v0.2.0-reconciliation: PaDiM->PatchCore card with measured OpenVINO parity

Reconciles model card with Phase A/B/Bx measured numbers from this session:
- Full PaDiM vs PatchCore ablation across 4 categories (bottle, cable, capsule, leather)
- Bottle/leather as mean +/- pstdev over 3 seeds; cable/capsule single-seed (legacy)
- Cable coreset sensitivity sweep (0.01, 0.10, 0.25)
- Latency tables on AMD Ryzen 9 9900X CPU and RTX 5070 CUDA
- Fresh OpenVINO parity measured on PatchCore (N=20 per category, all parity_clean,
max abs error <= 7.3e-05, zero label flips, zero mask pixel flips)
- License section makes MVTec AD CC-BY-NC-SA-4.0 propagation to checkpoints explicit
- model-index uses standard HF schema (no custom 'source' field); provenance lives
in PROVENANCE.md
- CHECKSUMS.sha256 covers all non-README shipped files

CHECKSUMS.sha256

@@ -0,0 +1,41 @@
+26ac21cef63b1a76c0001cfce3fa0d561128f4aff570efd14f6ca057d32d153b  ./artifact_index.json
+7c7c278d9e5339b0165a136aa216f313a81fa1a47de2ce599d9eeb5f8548b2b2  ./assets/release_visual.svg
+6a98a85a91505e9efe676a7b5905007e9cce9f18946a9ad693c25ca6d5500be7  ./claims_ledger.md
+38a0298f5db9f1a58eb10eaade5e792fb802bc756f8f69c0e8f17737f3b1247a  ./examples/prediction_classical_broken_large_000.json
+3e894b9074565231135675fef2e374cfbab6c14e06216b1007086852648cfa6e  ./examples/prediction_classical_good_000.json
+0d0a5fd5af0d6fdc83a3d074fe531a0bd94205ad9c89892c4c09cb363fe60019  ./examples/prediction_padim_broken_large_000.json
+97b84c75683cc310f796425fa897b616a6b9ce2b47a8b4cf78d600dcfeded410  ./examples/prediction_padim_good_000.json
+11ad7efa24975ee4b0c3c3a38ed18737f0658a5f75a0a96787b576a78a023361  ./.gitattributes
+3e3f3c19c25b8caca1f10ea5674e53107eb0e3e361edb805e25316db2c26e9f7  ./PROVENANCE.md
+6e08de6dca10ead6a284257fdf169557420d580a3eb0d6cffeb020b27f311a52  ./reports/anomalib_padim_export_smoke.json
+5947edb208b1be95de2480efa375ec49cf7d148dd7e404cd984e30c247119550  ./reports/anomalib_padim_export_status.json
+146b42cf2802e33145bdc5b8f81f86547307ad8d9d8e300e27aa9f727629773c  ./reports/anomalib_padim_mvtec_ad_bottle_result.json
+8d47b69d68f4bb4cdea67985e01ce10953decec318ccf2190f81feff707d9d97  ./reports/classical_patchdiff_rerun_mvtec_ad_bottle_result.json
+04dbdd95fb2c658c276b5c4756e46425576dfaef3e64073857fbfb24ded7bf3d  ./reports/dataset_provenance_mvtec_ad_bottle.json
+ab16dfee81105ce15cddeeb4a12caba5aaf66ae589358a54b2e0ca47a058cd9c  ./reports/eval_harness/openvino_parity_patchcore_bottle.json
+2e7d0d5109993ed89ebca2508cf9262ee4bf38d51fd11b900ad8dbb19b393b15  ./reports/eval_harness/openvino_parity_patchcore_cable.json
+1acd65a2919b34f5bd5092a69016e4b9a0ce6064d82d8129e2436716037dd362  ./reports/eval_harness/openvino_parity_patchcore_capsule.json
