JARVISLabs

Building

Roadmap

What the project is working on and what is waiting. Items move only when evidence moves them. Owner-only and security-sensitive work is not listed.

No dates are promised. Last updated 23 September 2026. Suggest something

Now 11

Being built, tested, integrated or proven.

  • INTEGRATING Advanced Systems

    Merge Advanced Systems into production

    The Advanced Systems branch now holds the spatial engineering workspace, governed CAD, print readiness, a what-if engine, contradiction review, evidence classes, cognition Cut B and adopted operator work. Several features have live proofs on an isolated diagnostic core with synthetic fixtures. None of it is merged into production.

    Blocked by

    • Live proofs used synthetic fixtures on isolated cores, not the owner’s everyday runtime
    • The last full test run (17 September) had 277 failures and 34 errors to classify
    • No production merge has been authorised

    Capabilities

    Updated

  • INTEGRATING Cognition

    Cognition Cut B: a live proof on the canonical branch

    On the night of 22 September, 32 cognition commits (“Cut B”, including request-specific bounded context selection) were adopted into the Advanced Systems branch, and a stack run gave 2,114 passed and 15 failed, with the real failures fixed. No canonical live proof exists yet; it is queued after Cut C.

    Blocked by

    • Live proof queued after Cut C
    • Worker commits after the adoption are not on the canonical branch

    Updated

  • TESTING Local model

    Local language model: behavioural viability study

    A bounded study of stronger local models under a frozen memory and latency envelope. Qwen3.5-9B came closest but failed transfer. About twelve single-variable Gemma 4 12B experiments on 22–23 September all failed behavioural Gate C; the best equalled the untrained base on transfer. The study found the prompt itself taught a fabricated check, and the owner chose contract v2. No model is development-qualified.

    Blocked by

    • Behavioural Gate C (v1) not passed by any candidate
    • Contract v2 battery and Gate C v2 not yet run: the next step compares the untrained base on both contracts

    Updated

  • TESTING Voice

    Speaking voice technical qualification

    The owner-selected local voice is integrated on the integration branch and assembled with the conversation loop on 2026-09-18. Its qualification is pending. Long-form non-regression failed, and first audio, streaming, acronym pronunciation, audible cancellation, restart recovery and fallback have not been measured in qualification.

    Blocked by

    • Long-form non-regression check failed
    • Audible first-audio measurement reserved for the owner-present sitting

    Capabilities

    Updated

  • TESTING Governance

    Production reach for the intent ledger

    A 2026-09-18 checkpoint found no production code constructing the intent ledger, so multi-step plans containing an irreversible command were safely refused and could never run. Explainable-plans work (A35) targeted the gap. Whether it is fully closed is uncertain.

    Blocked by

    • Closure not independently verified

    Capabilities

    Updated

  • BUILDING Local model

    Model-visible contract v2 with bounded tool discovery

    Contract v2, chosen on 23 September: persona 2.1.0 lets JARVIS say “no record” only after a search it actually ran; the deterministic body computes simple facts such as receipt validity; a regenerated training set and a separately frozen battery v2 with its own gate. Bounded discovery replaces offering every tool schema: at most two discovery calls, then one action completion.

    Blocked by

    • Persona 2.1.0 is tested but landed inactive
    • Context delivery defect recorded as unresolved
    • Body-derived facts cannot cover comparisons over owner-typed prose

    Capabilities

    Updated

  • BUILDING Advanced Systems

    Advanced Systems requirements census and atomic decomposition

    Break 417 source requirements into atomic capabilities with exact source quotes and parent-child links. On 23 September the ledger held 894 rows: 37 COMPLETE (36 atomic capabilities and 1 source requirement), up from 1 the day before. 1 of 417 source requirements is complete and 44 are partly built.

    Blocked by

    • 416 source requirements not yet complete
    • COMPLETE requires a live proof, not only code and tests

    Updated

  • BUILDING Computer operation

    Governed typing and filling on the frame-aware browser path

    The Advanced Systems DOM scene reads across frames and shadow roots and refuses stale targets, but only for reading. Typing and filling on this path are still pending validation. In production, browser click and fill remain authority-gated by design.

    Blocked by

    • Mutation validation pending on the branch
    • Depends on Advanced Systems integration

    Capabilities

    Updated

  • BUILDING Computer operation

    Installed-application journeys at release scale

    Calculator ran ten verified steps on 2026-09-19, and Paint, Notepad and Calculator passed a partial semantic proof. A 78-action workflow across three applications verified every action on the Advanced operator branch, but the requirement is four distinct applications. Full-universe installed-app journeys are owed at release.

    Blocked by

    • Mixed-application proof covered 3 of the required 4 applications
    • Partial UIA traversal is not full application-state coverage

    Capabilities

    Updated

  • BUILDING HUD

    HUD twelve-category fidelity gate

    The project's own HUD fidelity gate requires every one of twelve categories above 9.0. The latest recorded mean is 2.43, with a minimum of 1.42, and the scores were recorded against an earlier surface. The gate is not met.

