JARVISLabs

A spoken-style engineering question answered with exact, labelled calculations

Between 8 and 10 September a live engineering drive asked JARVIS to compare the pressure of 100 N on 1, 2 and 4 m square plates. All six recorded answers gave 100, 25 and 6.25 Pa, labelled as calculated, not measured. Turn time rose across the runs, cause not established.

LIVE-PROVEN What this run supports. Recorded 10 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

Compare the pressure a 100 N force puts on square plates of 1, 2 and 4 m, and show the engineering panel.

Result

What happened

The request arrived as a transcript through the conversation path (speech recognition was substituted). JARVIS computed the three pressures with its unit-checked engineering maths, answered 100, 25 and 6.25 Pa, labelled them as calculations, and opened the engineering widget. Turn times were 3,879 ms, 5,946.1 ms and 6,229.9 ms on successive days.

Pipeline

Architecture path

  1. Spoken-style request (transcript)
  2. Lane classifier
  3. Hosted cognition
  4. Engineering evaluator with unit checks
  5. Result labelled CALCULATED
  6. Engineering widget
  7. Spoken answer

What this proves

  • Two utterances times three runs all returned the oracle values (100, 25 and 6.25 Pa).
  • The answers were labelled as calculated, not measured.
  • The engineering widget opened as part of the turn.

What it does not prove

  • Speech recognition was substituted with a transcript, so the audio path is not covered.
  • A closed-form calculation is not a solver run, and a calculation is not a physical measurement.
  • Recorded 2026-09-08 to 2026-09-10; a recorded run is not a current reading.
  • The rising turn times are unexplained; they are not evidence of either a regression or noise.

Context

JARVIS’s engineering maths uses a Quantity type over the seven SI base dimensions. Converting one to a bare number raises an error: “A quantity is not a number.” Unit mistakes are rejected before any arithmetic happens. Units and dimensional analysis (CAP-ENG-01) is one of the nine capabilities the registry recorded as LIVE on 17 September. The engineering live drive is how the project exercised the conversational path to that code.

The run

The drive put a spoken-style request to a running JARVIS on isolated stores: compare the pressure of 100 N on square plates of 1, 2 and 4 m. Speech recognition was substituted with a transcript, and cognition was the hosted model. To pass, the answer had to be exactly 100, 25 and 6.25 Pa, labelled calculated, not measured, and the engineering widget had to open. Two utterances were run on each of three days, 8 to 10 September, and all six passed.

Turn times, one per day, were:

Date Revision Turn time
2026-09-08 4e90787c 3,879 ms
2026-09-09 4d47ca43 5,946.1 ms
2026-09-10 4aaaaeaf 6,229.9 ms

The times rose across the three runs. The project records this as a possible regression whose cause was not established.

The evidence

The drive’s result files are local and gitignored. The benchmark series cites them, and commit 991418c2f (12 September) records the engineering live drive restoring CAP-ENG-01 to LIVE. No footage exists yet.

The limits

This is a simple closed-form calculation, chosen because it can be checked exactly. A real solver run is a different claim. The registry records that a spoken request can drive a real CalculiX finite-element solve and light the engineering panel in tests (CAP-ENG-10), with no live probe, so this site keeps that at TESTED. The label on the answer is part of what was tested. A calculated number is never presented as a measured one.