+c0edf784a0a063cc45bb0090cd04c5b27e131940ba3cbd90076d5c8ce9064f2c  ./reports/eval_harness/openvino_parity_patchcore_leather.json
+a0016103b61a49a800140a68b8bac681fb31087011d2d02cad0eec2530d88556  ./reports/eval_harness/padim_bottle_repro.json
+2cd097ecfc4dad28f4665f653a88b99eb5aa69a6a9709ccd360641afae651e51  ./reports/eval_harness/padim_cable_repro.json
+be3ab8385426000c676e733e62ca6924582bed5a3b6e340272183763a5c83e85  ./reports/eval_harness/padim_capsule_repro.json
+5fd852806f6845d4b9c116222946a38e7b96b0a793464c6184aada92d43c4e90  ./reports/eval_harness/padim_leather_repro.json
+ac2d1a1b4dd74a196c185a8eb712a3bf2d6bce7f89c677ac18b8d9106330b72c  ./reports/eval_harness/patchcore_bottle.json
+5d2e611c1d03e18fe57957a52072762e5a922d090594248c191458c8e0f5f038  ./reports/eval_harness/patchcore_bottle_latency.json
+5d66812e78f96e01851d0597a34432fd566e7dd02071106b1bfa1be74d435be3  ./reports/eval_harness/patchcore_bottle_seed1.json
+701d35869ab6180d0032d58f3fede888e240ec51530e665476f3f106772112bf  ./reports/eval_harness/patchcore_bottle_seed2.json
+dd766d34d0f3ec5f513beae5b9094c125f57b0c4cdacc26ff70ceb461ab669c9  ./reports/eval_harness/patchcore_cable_coreset010.json
+8b96bd14afe5c8ee2fcef3d156c7c4f89a4e4e3c974ada0ac38fd99b559d4ef7  ./reports/eval_harness/patchcore_cable_coreset025.json
+ac0b8d3866aef237ba8d528b94c24cb2e9542698e1fbc55c006b7707b35a9723  ./reports/eval_harness/patchcore_cable.json
+14dd824301b855111e800d5e5073969f718b7cc7755e13a2c2a0f0d0cd08c323  ./reports/eval_harness/patchcore_cable_latency.json
+b3ef3252c04a357335453613a77987e7a1c28fcdbfa59c63c5086eea45fdd2fc  ./reports/eval_harness/patchcore_capsule.json
+9e3efe8128480cc7044e1375378277276a39180ef80625dd5b84387ce85be467  ./reports/eval_harness/patchcore_capsule_latency.json
+03fbfa815ee9b89b141d5176f9fcec7835d43c2cfad38695edcab1235d901eb6  ./reports/eval_harness/patchcore_capsule_latency_rerun1.json
+92041d03aeec4561a3fabe4b619994550ff7044cfb750c62f261f688b9b33213  ./reports/eval_harness/patchcore_capsule_latency_rerun2.json
+6eaf764e10fd15963c8d437938acb1a2fd9b3f3e8675a0c444fbdad684ab46b2  ./reports/eval_harness/patchcore_leather.json
+e626656236b92a657941697e6be46809a4c35d691c06a69718823303f70a85fe  ./reports/eval_harness/patchcore_leather_latency.json
+90c0a4ae66ad30d893602ea6f019835d08b435fcdc659320d1e223ea3b05627f  ./reports/eval_harness/patchcore_leather_seed1.json
+2110d56b0d84974f17fc64ba04caa4d31198b0c02231591f691cb0bd8246dc5d  ./reports/eval_harness/patchcore_leather_seed2.json
+e8d0831b27f0618ae16463adbf69474c5d4bce29d009c04e1ffe7168c58b6f50  ./reports/openvino_parity_investigation.json
+00058a1bafc3e314e60a2a680007f6668764917166e9a560efa60fad0b0e525b  ./requirements-agent_b_verified.txt
+341003c1ed0f9c438705d3f574804f616f263e80b90924d97e9bf4ad7104c49a  ./requirements.txt