    Blocked by

    • Scores need re-measuring against the current ONE HUD surface

    Capabilities

    Updated

  • BUILDING Research and reasoning

    Predict-before-act as one layer

    Pieces exist separately: calibration records predictions before outcomes, missions can be simulated as SIMULATED / not_proof, and the Advanced operator lane binds experiment predictions to observation receipts. A single layer that records a prediction before every dispatch is still open.

    Blocked by

    • 'Record prediction before dispatch' is open for integration on the Advanced branch

    Capabilities

    Updated

Next 8

Committed to, not started.

  • PLANNED Local model

    Production training campaigns and the SEALED evaluation

    Freeze an immutable production corpus, run at most five full campaigns and stop at the first development-qualified model. Open the 450-item SEALED test set once, for that one frozen checkpoint. None of the five authorised campaigns has been used.

    Blocked by

    • Behavioural Gate C not passed by any candidate (0 of 5 campaigns used)
    • Contract v2 not yet frozen into a production corpus

    Updated

  • PLANNED Local model

    Integrate a qualified local model into the normal provider path

    Serve the exact qualified package behind the same provider gateway the hosted model uses, bound to loopback, with the pinned base, revision and adapter digest checked on every streamed chunk and no silent fallback. Hosted cognition stays in production until then.

    Blocked by

    • No development-qualified model
    • SEALED evaluation not yet run

    Updated

  • PLANNED Qualification

    The owner-present physical sitting

    One session at the machine covers what software cannot: an audible reflex first-audio reading, the owner's own voice interrupting a live answer (journey J8), a hand in frame moving a panel (journey J10), and the camera acceptance runbook. J8 and J10 stay FIXTURE_TESTED until then.

    Blocked by

    • Scheduled after the release-candidate freeze, which waits on a qualified model

    Capabilities

    Updated

  • PLANNED Qualification

    Release-candidate freeze and qualification (Z08.P14)

    Freeze an exact candidate and qualify it serially: full suite, mandatory gates, security, mutation, journeys, soak, recovery, backup and restore, restart and shutdown. The release row Z08.P14 has been active since 2026-08-30, final_gate_passed is false, and the certification matrix reads 'JARVIS-READY-V1: NO'.

    Blocked by

    • No full suite recorded at the current head; the last full run (2026-09-17) had 277 failures and 34 errors before partition
    • Tier 1 mutation score 73.5% against the harness's own 90% target
    • Waits on the local-model decision

    Updated

  • PLANNED HUD

    Re-measure the ONE HUD on the final surface

    The ONE HUD's own performance acceptance still reads NOT YET MEASURED. Later UI edits invalidated the keyboard walk of 2026-09-18, and fifteen HUD instrument artifacts carry stale source stamps. The project re-measures rather than re-stamps.

    Blocked by

    • Several instruments need the native shell and thirty-minute soaks
    • Must follow the last UI change

    Capabilities

    Updated

  • PLANNED Missions

    Missions that survive a real machine reboot

    Missions persist before step one and resume after an application restart without duplicate effects. Surviving a full machine reboot (CAP-MSN-07) is implemented but not proven, and the project will not accept a process kill as a substitute.

    Blocked by

    • Requires an actual reboot of the owner's machine

    Capabilities

    Updated

  • PLANNED System health

    Live chaos and recovery matrix

    Bounded recovery and self-healing are tested, and child-process death and restart storms were read truthfully on /health. The live chaos matrix, which injects failures into a running system and verifies each recovery path, is still owed.

    Blocked by

    • Link-lost HUD recovery not yet verified live

    Capabilities

    Updated

  • PLANNED Qualification

    Prove JARVIS boots and runs without the build assistant

    The last step of the owner's completion order. After promotion to local main, JARVIS must boot and run with the AI build assistant absent, which shows the system does not depend on the tools used to build it.

    Blocked by

    • Follows release qualification and local-main promotion

    Updated

Later 1

Ideas and deliberately postponed work.

  • DEFERRED Fabrication

    Physical 3D printing

    Print preparation works in software. On 23 September, print readiness was proven across three boots of an isolated core with a real slicer: READY / NOT READY / WHY in 0.125 s, a job invalidated when the CAD revision changed, and no printer contacted. The physical start needs a fresh single-use approval bound to the G-code hash, a responsible adult present, a clear E-stop and a proven stop path. The final send refuses every caller today, and no evidence shows JARVIS has ever started a print.

    Blocked by

    • Start gate refuses until a pause has been observed on a real running print
    • Requires an owner-present check under existing governance

    Capabilities

    Updated

What the status words mean

PLANNED
Committed to, not started.
BUILDING
Being implemented.
TESTING
Implemented, under test.
INTEGRATING
Tested, being wired into the production path.
DEFERRED
Deliberately postponed.