresh configuration must bind the final Git SHA, exactly three feedback tasks, the frozen model/harness paths, and byte-frozen teacher_request and teacher_response paths. Historical operator_skill_path/span_skill_path inputs are no longer accepted by the live taught path.

Trusted JLens is optional. Without it the command can reach at most E2. To enable the E3 observation path, add this object to the frozen config and provide the secret only through the named process environment variable:

{
  "trusted_jlens": {
    "trace_root": "/absolute/path/to/canonical-jlens-traces",
    "secret_env": "JLENS_OBSERVER_SECRET",
    "key_id": "jlens-production-key-1",
    "implementation_id": "jlens-observer-v1",
    "implementation_sha256": "<64 lowercase hex>",
    "observer_config_sha256": "<64 lowercase hex>",
    "prediction_id": "role-commitment-v1",
    "expected_internal_effect": {
      "concept": "declared-role",
      "phase": "symbol-selection",
      "min_final_score": 0.7,
      "min_location_count": 2,
      "require_non_decreasing": true
    }
  }
}

The trace source must create one canonical JSON file named <plan_id>.json in trace_root before that plan's observation stage. Its only top-level field is locations; each location carries layer, token_position, phase, and a concept_scores object. The secret is never serialized into campaign artifacts.

legacy-feedback-e2e remains available only as a historical evidence-import compatibility command. Public AgentProgram campaigns support a deterministic fixture profile and one allowlisted local live profile. The bounded PortfolioOrchestrator connects one authoritative AgentProgram failure to an inactive validated Capability and a new live tournament without minting Claims or activating assets. See docs/MIGRATION-v3.md and docs/TRUST-BOUNDARIES.md for authority and threat boundaries.

License

MIT