PROVENANCE.md

@@ -0,0 +1,131 @@
+# PROVENANCE: Metric -> Source JSON + Field
+
+Every number on the model card traces to one of the JSONs under
+`reports/eval_harness/`. This file is the source-of-truth map.
+
+## YAML model-index metrics
+
+### inspectnet-cx-patchcore-bottle
+
+| metric (model-index) | value     | source JSON(s)                                                                                                | field        | aggregation                |
+|----------------------|-----------|---------------------------------------------------------------------------------------------------------------|--------------|----------------------------|
+| image_auroc          | 1.0       | `patchcore_bottle.json`, `patchcore_bottle_seed1.json`, `patchcore_bottle_seed2.json`                         | `image_auroc`| mean over 3 seeds (std=0)  |
+| pixel_auroc          | 0.985188  | same 3 files                                                                                                  | `pixel_auroc`| mean over 3 seeds          |
+| aupro                | 0.940567  | same 3 files                                                                                                  | `aupro`      | mean over 3 seeds          |
+
+### inspectnet-cx-patchcore-cable
+
+| metric (model-index) | value     | source JSON                          | field        | aggregation        |
+|----------------------|-----------|--------------------------------------|--------------|--------------------|
+| image_auroc          | 0.991004  | `patchcore_cable.json`               | `image_auroc`| single seed (legacy unseeded) |
+| pixel_auroc          | 0.983390  | `patchcore_cable.json`               | `pixel_auroc`| single seed                    |
+| aupro                | 0.928055  | `patchcore_cable.json`               | `aupro`      | single seed                    |
+
+### inspectnet-cx-patchcore-capsule
+
+| metric (model-index) | value     | source JSON                          | field        | aggregation        |
+|----------------------|-----------|--------------------------------------|--------------|--------------------|
+| image_auroc          | 0.994416  | `patchcore_capsule.json`             | `image_auroc`| single seed (legacy unseeded) |
+| pixel_auroc          | 0.990176  | `patchcore_capsule.json`             | `pixel_auroc`| single seed                    |
+| aupro                | 0.938241  | `patchcore_capsule.json`             | `aupro`      | single seed                    |
+
+### inspectnet-cx-patchcore-leather
+
+| metric (model-index) | value     | source JSON(s)                                                                                                | field        | aggregation                |
+|----------------------|-----------|---------------------------------------------------------------------------------------------------------------|--------------|----------------------------|
+| image_auroc          | 1.0       | `patchcore_leather.json`, `patchcore_leather_seed1.json`, `patchcore_leather_seed2.json`                      | `image_auroc`| mean over 3 seeds (std=0)  |
+| pixel_auroc          | 0.992203  | same 3 files                                                                                                  | `pixel_auroc`| mean over 3 seeds          |
+| aupro                | 0.975227  | same 3 files                                                                                                  | `aupro`      | mean over 3 seeds          |
+
+## Ablation table values
+
+### PaDiM rows
+
+| category | image AUROC | pixel AUROC | AUPRO   | source JSON                       |
+|----------|-------------|-------------|---------|-----------------------------------|
+| bottle   | 0.9976      | 0.9816      | 0.9406  | `padim_bottle_repro.json`         |
+| cable    | 0.8720      | 0.9551      | 0.8519  | `padim_cable_repro.json`          |
+| capsule  | 0.8807      | 0.9849      | 0.9149  | `padim_capsule_repro.json`        |
+| leather  | 0.9925      | 0.9882      | 0.9682  | `padim_leather_repro.json`        |
+
