JARVISLabs

Reflective tool-use controller

After using a tool, JARVIS checks whether the result was actually verified and revises the plan from receipts.

LIVE-PROVEN Recorded 23 September 2026. Advanced Systems branch, not merged into production.

Technical

What it is

jarvis/reasoning/judgment.py (jarvis.reasoning.plan-revision/1, tool-use-reflection/1) with an orchestrate.py reflection key.

Truth ladder

Where it stands

Rung 6 of 8: LIVE-PROVEN

A bounded probe ran the real code path on a booted JARVIS on the real machine and recorded evidence. The project treats live evidence as expiring after 24 hours, so every live reading here shows the date it was taken.

Qualified as PARTIAL: part of this capability is not at the rung shown.

Project record: PRODUCTION_REACHABLE + canonical live proof, PARTIAL

  1. CONCEPT Designed or specified. No working code yet.
  2. IMPLEMENTED Code exists. Nothing has demonstrated that it behaves correctly.
  3. TESTED Automated tests exercise it in isolation, often against fixtures.
  4. INTEGRATED Wired into the rest of JARVIS and tested across component boundaries.
  5. PRODUCTION-REACHABLE The owner can reach it through the real runtime path, not only through a test harness.
  6. LIVE-PROVEN A bounded probe ran the real code path on a booted JARVIS on the real machine and recorded evidence. The project treats live evidence as expiring after 24 hours, so every live reading here shows the date it was taken.
  7. FORMALLY VERIFIED Its critical property is checked by a proof or exhaustive verification.
  8. RELEASE-QUALIFIED Passed every qualification gate for a named release.

Limits

Limitations

  • The plans in the proof were supplied by the test harness, not proposed by a local model: plan proposal still depends on the language model.
  • Proven on an isolated diagnostic core with synthetic test fixtures, not on the owner’s everyday runtime.

Record

Status history

  1. TESTED Recorded state: IMPLEMENTED/TESTED; lane live proof PASS on an isolated core; INTEGRATED false; matrix UNDISPOSITIONED
  2. LIVE-PROVEN PARTIAL Canonical isolated-core proof: after a refused edit the controller diagnosed a changed state (stale target), revised the plan without replaying anything, and revalidated after a restart.

Only states recorded in the public export are listed. Earlier states may exist in the project’s private ledgers.

Connected work

Connected work

Neighbours

Related capabilities