September 2026 42 changes
- safety Advanced Systems
Print readiness: READY / NOT READY / WHY, and a handoff that only asks
A lane added print readiness across nine checks and a governed handoff bound to a READY receipt that requests approval and does not print. 'READY IS SOFTWARE READINESS AND NOTHING MORE.' No printer was contacted.
- capability Advanced Systems
Reflective controller revises plans from receipts
A cognition-lane controller checks whether tool use actually verified and revises plans from receipts. 'A dispatch that did not verify stays unverified.' It is tested in its lane and not yet integrated.
- capability Advanced Systems
Experiment predictions bound to observation receipts
The operator lane binds experiment predictions to immutable observation receipts, a first step toward predict-before-act. Recording a prediction before every dispatch is still open for integration.
- capability Advanced Systems
Sealed STEP cavities that leave outer geometry unchanged
FreeCAD can now hollow a solid into a sealed inward-offset cavity. On a 20 mm sphere with a 1 mm wall, the cavity volume matched the closed-form value to floating-point precision. This is a geometry calculation, not a physical measurement, and nothing was printed.
Governed FreeCAD STEP operations (booleans, split, offset, mirror, transform, fillet, faces, face extension, sealed hollow) Journal entry
- fix Advanced Systems
UI Automation geometry stays correct under display scaling
The operator lane now preserves physical geometry across DPI transitions. A disposable window moved across four displays with physical-pixel readback, and the E-stop refused a later move. The lane's combined acceptance passed 545 tests across 43 files.
Move window between monitorsPost-action geometry verification
- capability Advanced Systems
Browser grounding across frames and shadow roots, with stale-target refusal
A DOM scene reader records relationships across iframes and shadow roots, and dom-target-check/1 refuses targets that moved, were replaced, were covered or changed. In a fixture proof, 500 read-only reads verified across a restart, and a same-label replacement was refused.
Browser DOM scene with frames and shadow roots; stale-target refusal Demo Journal entry
- capability Advanced Systems
Specialist checkpoints and bounded advisory contracts
Missions can checkpoint specialist work with immutable revisions, and 11 specialist roles have explicit contracts. 'Scope is a ceiling, never execution authority.' Live proofs used fixture tasks, and no autonomous specialist dispatch is claimed.
Bounded specialist decompositionSingle JARVIS identity over silent specialists
- capability Advanced Systems
STEP to verified slicing mesh, with slicer estimates bound to engineering runs
The installed OrcaSlicer 2.4.2 rejected STEP input, and that failure was kept as a receipt. FreeCAD tessellation now produces a verified closed mesh. A synthetic 10 mm cube sliced to a 325 s and 0.73 g estimate. These are slicer estimates, not a print.
STEP to verified slicing mesh and slicer estimates Journal entry
- capability Advanced Systems
Recorded engineering trade spaces
Engineering runs can be compared across a recorded trade space with explicit objectives, constraints, archived evidence, HUD plots and resumable sweeps. 'Nominal frontier is not feasibility' or approval.
- capability Advanced Systems
Governed FreeCAD STEP operations
Fixed, isolated FreeCAD scripts compute STEP booleans and, later the same day, split, components, transform, mirror, fillet, offset and face operations. Each result is journalled through PREPARED, CALCULATED, FILE_VERIFIED and REGISTERED as a CALCULATION. Proofs used synthetic solids.
Governed FreeCAD STEP operations (booleans, split, offset, mirror, transform, fillet, faces, face extension, sealed hollow) Demo Journal entry
- hud Advanced Systems
A revision-bound 3D scene inside the ONE HUD
A persistent scene store (at most 64 objects and 4,096 revisions) imports STL, OBJ and 3MF, displays STEP, and supports compare, labels and selective restore. A verified spin preview renews a short E-stop lease, and the measured foreground stop took 31 ms. The project labels that 'not worst-case 150 ms qualification'.
Spatial engineering workspace (scene)Load STEP designs into the spatial scene Journal entry
- capability Advanced Systems
Obsidian projection of project memory and a Failure Atlas
Project memory can be exported to an Obsidian vault as 'a human-readable projection, never a second memory authority'. Failure history is kept with evidence links, and 'no reported root cause is silently upgraded into verified causation.'
Governed Obsidian projection of project memoryFailure Atlas (diagnosis history and cross-trajectory patterns)
- infrastructure Advanced Systems
Advanced Systems begins: six branches and bounded capability discovery
Six parallel branches started from production head 893913a9c. The first commits added a temporal world blackboard, a lossless requirements census and read-only capability discovery, which 'has no execution or grant method'. None of this work is merged into production.
Bounded capability discovery over the body interface Journal entry
- training Post-LM stronger-local study
Gemma 4 12B passes the training-safe endpoint check (Gate C)
One real optimizer update with a 4,096-token backward pass, exact loss parity and a same-base checkpoint reload passed independent review. This shows the model can be trained on the hardware. It says nothing about behaviour, and its behavioural gates are pending.
