Skip to content

ContextLattice v3.13.0 - Memory That Learns Without Grabbing the Wheel

Choose a tag to compare

View all tags
@sheawinkler sheawinkler released this 12 Jul 01:41
v3.13.0
bd8e399
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

ContextLattice v3.13.0 - Memory That Learns Without Grabbing the Wheel

v3.13 turns outcomes into policy evidence and repeated wins into reusable
skills. It does not confuse learning with permission.

What changed

  • Added an outcome-trained context-policy ledger. Only calibration-eligible
    outcomes can seed a candidate.
  • Added one-step candidate -> shadow -> canary -> promoted|rolled_back
    lifecycle gates. Phases cannot be skipped.
  • Added controlled control/canary comparisons across first-pass success, repair
    rate, follow-up tokens, and provider tokens. Evidence is candidate-, project-,
    and phase-scoped; mixed operator/persisted arms are rejected.
  • Added explicit policy ID, arm, and phase attribution to adapter and task-worker
    outcome reports.
  • Added Skill Foundry. Three repeated verified runs can become a draft; three
    separate holdouts must reproduce it; export requires named human approval.
  • Added explicit skill version, supersession, collision, and non-automatic
    retirement metadata so a generated update cannot silently replace behavior.
  • Bound candidate IDs to evidence digests and skill evaluations to immutable
    draft fingerprints. Holdouts require explicit identities and evidence refs.
  • Made policy and skill lifecycles replay-safe: stale policy transitions fail,
    candidate regeneration cannot reset phases, and identical draft replay cannot
    regress evaluated/exported state.
  • Treats promoted and rolled_back as terminal policy phases, and parses
    Skills Index root lists with the OS-native separator so Windows drive letters
    are not split as Unix path lists.
  • Added seven primary CLI commands for policy and skill workflows, plus native
    HTTP/tool/telemetry routes and five bounded contracts.
  • Expanded native-ownership, context-boundary, installer, audit, preflight, and
    public-core parity coverage from 17 to 19 capabilities.

The boundary

  • Public policy records are advisory. Even a promoted record has
    runtime_activation=false.
  • Skill exports are inactive artifacts. Public ContextLattice never writes them
    into an active skill root.
  • Infrastructure failures remain observable but cannot train context policy.
  • Canary promotion needs controlled evidence and a measurable benefit while
    every guardrail passes.
  • A material regression recommends rollback.
  • No model call, external network call, Python application service, or gateway
    subprocess was added.

Measured behavior

Five-run component benchmarks on the release development host:

  • Candidate generation from 100 outcomes: median 261342 ns/op, 284760 B/op,
    1079 allocs/op.
  • Canary gate: median 1082 ns/op, 2320 B/op, 24 allocs/op.
  • Skill draft from 20 verified runs: median 91512 ns/op, 58270 B/op,
    1004 allocs/op.
  • Model calls: 0.
  • External network calls: 0.

These are bounded administrative component benchmarks, not universal latency
promises. Reproduce them with docs/evals/v3.13-outcome-policy-skill-foundry.json.

Try it

contextlattice_policy_candidate --project contextlattice --pretty
contextlattice_policy_status --pretty
contextlattice_skill_foundry_status --pretty

The full workflow and payload shapes are documented in
docs/outcome-policy-skill-foundry.md.

Assets 5
Loading