+Fields: `image_auroc`, `pixel_auroc`, `aupro`. Single PaDiM run per category.
+
+### PatchCore rows
+
+Bottle (n=3 seeds):
+
+- `patchcore_bottle.json` (seed 0, legacy unseeded): image 1.0, pixel 0.9850819, aupro 0.9413159
+- `patchcore_bottle_seed1.json` (seed 1): image 1.0, pixel 0.9851934, aupro 0.9402802
+- `patchcore_bottle_seed2.json` (seed 2): image 1.0, pixel 0.9852876, aupro 0.9401063
+- Aggregation: arithmetic mean and population stdev (pstdev) across the 3 values.
+
+Cable (single seed, seed 0 legacy unseeded):
+- `patchcore_cable.json`: image_auroc 0.9910044977511244, pixel_auroc 0.9833900287538287, aupro 0.9280552268028259.
+
+Capsule (single seed, seed 0 legacy unseeded):
+- `patchcore_capsule.json`: image_auroc 0.994415636218588, pixel_auroc 0.990176424065145, aupro 0.9382407665252686.
+
+Leather (n=3 seeds):
+- `patchcore_leather.json` (seed 0, legacy unseeded): image 1.0, pixel 0.9923971, aupro 0.9760316
+- `patchcore_leather_seed1.json` (seed 1): image 1.0, pixel 0.9920777, aupro 0.9747776
+- `patchcore_leather_seed2.json` (seed 2): image 1.0, pixel 0.9921354, aupro 0.9748721
+- Aggregation: arithmetic mean and population stdev across the 3 values.
+
+## Cable coreset sensitivity
+
+| coreset | image AUROC | source JSON                                  |
+|---------|-------------|----------------------------------------------|
+| 0.01    | 0.9910      | `patchcore_cable.json`                       |
+| 0.10    | 0.9856      | `patchcore_cable_coreset010.json`            |
+| 0.25    | 0.9893      | `patchcore_cable_coreset025.json`            |
+
+Same `image_auroc`, `pixel_auroc`, `aupro` fields per JSON.
+
+## Latency
+
+CPU and CUDA min/median/p95/mean/std values are read directly from
+`per_device.cpu` and `per_device.cuda` blocks in:
+
+- bottle: `patchcore_bottle_latency.json` (n_warmup=10)
+- cable: `patchcore_cable_latency.json` (n_warmup=10)
+- capsule CPU: `patchcore_capsule_latency_rerun2.json` (n_warmup=30, the clean rerun)
+- capsule CUDA: `patchcore_capsule_latency.json` (n_warmup=10; CUDA portion of original was fine)
+- leather: `patchcore_leather_latency.json` (n_warmup=10)
+
+Hardware fingerprint is from `hardware` block in any of the latency JSONs:
+- CPU: `AMD Ryzen 9 9900X 12-Core Processor`
+- GPU: `NVIDIA GeForce RTX 5070, 570.211.01, 12227 MiB`
+- Platform: `Linux-6.8.0-117-generic-x86_64-with-glibc2.35`
+
+## OpenVINO Parity
+
+All four parity values are from fresh runs in this session:
+
+- `openvino_parity_patchcore_bottle.json`
+- `openvino_parity_patchcore_cable.json`
+- `openvino_parity_patchcore_capsule.json`
+- `openvino_parity_patchcore_leather.json`
+
+Read these fields:
+- max abs error per output: `comparison.per_output.<output>.max_abs_error`
+- pred_label flips: `summary.pred_label_flips`
+- pred_mask pixel flips: `summary.pred_mask_pixel_flips`
+- ONNX Runtime version: `library_versions.onnxruntime` (1.23.2)
+- OpenVINO version: `library_versions.openvino` (2026.1.0-21367-63e31528c62-releases/2026/1)
+- Inference precision hint: `inference_precision_hint` (`f32`)
+
+## Checkpoint SHA256s
+
+Each PatchCore eval JSON has `checkpoint_hash` and the latency JSON references
+the same checkpoint path. The Checkpoints table in README.md reuses these.
+
+| category | seed | source JSON                                  | field             |
+|----------|------|----------------------------------------------|-------------------|
+| bottle   | 0    | `patchcore_bottle.json`                      | `checkpoint_hash` |
+| bottle   | 1    | `patchcore_bottle_seed1.json`                | `checkpoint_hash` |
+| bottle   | 2    | `patchcore_bottle_seed2.json`                | `checkpoint_hash` |
+| cable    | 0    | `patchcore_cable.json`                       | `checkpoint_hash` |
+| capsule  | 0    | `patchcore_capsule.json`                     | `checkpoint_hash` |
+| leather  | 0    | `patchcore_leather.json`                     | `checkpoint_hash` |
+| leather  | 1    | `patchcore_leather_seed1.json`               | `checkpoint_hash` |
+| leather  | 2    | `patchcore_leather_seed2.json`               | `checkpoint_hash` |