- infrastructure Post-LM stronger-local study
Model, speech recognition, voice, HUD and body fit together on one GPU
One clean run loaded the Gemma inference recipe with speech recognition, the voice worker, the HUD, body services and a browser, with a whole-GPU peak of 11,025,776,640 bytes, then shut down cleanly. It shows memory fit only. It is not latency-certified, uncontended or production routing.
- decision Post-LM final LM
Negative architecture decision for the local language model
No tested configuration met the frozen behavioural viability requirements, so the project decided not to start another full training run. It frames this as 'a negative feasibility decision, not an assertion that every possible local model or adapter is incapable.' Hosted cognition stays in production.
- training Post-LM stronger-local study
Qwen3.5-9B adapters fail the transfer gate
A 16-update adapter reached 11/1/0 exact and 8/0/0 anchors but 24/7/9 on transfer. A 48-update continuation reached 32/1/7 on transfer, still above the limit of 4 failures. Exact fit improved, and generalisation did not follow.
- performance Post-LM stronger-local study
Speech recognition at int8_float16 frees about 1.9 GiB of GPU memory
On a matched public-synthetic screen, the local recogniser at int8_float16 kept accuracy on the frozen cases while using 2,209,906,688 bytes instead of 4,256,743,424. The screen used synthetic speech, not the owner's acoustics.
- training Post-LM final LM
Final architecture candidate fails its development gates
Architecture001 trained 125 updates on corpus edition008. Checkpoint 96 was selected by a rule declared in advance and passed 8 of 26 validation and 5 of 25 adversarial cases. It proposed approval while authority was withheld and accepted an expired receipt. The SEALED set stayed unopened.
- infrastructure Post-LM
Audible first-audio instrument bound to the local voice
The reflex first-audio probe gained a mode that rejects any fallback voice and binds the exact voice package. The project describes it as an implemented and tested instrument, not an audible qualification. The measurement is reserved for the owner-present sitting.
- training Post-LM final LM
Recovery training runs on reviewed, frozen corpora; none qualified
Watched recovery runs trained against reviewed corpus editions, with package qualification blocked whenever semantic review was missing. Recovery001 to Recovery004 trained and reloaded, and none was development-qualified.
- capability Z08.P14
Paint, Notepad and Calculator inspected and controlled with verified outcomes
A bounded proof on a real loopback server verified semantic UI Automation inspection of three native apps, same-state restore for Paint and Calculator, and a verified Calculator close. The record labels itself a partial development proof, not RC-certified.
UI Automation-first interactionApplication-state understanding with fresh identity Demo
- docs Post-LM A-rows
All 44 post-LM body rows have integrated code and cited evidence
A reconciliation matrix shows every A-row with integrated code and evidence paths: 36 IMPLEMENTED, 7 IMPLEMENTED_PRODUCTION_REACHABLE and 1 IMPLEMENTED_INTEGRATED. The matrix states that it 'does not classify all rows LIVE or formally verified'.
- training Post-LM final LM
Model-visible contract (CP-7) and final corpus frozen; first final candidate fails validation
The contract binds 134 commands, their schema projections, the persona prompt, and the tokenizer and template across 373 source files by hash, and a reviewed final corpus was frozen. Candidate final-cp7-001 trained and reloaded, then failed validation (6 of 26).
- capability Z08.P14
Calculator operated end to end, ten steps independently verified
An isolated real core launched a new Calculator window, then maximised, restored, inspected and queried it semantically (249.1 ms lookup), invoked a control, set an exact rectangle, closed it and shut down. Every step was verified. This is one application and one run.
UI Automation-first interactionPostcondition verification Demo
- safety Post-LM A06
E-stop settles a GPU render's owned processes in 374.68 ms
During a local image-generation render, an E-stop was acknowledged in 19.4 ms, and all three owned processes had settled by 374.68 ms with zero survivors. No artifact was published. This was a single run.
E-stop (hotkey/HUD/voice)Local image generation and viewing as a governed mission Demo Journal entry
- capability Z08.P14.72
Fabrication software, durable modes, document intelligence and gesture personalisation merged
Wave three integrated A01 fabrication software, A22 durable owner modes, A43 local document intelligence and A31 gesture personalisation (recognition only). Defects found at integration, including a printer dial from a test and a document-chunk privacy leak, were fixed with tests.
Local document intelligenceFabrication workshop: installed slicer, printer staging and monitoring
- capability Post-LM A44
Missions can be simulated before they run
POST /missions/simulate walks every step without effects and labels its output SIMULATED / not_proof. A read-only probe of a running core the same day confirmed that label. 'Simulation never grants authority or certifies effects.'
- capability Post-LM
The conversation loop assembles with the local voice
In owner-authorised runtime testing, the conversation loop's check mode assembled with the local C013 voice, and the launcher's check reported the core and HUD answering. The voice's technical qualification is still pending.
