JARVISLabs

Journal

Engineering journal

Write-ups of decisions, measurements, failures and corrections. Each entry lists the evidence it rests on and what it does not prove.

Subscribe with RSS 9 entries

  1. Latest 5 min read

    Advanced Systems, 22 September: a spatial engineering workspace on an unmerged branch

    In one day, six parallel branches added a revision-bound 3D scene, governed FreeCAD STEP operations, a STEP-to-slicer path, and browser grounding across frames and shadow roots. None of it is merged into production, and its live proofs used synthetic fixtures.

    • advanced-systems
    • engineering
    • cad
    • browser
    • branch

  2. 3 min read

    Filming JARVIS at sixty frames a second: one evening, one take, every stumble kept

    The site's footage was recaptured on 22 September from the booted production system in one continuous 1080p60 recording. Forty-five typed requests later: a real solve, cited research, a forget that held, a latched E-stop, and eleven things that did not go to plan.

    • footage
    • method
    • hud
    • honesty

  3. 5 min read

    Fitting is not generalising

    Across three model families, training loss on JARVIS's own examples kept falling while the transfer gate kept failing. This entry sets the likelihood numbers beside the behaviour numbers and explains why the project will not promote a model on its loss curve.

    • training
    • evaluation
    • local-model

  4. 5 min read

    A negative decision on the local language model

    On 21 September 2026 the project recorded that no local model candidate had passed its frozen gates and refused to start another full training run. This entry covers what was tried, what the gates measured, why more training was refused, and what came next.

    • training
    • decision
    • local-model

  5. 4 min read

    The governance chain as built, and why the E-stop is checked at every layer

    The simple story is proposal, policy, approval, E-stop, execution. The code works differently. The E-stop comes first and is re-checked at four points, approvals are consumed inside the policy decision, and nothing is done until the world has been read back.

    • governance
    • safety
    • architecture

  6. 5 min read

    Why nine of 342 capabilities are live

    The capability registry lists 342 entries and nine of them LIVE. On 28 July the same count read 256. The difference is not lost features. The project tightened what LIVE means, and most of those readings could not meet the new rule.

    • truth-ladder
    • qualification
    • capability-registry

  7. 4 min read

    Voice without samples

    JARVIS listens and speaks locally through a fourteen-stage pipeline. This entry walks through that pipeline, explains how the speaking voice was chosen and what its qualification still lacks, and says why this site publishes no voice audio.

    • voice
    • speech
    • qualification

  8. 4 min read

    From 3.1 s to 1.3 s: the 11 September latency work

    Two changes on 11 September cut the median conversation turn from 3,083 ms to 1,270 ms on a fixed set of 21 journeys. One took the spoken acknowledgement off the critical path. The other stopped sending all 101 tool schemas on every hosted turn.

    • performance
    • latency
    • conversation

  9. 4 min read

    The ONE HUD truth contract

    JARVIS has one HUD page, and it follows one rule: nothing renders without a source and an observed instant. This entry covers how the rule is enforced, what the accessibility and soak measurements show, and why the project publishes a fidelity score of 2.43 out of 10.

    • hud
    • truth
    • accessibility