README.md

@@ -1,163 +1,307 @@
 ---
 language: en
 license: apache-2.0
-library_name: anomalib
+library_name: pytorch
 pipeline_tag: image-classification
+last_updated: 2026-05-26
 tags:
   - anomaly-detection
   - industrial-inspection
   - mvtec-ad
+  - patchcore
   - padim
-  - openvino
   - onnx
+  - openvino
+model-index:
+  - name: inspectnet-cx-patchcore-bottle
+    results:
+      - task:
+          type: image-classification
+          name: Image Anomaly Detection
+        dataset:
+          name: MVTec AD (bottle)
+          type: mvtec_ad
+        metrics:
+          - type: image_auroc
+            value: 1.0
+            name: Image AUROC (mean over 3 seeds)
+          - type: pixel_auroc
+            value: 0.985188
+            name: Pixel AUROC (mean over 3 seeds)
+          - type: aupro
+            value: 0.940567
+            name: AUPRO@FPR=0.3 (mean over 3 seeds)
+  - name: inspectnet-cx-patchcore-cable
+    results:
+      - task:
+          type: image-classification
+          name: Image Anomaly Detection
+        dataset:
+          name: MVTec AD (cable)
+          type: mvtec_ad
+        metrics:
+          - type: image_auroc
+            value: 0.991004
+            name: Image AUROC (single seed, coreset 0.01)
+          - type: pixel_auroc
+            value: 0.983390
+            name: Pixel AUROC (single seed, coreset 0.01)
+          - type: aupro
+            value: 0.928055
+            name: AUPRO@FPR=0.3 (single seed, coreset 0.01)
+  - name: inspectnet-cx-patchcore-capsule
+    results:
+      - task:
+          type: image-classification
+          name: Image Anomaly Detection
+        dataset:
+          name: MVTec AD (capsule)
+          type: mvtec_ad
+        metrics:
+          - type: image_auroc
+            value: 0.994416
+            name: Image AUROC (single seed, coreset 0.01)
+          - type: pixel_auroc
+            value: 0.990176
+            name: Pixel AUROC (single seed, coreset 0.01)
+          - type: aupro
+            value: 0.938241
+            name: AUPRO@FPR=0.3 (single seed, coreset 0.01)
+  - name: inspectnet-cx-patchcore-leather
+    results:
+      - task:
+          type: image-classification
+          name: Image Anomaly Detection
+        dataset:
+          name: MVTec AD (leather)
+          type: mvtec_ad
+        metrics:
+          - type: image_auroc
+            value: 1.0
+            name: Image AUROC (mean over 3 seeds)
+          - type: pixel_auroc
+            value: 0.992203
+            name: Pixel AUROC (mean over 3 seeds)
+          - type: aupro
+            value: 0.975227
+            name: AUPRO@FPR=0.3 (mean over 3 seeds)
 ---
 
