JARVISLabs

A governed STEP union through FreeCAD, verified across a restart

On 22 September, on the unmerged Advanced Systems branch, a real JARVIS core discovered the STEP commands, ran a FreeCAD union of two 1,000 mm³ boxes in about 0.39 s, published and registered a 1,500 mm³ result, kept its lineage across a restart, and refused changed output bytes.

INTEGRATED What this run supports. Recorded 22 September 2026.

Footage pending

This run is documented from its recorded evidence. Footage has not been published: capture waits for the machine to be free of local model measurements, and every clip is reviewed for private information first.

Asked

What was asked

Combine two STEP parts into one, keep a record of where the result came from, and never accept a file that has changed since.

Result

What happened

A diagnostic core booted and selected two complete STEP schemas from a 144-command catalogue. engineering.step.derive ran a fixed FreeCAD union of two overlapping 1,000 mm³ boxes, then published and registered the 1,500 mm³ result with source lineage. Verification came from a separate read of the bytes and the registration, not from the dispatch response. After a second boot the lineage was intact, and changed output bytes failed verification.

Pipeline

Architecture path

  1. Command discovery
  2. engineering.step.derive
  3. Core auth and E-stop generation
  4. FreeCAD fixed script
  5. PREPARED
  6. CALCULATED
  7. FILE_VERIFIED (governed file write)
  8. REGISTERED (revision CAS)
  9. Independent read
  10. Restart
  11. Changed-bytes refusal

What this proves

  • On the Advanced Systems branch, a STEP boolean runs through a fixed FreeCAD script and is verified by independent observation.
  • The derivation's lineage survives a core restart.
  • A derived file whose bytes changed afterwards is refused.
  • The result is recorded as a CALCULATION, with input and output digests and reloaded volumes.

What it does not prove

  • This is branch work that is not merged into production.
  • The inputs were synthetic boxes; this is not a real project part.
  • A geometry calculation is not a physical measurement, a strength claim or printability.
  • It is not a general CAD feature engine or natural-language CAD.

Context

The Advanced Systems branch gives JARVIS governed CAD operations through FreeCAD. They run only as fixed, isolated scripts (“No caller-supplied Python is executed”), and every result is journalled through four phases: PREPARED, CALCULATED, FILE_VERIFIED, REGISTERED. A project revision is updated by compare-and-set with an E-stop commit fence, so a newer revision is never overwritten, and a trip between calculation and publication is detected.

The run

The proof booted a real diagnostic JARVIS core twice on the branch.

  1. Discovery. Bounded capability discovery selected two complete STEP schemas from a 144-command catalogue.
  2. Derive. engineering.step.derive authenticated, bounded the request and carried the admission E-stop generation into the worker. FreeCAD computed the union of two overlapping 1,000 mm³ boxes in about 0.39 s.
  3. Publish and register. The derived STEP was written through the governed file owner and registered with its source lineage.
  4. Verify. Success required a separate read of the file bytes and an exact canonical registration. The command’s own dispatch response could not establish success.
  5. Restart. On the second boot the lineage was still present.
  6. Refuse. When the output bytes were changed, verification refused.

A companion proof exposed the calculation receipt: two 1,000 mm³ inputs, a 1,500 mm³ union with 15 × 10 × 10 mm bounds, input and output digests, reloaded volumes, the kernel version and evidence_class: CALCULATION. The derived shape was displayed in the real HUD scene, and a partial geometry restore returned the original shape while keeping a later placement.

The evidence

The branch’s running report (docs/ledger/ADVANCED_SYSTEMS_FINAL_REPORT.md) and its live proof records describe both runs. The inputs were deliberately simple so the result can be checked by hand. A 1,500 mm³ union that exactly fills 15 × 10 × 10 mm bounds is what two 1,000 mm³ cubes overlapping by half produce. No footage exists yet.

The limits

This is branch work and not merged into production, so it is shown here as INTEGRATED. The fixtures were synthetic. The report lists what remains open: scene promotion, arbitrary geometry features, a geometry rollback workflow, fabrication readiness, and natural-language interaction with the final model. A calculated volume is only geometry. It says nothing about mass, strength or whether the part can be printed.