The spatial engineering work that opened on 22 September could already load a STEP part into the HUD and run governed CAD operations. The next two days were about something harder: making engineering answers carry enough context to be trusted. Each of the features below was proven by booting a real JARVIS core, and most by restarting it and checking that the answer survived.
What changed
A what-if engine. Ask “what if we make this 15% thinner?” and JARVIS freezes the current revision, makes an exact clone as the baseline, and derives the changed design along one axis. It never promotes the candidate to current. It then recomputes only the calculations the change affects. On a 40 × 20 × 10 mm test part, −15% gave a thickness of 8.5 mm and a mass of 8.432 g, and “halfway between those two” produced the midpoint of two verified changes. The proof ran 213 checks against a real CAD kernel across two boots, and all 213 passed.
A contradiction review. A read-only watcher looks for two kinds of disagreement: a design that breaks a declared limit (“6 mm over the envelope limit”), and an analysis that belongs to a different revision from the one on screen (“the latest volume result belongs to revision 1, but the HUD is showing revision 2”). It stayed quiet when nothing changed and kept identical findings across a restart.
Calculated is not measured, now enforced. Every engineering number is either CALCULATED, with its method and inputs, or MEASURED, with instrument, method, observer, time and uncertainty. A run without that evidence is MEASUREMENT_UNVERIFIED, and the “measured” label is refused. The HUD draws the two classes differently. This is the same distinction this website makes, now held by JARVIS itself.
Honest compromises. “Show me a reasonable compromise” considers only designs that no other design beats on every objective, normalises each objective over those candidates, and picks the one nearest the ideal under weights it states. In the proof the balanced design scored 0.3877, against 0.7071 for either extreme. Every score is shown, and a tie is reported as a tie, not broken by list order.
Print readiness, proven. Before anything could reach a printer, JARVIS now answers READY, NOT READY or WHY from checks on the mesh, real dimensions, bed fit, orientation, supports, the slice itself, the exact profiles, and whether the job still matches the current revision. The canonical proof booted a core three times with a real slicer. It answered in 0.125 s, returned identical receipts after a restart, called a job NOT VALID after the CAD revision changed, and reported NOT READY when the real slicer rejected a part. No printer was contacted at any point.
Orientation that is real. A part rotated only on screen can no longer be sent to the slicer. JARVIS answers that the display is transformed and names the step that bakes the rotation into real geometry. Laid flat, the test plate had 2,349.9 mm² of bed contact against 600 mm² standing, over 85 checks.
“Is this the latest model?” JARVIS answers only from what it can verify now: “Revision 2 is the newest file I can verify. Revision 3 is referenced in the project notes, but I cannot currently locate its artifact.” It does not guess, and the answer survived a restart.
What failed along the way
Building and proving these features turned up real bugs:
- The first what-if run treated an exact clone as a changed design, because two bounding boxes differed by 2 × 10⁻⁷ mm and the comparison was exact. The tolerance is now stated.
- The evidence-class proof rejected a valid observation made in the same whole second as the run’s close.
- A scale factor applied and then undone left 0.9999999999999999 for 36 of 299 factors, and those parts could never be sliced.
- The 3D scene route created its emergency-stop fence but never entered it, so model bytes could still leave after a stop.
Each one is written up in the failure and recovery lab.
What was verified, and what was not
On the branch, 37 requirement rows are now complete (36 atomic capabilities and 1 source requirement), up from 1 on the afternoon of 22 September; 20 of them are the print-readiness chain. That is still only 1 complete source requirement out of 417, with 44 partly built. The orientation check proves the chosen up direction was kept, not that it is a good way to print. Readiness is software readiness only. None of this is merged into production, and none of it has run on the owner’s own projects.
What remains
Continuous scene refresh and conversational control of the 3D workspace. A simulated evidence class for solver results. Assemblies of separately placed parts in the contradiction review. And the merge: every one of these capabilities still has to reach the production branch and pass a live proof there.