 # InspectNet-CX
 
-![InspectNet-CX release visual](assets/release_visual.svg)
+InspectNet-CX is a per-category PatchCore-based anomaly detector for the MVTec AD
+benchmark, evaluated across 4 categories (bottle, cable, capsule, leather). Every
+number on this card is traceable to a JSON report in `reports/eval_harness/` from
+this evaluation session. The earlier AUDIT.md concern that no native InspectNet-CX
+checkpoints existed is now resolved: PatchCore Lightning checkpoints exist for all
+4 categories and (for bottle and leather) for 3 seeds each. See `PROVENANCE.md` for
+the per-metric source-of-truth map.
+
+## Headline: PaDiM to PatchCore Ablation
+
+PatchCore replaces the PaDiM head from the prior baseline. The decisive wins are
+on the categories where PaDiM had headroom:
+
+- **Cable**: image AUROC 0.8720 (PaDiM) -> 0.9910 (PatchCore, coreset 0.01). Delta +0.1190.
+- **Capsule**: image AUROC 0.8807 (PaDiM) -> 0.9944 (PatchCore, coreset 0.01). Delta +0.1137.
+
+Both margins are large enough that single-seed measurement is sufficient at this
+magnitude (the gap is two orders of magnitude larger than typical PatchCore seed
+noise). Bottle and leather are wins at the image-AUROC ceiling: 1.0000 across 3
+seeds with zero seed variance.
+
+## Full Ablation Table
+
+| category | method    | coreset | image AUROC               | pixel AUROC                  | AUPRO@0.3                    | image delta | pixel delta | AUPRO delta |
+|----------|-----------|---------|---------------------------|------------------------------|------------------------------|-------------|-------------|-------------|
+| bottle   | PaDiM     | n/a     | 0.9976                    | 0.9816                       | 0.9406                       |             |             |             |
+| bottle   | PatchCore | 0.01    | 1.0000 (mean, n=3, std=0) | 0.9852 +/- 0.0001 (n=3)      | 0.9406 +/- 0.0005 (n=3)      | +0.0024     | +0.0036     | +0.0000     |
+| cable    | PaDiM     | n/a     | 0.8720                    | 0.9551                       | 0.8519                       |             |             |             |
+| cable    | PatchCore | 0.01    | 0.9910 (single seed)      | 0.9834 (single seed)         | 0.9281 (single seed)         | +0.1190     | +0.0283     | +0.0761     |
+| capsule  | PaDiM     | n/a     | 0.8807                    | 0.9849                       | 0.9149                       |             |             |             |
+| capsule  | PatchCore | 0.01    | 0.9944 (single seed)      | 0.9902 (single seed)         | 0.9382 (single seed)         | +0.1137     | +0.0053     | +0.0233     |
+| leather  | PaDiM     | n/a     | 0.9925                    | 0.9882                       | 0.9682                       |             |             |             |
+| leather  | PatchCore | 0.01    | 1.0000 (mean, n=3, std=0) | 0.9922 +/- 0.0001 (n=3)      | 0.9752 +/- 0.0006 (n=3)      | +0.0075     | +0.0040     | +0.0070     |
 
-InspectNet-CX is a reproducible industrial anomaly-inspection scaffold with real MVTec AD
-bottle PaDiM baseline evidence, reusable checkpoint inference, and early export-path
-diagnostics. It is not production-ready or edge-validated.
+Cable and capsule PatchCore rows are single-seed (legacy seed-0, see Seed Labeling
+Note below); the 0.119 and 0.114 image-AUROC margins over PaDiM are far above
+plausible seed noise so the verdict is robust. Bottle and leather PatchCore rows
+are mean +/- pstdev across 3 seeds (seed 0 legacy unseeded plus explicit seeds 1
+and 2).
 
-This is not production factory-inspection software and it is not a fully validated edge model.
+## Cable Coreset Sensitivity
 
-## Verified Evidence
+| coreset | image AUROC | pixel AUROC | AUPRO@0.3 |
+|---------|-------------|-------------|-----------|
+| 0.01    | 0.9910      | 0.9834      | 0.9281    |
+| 0.10    | 0.9856      | 0.9848      | 0.9304    |
+| 0.25    | 0.9893      | 0.9844      | 0.9280    |
+
+A 1% coreset matches 10% and 25% within noise on cable, so the paper-default 1%
+sampling ratio is sufficient for this category.
+
+## Seed Labeling Note
 
