JARVISLabs

Operating Calculator end to end, ten steps verified

On 19 September an isolated real JARVIS core launched a new Calculator window and ran ten steps against it: scope, maximise, restore, inspect, find a control by name in 249.1 ms, invoke it, set an exact rectangle, close and shut down. Each step was independently verified.

LIVE-PROVEN What this run supports. Recorded 19 September 2026.

Footage pending

This run is documented from its recorded evidence. Footage has not been published: capture waits for the machine to be free of local model measurements, and every clip is reviewed for private information first.

Asked

What was asked

Operate a real Windows application end to end through JARVIS's own command path, verifying every step.

Result

What happened

An isolated real core launched Calculator and scoped the new window. It then maximised and restored it, inspected 27 controls within a native-call budget, found one control by name in 249.1 ms, invoked it, set an exact window rectangle, closed the window and shut down cleanly. Every step reported an independent verification.

Pipeline

Architecture path

  1. Owner-style command
  2. Command table
  3. Governance chain
  4. Launch and scope
  5. UI Automation effects
  6. Semantic query
  7. Semantic command
  8. Independent re-read per step
  9. Close
  10. Clean shutdown

What this proves

  • JARVIS's semantic operator can drive a real installed Windows application through launch, geometry, semantic query, semantic command and close, with each step verified.
  • A named control was located in 249.1 ms within a 750 ms deadline.
  • The window close was verified and the core shut down cleanly.

What it does not prove

  • One application and one run are not universal application control.
  • Recorded on 2026-09-19; a recorded run is not a current reading.
  • Inspection was truncated at its budget, so this is partial application-state coverage.
  • Isolated stores were used; this was not the owner's everyday runtime session.

Context

Operating someone’s real computer is where assistants usually fail quietly: the wrong window, a control that moved, success reported because the click was sent. JARVIS’s desktop operator prefers native APIs and UI Automation over synthetic input. It treats a dispatched action as unverified until a separate read of the world agrees with the expected state.

The run

The probe started an isolated real JARVIS core and launched a new, owned Calculator window, so no existing owner window was touched. It then ran ten steps:

  1. launch (present, focus verified);
  2. scope to the new window;
  3. maximise (VERIFIED);
  4. restore (VERIFIED);
  5. inspect: 27 controls, truncated at the budget, 1,014 native calls;
  6. semantic query: one control found by name in 249.1 ms;
  7. semantic command on that control (VERIFIED);
  8. set an exact window rectangle (VERIFIED);
  9. close (verified);
  10. clean shutdown.

All ten steps were verified. The next night a bounded development proof inspected Paint, Notepad and Calculator semantically, verified same-state restore for Paint and Calculator, and verified Calculator’s close through the canonical command.

The evidence

The probe record docs/ledger/evidence/UNIVERSAL_APP_OPERATOR/calculator.json was measured at 02:51 UTC on 19 September, starting from head 71fbf83e, and binds the hashes of the source files it exercised. No footage exists yet.

The limits

The record lists its own limits. Notepad and Paint focus were unverified in earlier bounded probes. The external computer-use pipe was unavailable, so nothing here depends on it. Universal multi-application qualification is pending, and no release qualification is implied. The follow-up proof on 20 September labels itself “PARTIAL_DEVELOPMENT_PROOF_NOT_RC_CERTIFIED”. It records no typing, no semantic click and no browser journey.