docs(perception): require visual LAB evidence

docs/13_LIDAR_WORKER_PRODUCT_AND_ROADMAP.md

@@ -641,6 +641,10 @@ changes.
   solid-state-LiDAR result is not independently reproducible, KITTI is
   cross-domain, and `RAVNOVES01` without independent cuboids cannot establish
   detector accuracy.
+- [x] Publish a bounded visual-audit derivative of the same sealed transfer
+  run: `18` deterministic TP/FP/FN/class-coverage frames, original LiDAR
+  samples, KITTI truth, model predictions and shared `3D ↔ BEV` inspection in
+  Mission Core without repeating inference.
 - [ ] Freeze a separate training/adaptation admission before downloading a
   trainable checkpoint or starting a training runtime.
 - [ ] Train only on the frozen `3,712` KITTI train frames; keep all `3,769`
@@ -709,6 +713,10 @@ an uncovered volume.
 
 The exact contract, metrics, rejection decision and reproduction notes are in
 `experiments/perception/L3_POINTPILLARS_TRANSFER_REPORT_2026-07-31.md`.
+The current visual-audit derivative is
+`l3-pointpillars-visual-audit-35a5bfef788e40080cac1b8ca82ba65176f158326fa2fc4570b522c97d6a2f8d`.
+It is an append-only read-only projection of the same frame-result identity,
+not a new detector result or a change to the measured gate.
 
 ### L4 — live shadow integration
 

docs/15_LABORATORY_RUN_CANON.md

@@ -140,6 +140,13 @@ Each LAB result must make these sections available to the operator:
 5. **Conclusion** — what was proved, what was not proved, limitations and the
    next gate.
 
+If an algorithmic result has a spatial, image-plane, temporal, geometric or
+otherwise directly visualizable interpretation, metrics alone are not a
+complete LAB record. The published run must include bounded visual evidence
+derived from the same immutable input and result identities. That evidence
+must expose the operator question and the relevant truth/prediction/error
+layers without recomputing, retuning or substituting the measured result.
+
 The result must explicitly state:
 
 - ground-truth availability and its identity;

experiments/perception/L3_POINTPILLARS_TRANSFER_REPORT_2026-07-31.md

@@ -175,6 +175,37 @@ The compact report SHA-256 is:
 
 `f53d8d03a7c1e092d87b732872b26fea16447f373cf567bfe57737cb3ccbebba`.
 
+## Visual audit
+
+The numerical result now has a separate append-only visual derivative:
+
+`l3-pointpillars-visual-audit-35a5bfef788e40080cac1b8ca82ba65176f158326fa2fc4570b522c97d6a2f8d`.
+
+The derivative revalidates the complete source manifest and the unchanged
+frame-result identity before reproducing the same global score-descending
+oriented-3D-IoU matching. It then selects `18` deterministic evidence frames:
+all three frames containing a true positive, the highest false-positive and
+false-negative frames, and target-class coverage.
+
+Each selected frame contains:
+
+- an evenly bounded sample of at most `12,000` original XYZI points inside the
+  shared evaluation volume;
+- independent KITTI truth cuboids marked as matched or false negative;
+- evaluated PointPillars cuboids marked as true or false positive;
+- the original inference latency and complete per-frame accounting.
+
+Mission Core reads only the catalog and one selected frame at a time. The
+operator can inspect the same case in `3D` and `BEV`, move sequentially through
+the evidence set and expand/restore the shared viewer. No inference, threshold
+retuning or metric recomputation occurs in the browser. The original KITTI
+archives remain only on Worker 006; the compact visual derivative is about
+`15 MiB` on the operator host.
+
+The visual layer makes the measured domain mismatch inspectable. It does not
+convert the cross-domain probe into K1 accuracy and does not authorize
+navigation or safety.
+
 ## Interpretation
 
 The worker seam, TensorRT engine, Triton serving path, bounded postprocess and