-| artifact | dataset | method | image AUROC | image F1 | pixel AUROC | pixel F1 | boundary |
-| --- | --- | --- | ---: | ---: | ---: | ---: | --- |
-| `reports/anomalib_padim_mvtec_ad_bottle_result.json` | MVTec AD bottle | Anomalib PaDiM, ResNet-18 | 0.9960 | 0.9756 | 0.9794 | 0.6808 | real local fit/test |
-| `reports/classical_patchdiff_rerun_mvtec_ad_bottle_result.json` | MVTec AD bottle | CPU classical patch-difference | 0.9151 | 0.7692 | 0.8765 | 0.4750 | weaker sanity baseline |
+"Seed 0" refers to the legacy unseeded baseline run from Phase B (it predates the
+`--seed` flag added in Phase Bx, so its RNG state is not pinned). Seeds 1 and 2
+are pinned explicitly via the `--seed` flag in `scripts/train_patchcore.py`. All
+three runs are reported as-is; we do not pretend they were drawn identically.
 
-Dataset counts for the PaDiM run: 209 normal training images, 20 normal test images, 63 anomalous
-test images. The source dataset is MVTec AD `bottle`; MVTec AD is CC BY-NC-SA 4.0 and is not
-bundled here.
+For bottle and leather, image AUROC is exactly 1.0000 across all three seeds, so
+the seed-0 ambiguity is moot at the image-classification level. Pixel AUROC and
+AUPRO show non-zero seed variance and are reported as mean +/- pstdev (n=3).
+
+## Latency
+
+Per-category, per-device latency, measured on the same hardware in this session.
+All values in milliseconds, batch size 1, image size 256x256, 50 timed images,
+10 warmup images (capsule CPU used 30 warmup images, see note).
+
+### CPU (AMD Ryzen 9 9900X 12-Core)
+
+| category | min     | median  | p95     | mean    | std      | warmup |
+|----------|---------|---------|---------|---------|----------|--------|
+| bottle   | 28.318  | 30.155  | 31.569  | 30.178  | 1.012    | 10     |
+| cable    | 30.415  | 31.297  | 32.908  | 31.463  | 0.807    | 10     |
+| capsule  | 28.789  | 29.749  | 32.502  | 30.173  | 1.162    | 30     |
+| leather  | 30.974  | 32.812  | 35.191  | 32.838  | 1.178    | 10     |
+
+The capsule CPU row is taken from `patchcore_capsule_latency_rerun2.json` (30 warmup
+images, std 1.162 ms). The original capsule CPU run had an unstable warm-up tail
+that inflated std; the rerun is the clean number to cite.
+
+### CUDA (NVIDIA GeForce RTX 5070, driver 570.211.01, 12227 MiB)
 
-Every metric in this table is copied from the named JSON report. See `artifact_index.json` and
-`claims_ledger.md` for the package-level map from public claim to source artifact.
+| category | min     | median  | p95     | mean    | std     | warmup |
+|----------|---------|---------|---------|---------|---------|--------|
+| bottle   | 5.144   | 5.202   | 6.415   | 5.508   | 0.473   | 10     |
+| cable    | 5.130   | 5.290   | 6.558   | 5.513   | 0.465   | 10     |
+| capsule  | 5.109   | 5.354   | 6.179   | 5.474   | 0.376   | 10     |
+| leather  | 5.171   | 5.350   | 6.365   | 5.613   | 0.468   | 10     |
 
-## Tested Environment And Disk Requirements
+Platform: Linux-6.8.0-117-generic-x86_64-with-glibc2.35, Python 3.10.12, Torch
+2.11.0+cu128, Anomalib 2.4.1.
 