- safety Z08.P14
Nine capabilities recorded LIVE under the 24-hour rule
Bounded health-route probes against booted processes recorded nine capabilities LIVE between 22:33 and 22:35 UTC: runtime infrastructure, fail-closed policy, the E-stop, the prompt-injection detector and unit-checked engineering maths. Under the project's own rule these readings are historical after 24 hours.
Async event busCorrelation IDs across every surfaceHealth & readiness probesLoopback-only binding assertionPolicy engine (fail-closed)E-stop (hotkey/HUD/voice)Prompt-injection detectorUnits & dimensional analysis with unit-error rejection Journal entry
- infrastructure Z08.P14
Tier 1 mutation testing: 449 of 611 mutants killed
All 44 Tier 1 modules were measured, and 73.5% of mutants were killed. That is below the harness's own 90% campaign target, and the project reports it that way.
- infrastructure Z08.P14
Full suite at bb54722e: 23,984 tests collected
The run recorded 23,588 passed, 277 failed, 34 errors and 85 skipped. The failures were partitioned rather than hidden: 5 worktree-only families were identified, 85 genuine failures were repaired in 12 commits, and 8 were left open with reasons. No full suite has been recorded at the current head.
- docs Post-LM closure
The interaction capability model: what works end to end on the owner's machine
Each interaction category was classified PROVEN_LIVE, IMPLEMENTED_UNPROVEN, BROKEN, BLOCKED or NOT_IMPLEMENTED, with the desktop, browser, files, research, missions and HUD driven live. 'Open Notepad' verified in 2.3 s, and a four-step mission completed in 10.8 s.
Open/close appsRequest→durable missionPostcondition verification Demo
- capability Post-LM 2.61
The owner-selected speaking voice runs locally
The speaking voice moved to a locally run, open-source text-to-speech model with no fine-tuning, bound to a hash-checked package that fails closed on mismatch. Any fallback voice is announced. Its technical qualification is pending, and the site publishes no voice samples.
- hud Post-LM 2.50 to 2.52
A three.js spatial layer inside the ONE HUD; journeys 9, 14 and 16 LIVE
The pre-spatial baseline of the ONE HUD was measured first, then three.js was added as its spatial layer. The mouse moves none of it. Acceptance journeys 9 (shared selection), 14 (project-scoped reminders) and 16 (rollback) were recorded LIVE.
- capability Post-LM 2.45
Research produces a cited artifact that opens correctly
Journey J5 went LIVE. Research from two named sources produced a Markdown report and an HTML copy, read back by digest, with six sourced findings and one assumption marked NOT ESTABLISHED. A live defect, summarising page openings instead of the question, was repaired.
Research reports with confidence and open questionsSupporting AND contradicting evidence collection Demo
- capability Post-LM 2.47
Objectives, recovery, memory correction and injection resistance hold across a real restart
Journeys 2, 3, 4 and 12 went LIVE across a real stop and relaunch. A corrected fact was recalled after the restart, and forgetting yielded 'I don't know' after a repair. Injected text in a test's output did not grant authority.
Memory consolidation, supersession and forgettingPrompt-injection detectorSurvive app restart Demo
- capability Post-LM 2.28
The engineering live drive restores unit-checked maths to LIVE
A spoken-style engineering question, delivered as a transcript, was answered with exact, calculated-not-measured pressures, and the engineering widget opened. The drive restored CAP-ENG-01 to LIVE in the registry.
- performance Post-LM 1.2 and 1.3
Median conversation turn cut from 3,083 ms to 1,270 ms
The spoken acknowledgement stopped blocking the provider call, and hosted turns now carry only task-relevant tool schemas instead of all 101. On 21 fixed typed journeys, first answer audio p50 fell from 2,749 ms to 1,107 ms.
- decision Post-LM 1.1
Post-LM gate opens; the first sealed local-model qualification fails
The Post-LM gate opened with the Latency Lab and a typed inlet. The first sealed local-model qualification was recorded as a FAIL on hard axes, and hosted cognition stayed in production. The body work went ahead without waiting on the model.
- hud Z08.P14.256
ONE HUD, slice A
One page at '/', with the intelligence core at the centre, identity above it and four permanent indicators. Workspaces became layout contexts inside the page instead of windows. Nothing renders without a source and an observed instant.
Truthful binding to canonical sourcesZero synthetic/placeholder data Journal entry
August 2026 2 changes
- safety Z11
Z11 safe local live proof
The real core and HUD were booted on isolated loopback ports. Unauthenticated gesture mutations were refused with 403, an E-stop blocked the gesture pipeline, and the HUD rendered that truthfully. No camera was opened and no window was moved.
Truthful binding to canonical sourcesLow-confidence safe failure Demo Journal entry
- decision Z08.P03
LIVE must be reached by asking the process
The evidence rule was tightened so that a LIVE state must come from asking a booted process, not from proximity. The registry records 108 demotions from LIVE stamped 2026-08-06T04:59:07Z, and the recorded LIVE count fell from 145 to 18 that day.