# What changed, and why

All work belongs to the new **Precision PCBA · datum-pin fixture loading** project. The earlier camera-module project is preserved.

## Research and geometry

The operation comes from the documented BTL / Test-it-off PCBA test-cell workflow. Original high-resolution industrial photographs, a Raspberry Pi manufacturer photograph, the mechanical drawing and the original STEP archive are preserved. The workpiece, robot and fixture are an explicit adaptation; no Samsung production geometry was supplied.

| Backend scene | Change |
|---|---|
| Version 1 | Populated manufacturer PCB, raised input/output nests, round and relieved locating pins, open test fixture and a complete electronics laboratory. |
| Version 2 | Explicit thin solid collision geometry and updated room labels. The first simulation preparation exposed the platform’s GLB size limit. |
| Version 3 | Reduced the display geometry while keeping the original STEP and PCB hole geometry. The scene became small enough for the simulation backend. |
| Version 4 | Narrowed support posts to clear the gripper, matched both cooler-hole radii to the CAD, mapped observed board markings from the original photograph, and corrected connector-shell appearance. |

The simulator originally combined authored collision pieces before decomposition. That could fill precision holes. The updated path preserves already-convex authored pieces as a rigid compound and apportions mass by volume. Explicit thin solid sectors also survive scene-physics filtering. Thirteen scene-contract tests and three collision-helper tests passed. A separate prescribed-pose collision test confirms that the aligned pins clear the holes while a 1 mm offset or a 1° rotation interferes. This unit test is not robot-execution evidence.

## Engineering trials

Eleven scripted trials are recorded in `evidence/engineering-trials.json`, with individual results under `evidence/engineering-trials/`. These use the real Runtime controller and free physical objects, without attachment or teleportation.

Initial grasps were too low or interfered with populated components. Moving to an unobstructed PCB edge and narrowing the support posts provided finger clearance. One completed sequence still forced the board onto the pins. A calibrated 0.35 mm approach correction and earlier release reduced the scripted trial’s peak board–fixture contact to 7.16 N, with no robot–scene contact. That established the nominal commissioned procedure; it was not an LLM result.

## Native LLM evaluations

Acceptance limits were written before the first native evaluation. Each attempt uses GPT-6 Astra, actual camera checkpoints and paused inference timing. Its manifest identifies the exact frozen scene. Every original evaluation bundle is retained.

1. **`fd1a01dd96a247758b130517e50271c2`** — failed. The model omitted X/Y during vertical moves, preserving accumulated servo displacement. It subsequently noticed a failed retrieval and stopped. Replay rejected seating and contact forces.
2. **`2b98597a830741219df2c7b93abaface`** — failed. Explicit X/Y completed the sequence, but repeated yaw commands also reset pitch/roll and disturbed the grasp. The model reported completion; the audit rejected the tilted seat, excess contact and output error. This is why visual completion alone is insufficient.
3. The next procedure specifies absolute X/Y/Z on every move and sets yaw only when a turn is actually required. The final aligned-input procedure uses yaw only on the first approach. The selected run and its measured outcome are recorded in `FINAL_REPORT.json`; a finish message is never used as the acceptance test.

All evaluations are of a commissioned nominal layout. They do not establish robustness across randomized fixtures or manufacturing tolerances. Simulator forces are evaluator-only measurements; the policy has no force-feedback sensor.

First passing evaluation on scene version 4: `fa620d0834294454b2d0ad5d916670dc`. Independent replay matched all 551 states exactly. Maximum seated hole-centre error 0.1345 mm; valid released dwell 5.564 s; peak board–fixture contact 8.4895 N; robot–scene contact 0 N; final output centre error 0.8009 mm. Five model calls, 27.5 simulated seconds.

Version 5 restores complete CAD material groups below 200,000 faces so display reduction does not remove small chip housings, and corrects the upper USB shell colour. Version 6 replaces major component envelopes with original CAD component bounds plus a 0.05 mm per-side allowance, adding the ribbon, radio and fan/POE housings. Mounting holes and fixture geometry remain unchanged. The final selected result is in FINAL_REPORT.json.

The CAD-bounded scene exposed HDMI and fan-connector clearance issues while turning the board. Additional edge-grasp and rotation trials are retained. Version 7 uses an aligned input nest, consistent with a fixture-fed production setup, and focuses this demo on precision pin engagement, released seating, retrieval and unloading. The rotated-start failures are not presented as solved; arbitrary starting orientations remain outside the final nominal demonstration.

## Final selected evaluation — scene version 7

`8396350ebf5a42daa8ed6a2838220e38` passes all frozen acceptance clauses. Its 549 recorded states and timestamps match independent command replay exactly. Maximum seated hole-centre error is 0.1440 mm; valid released dwell 5.560 s; peak board–fixture contact 8.6529 N; robot–scene contact 0 N; final output centre error 0.6522 mm. Five model calls, 27.4 simulated seconds and 65.39 evaluation wall seconds. Physics pauses during inference. Recorded gripper contacts involve only PCB edge sectors, with no modeled component contact.

The six native evaluations, in order:

| Run | Scene | Audit | Reason or change |
|---|---|---|---|
| fd1a01dd96a247758b130517e50271c2 | V4 | Rejected | Incomplete coordinates caused accumulated displacement; failed retrieval. |
| 2b98597a830741219df2c7b93abaface | V4 | Rejected | Repeated orientation reset disturbed the grasp and produced a tilted seat. |
| fa620d0834294454b2d0ad5d916670dc | V4 | Passed | Corrected absolute-coordinate and orientation procedure. |
| c06a8070c5814527894c7d67954fbbbe | V5 | Passed | Restored display details; same recorded physical outcome as V4. |
| f48a47d804ef428f95c37dad5aea8a4c | V6 | Rejected | CAD-bounded component collisions exposed interference during rotation. |
| 8396350ebf5a42daa8ed6a2838220e38 | V7 | Passed | Aligned input nest; complete CAD collision bounds and edge grasp. |

These are development iterations with different scenes and guidance, not trials from a fixed statistical benchmark. The final demonstration proves one commissioned nominal arrangement. Source photographs are retained alongside the additional CAD, drawing, engineering brief, explicit policy guidance and existing robot/room assets.