On 22 September 2026 the project ran a different kind of campaign. Six branches started from the production head 893913a9c: one canonical integration branch (build/jarvis-advanced-systems-20260922) and five lanes covering cognition, computer operation and three executors. By the end of the day each branch held roughly 250 commits past that base, 192 of them shared. The canonical branch had reached f0738a5b9.
This entry describes what that work added and what it proves. It is all Advanced Systems branch work. None of it is merged into production, and the site labels it that way everywhere it appears.
Why a separate branch
Production was in a model-first phase: its critical path waited on a qualified local language model. The body work that did not depend on the model (geometry, browser grounding, memory projection, specialist checkpoints) could advance in parallel, but it could not be allowed to disturb the production checkout or the model-visible contract. The branch report opens with its own disclaimer: “This is a running report, not a completion declaration.”
A revision-bound 3D scene
The spatial engineering workspace is a SceneStore (commit 26e55cacd). It holds at most 64 objects, 64 KiB and 4,096 revisions per scene. It supports compare, labels and selective restore, imports STL, OBJ and 3MF, and displays STEP. It renders through three.js inside the ONE HUD, not in a separate application.
The scene draws a firm line: “Object transforms describe display placement, not CAD edits, fabrication dimensions, or observations of physical objects.” Rotating a part on screen changes nothing about the part.
“Load STEP designs into the scene” is the only row of 857 in the Advanced Systems matrix marked COMPLETE.
A spin preview shows how the project handles something as small as an animation. The command spatial.scene.spin performs a 360-degree display rotation and returns to the saved pose. It verifies only when a matching renderer receipt reports an observed intermediate frame, so admission alone is not success. The renderer must renew a short E-stop lease, at most 125 ms, while it animates. After a repair, the measured stop from a diagnostic trip took 31 ms in the foreground. The project labelled that figure “not worst-case 150 ms qualification”.
Governed FreeCAD STEP operations
The CAD work lives in jarvis/engineering/freecad.py and runs only fixed, isolated scripts: “No caller-supplied Python is executed.” By the end of the day it covered booleans, centre and plane split, component extraction, transform and mirror, edges and fillet, offset, face listing and extension, and a sealed hollow.
Every derivation passes through four journalled phases: PREPARED, CALCULATED, FILE_VERIFIED, REGISTERED. Each result carries evidence_class: CALCULATION. Project revisions use compare-and-set with E-stop commit fences, so a trip and re-arm between calculation and file publication is detected. Interrupted work is not retried automatically: “Incomplete attempts retain their last durable phase for explicit reconciliation; they are never replayed blindly.”
The fixtures were deliberately simple, so the answers could be checked by hand. Two overlapping 1,000 mm³ boxes gave a 1,500 mm³ union and 500 mm³ difference and intersection, and all three reloaded correctly. A 1,000 mm³ cube split into two 500 mm³ halves. Hollowing a 20 mm sphere with a 1 mm wall gave a cavity volume that matched the closed-form 4/3·π·9³ to floating-point precision. A non-uniform transform of a curved solid was refused, because the calculated and reloaded volumes (21,948.37 and 21,945.46 mm³) disagreed. No tolerance was relaxed to let it through.
There is an open contradiction here. Production’s own solver survey found FreeCAD absent from its PATH, while these proofs used an installed FreeCAD 1.1. It is probably installed or located differently now, but that is uncertain.
From STEP to slicer, and a door that stays shut
The installed OrcaSlicer 2.4.2 rejected STEP input. The branch kept that failure as a receipt and added FreeCAD tessellation into a verified closed mesh (commit f628d3f3d). A pinned synthetic 10 mm cube then sliced successfully, with a slicer estimate of 325 s and 0.73 g. Those figures are estimates, not a print.
Print readiness returns READY, NOT READY or WHY across nine checks. Its rule is written in capitals in the source: “READY IS SOFTWARE READINESS AND NOTHING MORE.” A governed handoff can only ask: “PREPARING A HANDOFF IS ASKING, NOT DOING.” The physical start requires a fresh single-use approval bound to the G-code hash, a responsible adult, a clear E-stop, re-validated G-code, a proven stop path and no print already running. Today the final send always refuses. No printer was contacted.
Browser grounding across frames and shadow roots
Production has no iframe or shadow-DOM awareness. The operator lane added a DOM scene reader that records relationships across frames and shadow roots, and a target check (dom-target-check/1) that refuses a target that moved, was replaced, was covered or changed. Two rules keep observation separate from authority: “CSS viewport rectangles are observation data, never native screen leases” and “CURRENT/STALE is an observation comparison, never click authority.”
In one recorded run on the operator branch, 500 read-only target reads in an isolated Chromium completed in 2.81 s and verified across a restart. A same-label replacement was then refused as target_missing_or_replaced. Typing and filling on this path are still pending validation.
The same lane fixed UI Automation geometry under display scaling. A disposable window moved across four displays with physical-pixel readback, and the E-stop refused a later move. The lane’s combined acceptance run passed 545 tests across 43 files.
What else landed
- Bounded capability discovery. A read-only body interface lets JARVIS ask what it can do right now. The interface “has no execution or grant method”, so asking changes nothing.
- An Obsidian projection of project memory: “a human-readable projection, never a second memory authority”. Human edits count as conflicts.
- A Failure Atlas of past failures, whose patterns are marked as inferred from reports: “No reported root cause is silently upgraded into verified causation.”
- Specialist checkpoints in the mission store, where “scope is a ceiling, never execution authority.”
- A reflective controller that revises plans from receipts: “A dispatch that did not verify stays unverified.”
How to read this
The branch’s stacked labels (“IMPLEMENTED; TESTED; INTEGRATED; LIVE-PROVEN”) are local to each lane. On this site the Advanced capabilities are shown as INTEGRATED: wired and tested on the branch, not reachable in production. The live proofs are real runs of real tools on isolated cores. The inputs were synthetic fixtures and nothing touched owner hardware. The next step is canonical integration into production, and that has not happened.