# ETA 6498-1 reconstruction — verification

The final reference was opened in Chrome, with all 49 groups visible to inspection and the independent clock paused. The original `watchmaking-bench:v1` save was compared before and after opening: **2,499 bytes, byte-for-byte unchanged**. The Codex browser-panel opening was also requested; the app returned a queued panel status. Actual Chrome navigation and the 43/43 source inventory were confirmed.

13 November 2026. This is the **best documented reconstruction** approved by the user after public-source research. It is not a fully verified factory digital twin. The distinction applies to every result below.

## Part-by-part result

The [recorded actual-browser audit](part-audit.json) covers **all 43 positions** in ETA's May 2025 spare-parts table, represented by **49 selectable component roots**. Each row retains the source identity, expected/represented quantity, current finite vertex-envelope check and explicit geometry unknowns. All 43 inventory checks pass; **exact part geometry remains unverified**. Aggregate 28 is checked through all five children. Assembled plate 1 is allocated to its body plus separately indexed bottom shock assembly and two dial fasteners. Supplied barrel/balance/shock/bridge groups contain further constituents;49 is a reconstruction allocation, not ETA's elementary-piece total.

The physical inventory includes 13 separate screws with the source allocation,17 unique ruby elements, all source-listed keyless and winding members, complete motion/going trains, pallet/balance assemblies and the five balance-bridge subparts. Every source row was selected through the actual browser UI. The geometry notes provide the corresponding part-by-part construction and remaining evidence gaps.

## Independent numerical/construction checks

`node --test tests/*.test.mjs`: **110/110 pass**. This retains all 78 previous checks and adds32 for the new reference:

| File | Tests | Evidence |
|---|---:|---|
| eta-reference-data.test.mjs | 8 |43 positions, source dates/variants, repeat screw quantities, parent accounting, invalid/duplicate/missing inventory, independent tooth-ratio periods/signs |
| eta-reference-geometry.test.mjs | 12 |49 real Three.js roots, 17ruby identities, finite geometry, authoritative part identities, specimen tooth counts/major envelopes, source interfaces, full support graph, internal/root separation, bearing/journal spans, barrel cutaway, cap bores, stem channel and external mesh core/face checks |
| eta-reference-state.test.mjs | 8 |Paused restoration, malformed input, retention, assembly validity, independent winding/setting, non-hacking playback, stored-input handling guard, elapsed-time behaviour |
| eta-reference-handling.test.mjs | 4 |Dial-side negative-Z extraction, outside-footprint travel, reverse installation, stem axial withdrawal and bridge-side lift |

New checks were exercised red before implementation or a corresponding repair. Tooth counts are firsthand specimen evidence; the common nominal modules and most arbor coordinates are reconstructed. Nine parallel external pairs have matching declared pitch-centre relationships, overlapping axial faces and no tooth-root core crossings. Unrelated parallel wheel cores are checked as well. This is not a swept tooth/stone contact certification. Two perpendicular winding/setting interfaces remain prescribed and explicitly unvalidated.

The model matches the stated37.20mm overall flange,36.60mm case-fitting interface,10.25mm small-seconds offset and30mm raw stem in the chosen coordinate system. Its reconstructed body datum is0mm and highest shock-setting face approximately4.495 mm within the published4.50mm envelope; hand-fitting pipes extend below that body. These nominal checks do not validate all internal depths or manufacturer tolerances.

## Real Chrome journeys

`eta-reference-browser.js`: **70 assertions passed**. All tests run in an isolated browser context and use actual controls with read-only inspection:

- All 43 source rows select corresponding real roots; repeated instances are separately accessible.
- Every row has passing quantity/render checks while retaining unverified geometry.
- Crown-wheel screw 903 retains its documented left-hand identity.
- Dial view places the stem at 3 and small seconds at 6; isolation affects presentation only.
- One beat advances 0.2s; setting leaves the non-hacking going train running; reverse input/click and separate winding/setting angles behave as described.
- Represented input blocks opening until release. Parts first withdraw from their mounting side. All 48 removable groups strip and rebuild through the complete dependency graph, at scale 1 in the physical tray.
- A full partial teardown restores paused; undo/redo restores assembly changes.
- Downloaded JSON is parsed and checked for all 43 rows,49 actual envelopes and per-row unresolved geometry.
- Sources dialog/Escape, independent reset/storage and 390px operation pass without browser errors or shader warnings.

All **177 previous browser assertions** also passed, for **247 total regression assertions**. Adding the reference link initially exposed a phone-header overflow; the navigation now wraps and the full affected original journey passed again. The original bench, Learn course and Design lab remain operational.

The per-part vertex bounds use precise transformed geometry. The initial fast Box3 envelope expanded rotated circular wheels unnecessarily and falsely suggested a tray overhang; the actual-vertex check corrected that classification. Neither kind of AABB is a solid collision solver.

## Visual/quality/performance evidence

Actual bridge, overhead, dial, exposed, exploded and stripped views were captured and visually inspected. Source diagrams and external interface drawings were rendered locally for comparison. The independent reviewer found and rechecked three important repairs: pierced barrel caps around the arbor; corrected setting/motion wheel centres, axial layers and angular ratios; and mounting-side/stem-specific handling routes. A historical-source traceability mismatch was corrected by exposing CT 2020 as the silhouette source while retaining CT 2025 as BOM authority. The final review found no remaining important actionable issue within this declared scope.

`eta-reference-quality.js`: bridge/overhead/dial views at 1600×1050, the sources dialog and 390 × 844 mobile all report **zero axe WCAG 2 A/AA and 2.1 AA violations, no page horizontal overflow, no console/page/shader errors and no external runtime requests**. Original mode, Learn and Design lab quality checks were repeated with the same clean outcomes. Automated checks are not full assistive-technology certification or field Core Web Vitals.

Final local unthrottled Chrome samples, approximately 1.2 s each: exposed motion **54.7 scene updates/s**, exploded view **54.6/s**. The assembled render reports364 calls,493,130 triangles and181 geometries including the environment. Rendering is capped at 60 Hz and DPR 1.5. These samples do not guarantee every laptop/browser's performance.

## Offline Blender inspection

Actual final browser meshes were exported through a one-shot loopback development receiver. `tools/inspect-eta-blender.py` ran Blender 5.2.1 LTS in a new background factory session, converted mm to metres and saved `eta6498-inspection.blend`, `blender-assembly.png` and `blender-inspection.json`. No existing Blender window/scene was modified.

The final snapshot contains **49 source groups,363 mesh objects,1,446,766 vertices and 493,118 faces**. After a report-only 1 nm seam weld, 174 boundary/nonmanifold edges remain. These are render meshes with open seams, not watertight manufacturing solids. The full bounds include the raw stem and hand-fitting extensions: X −38.55..18.60,Y −18.60..18.60,Z −1.41..4.495 mm. Render, body-interface and topology checks establish no material properties, actual fit or contact forces. A sandbox-denied global thumbnail-cache write did not prevent saving the local scene, report or render.

## Remaining accuracy limits

No matching complete factory CAD/tolerance dataset or traceable measured specimen was obtained. Most arbor coordinates, all unspecified recesses/depths, working tooth profiles, threads, bearing fits, springs/preload, regulator geometry and exact production variant remain reconstructed. Continuous escape-wheel rotation, pallet/balance motion and crown coupling are prescribed; no full escapement lock/impulse/force solver is claimed. Physical reserve, friction, amplitude, wear, shock resistance, manufacturing interchangeability and real watch rate cannot be certified from this model. See [sources](source-audit.md), [geometry](geometry-notes.md) and [research](research-log.md).