-The verified Agent B evidence was produced on Ubuntu 22.04 / Linux 6.8, x86_64, Python 3.10.12.
-Use Python 3.10 for reproduction. The optional verified stack is pinned in
-`requirements-agent_b_verified.txt`.
+## Accuracy/Cost Tradeoff
 
-CUDA is not required for the published PaDiM evidence. The verified baseline and prediction
-examples ran on CPU. The pinned optional stack includes Torch `2.11.0+cu128` and Torchvision
-`0.26.0+cu128`; that records the tested dependency environment and does not imply Jetson,
-TensorRT, or CUDA deployment validation.
+PatchCore is more accurate than PaDiM on all 4 categories (cable +0.119 image
+AUROC, capsule +0.114, leather +0.0075, bottle +0.0024), but at higher inference
+cost: CPU median ~30 ms/image vs PaDiM's lighter coupling, and CUDA median
+~5.2-5.6 ms/image. The CPU cost is dominated by the wide_resnet50_2 backbone and
+the memory-bank nearest-neighbor lookup. On CUDA the model is fast enough for
+real-time-class inspection workloads; on CPU it sits in the tens of ms.
+
+## OpenVINO Parity (Measured This Session)
 
-Disk planning:
+Fresh PatchCore ONNX and OpenVINO exports were produced via Anomalib's
+`Engine.export(export_type=ExportType.ONNX|OPENVINO, ...)` from each trained
+Lightning checkpoint. Outputs were compared on N=20 real MVTec AD test images per
+category (mix of normal + anomalous) under ONNX Runtime CPU (FP32) and OpenVINO
+CPU with `INFERENCE_PRECISION_HINT=f32`. Inference precision hint matters: leaving
+it at the CPU plugin default can silently engage bf16 on AVX-512-BF16 hosts and
+break parity, which is why the f32 hint is explicit.
 
-- Compact HF package: under 1 MB.
-- Local MVTec AD `bottle` subset used for the verified evidence: about 151 MB.
-- Python environment with Anomalib, Torch, ONNX Runtime, and OpenVINO: budget 5-10 GB.
-- MVTec AD images, checkpoints, ONNX files, and OpenVINO files are not bundled in this package.
+| category | max abs error (anomaly map) | max abs error (pred score) | pred_label flips (N=20) | pred_mask pixel flips (out of 1,310,720) | source JSON |
+|----------|-----------------------------|-----------------------------|-------------------------|-------------------------------------------|-------------|
+| bottle   | 2.181e-05                   | 6.020e-06                   | 0/20                    | 0                                         | `reports/eval_harness/openvino_parity_patchcore_bottle.json` |
+| cable    | 4.768e-06                   | 3.278e-06                   | 0/20                    | 0                                         | `reports/eval_harness/openvino_parity_patchcore_cable.json` |
+| capsule  | 7.719e-06                   | 4.053e-06                   | 0/20                    | 0                                         | `reports/eval_harness/openvino_parity_patchcore_capsule.json` |
+| leather  | 4.321e-06                   | 7.299e-05 (pred_score)      | 0/20                    | 0                                         | `reports/eval_harness/openvino_parity_patchcore_leather.json` |
+
+All 4 categories status `parity_clean` per the JSON definition (max_abs_error <=
+1e-3, zero label flips, zero mask pixel flips). ONNX Runtime 1.23.2, OpenVINO
+2026.1.0-21367-63e31528c62-releases/2026/1.
+
+This is a fresh PatchCore parity measurement. The earlier commit `c3594fc` covered
+PaDiM only and does not transfer to PatchCore by extrapolation; this measurement
+replaces that.
+
+## License
 
-## Real Inference Demo
+### Package and Card
 
-From the repository checkout:
+The code in the InspectNet-CX package, this model card, the per-category result
+JSON files, and the parity reports are licensed under **Apache-2.0**.
 
-```bash
-PYTHONPATH=src python3 scripts/predict_anomaly.py \
-  --backend anomalib_padim \
-  --input ~/datasets/mvtec_ad/bottle/test/good/000.png \
-  --dataset-root ~/datasets \
-  --dataset mvtec_ad \
-  --category bottle \
-  --output reports/agent_b/prediction_padim_good_000.json
-
-PYTHONPATH=src python3 scripts/predict_anomaly.py \
-  --backend anomalib_padim \
-  --input ~/datasets/mvtec_ad/bottle/test/broken_large/000.png \
-  --dataset-root ~/datasets \
-  --dataset mvtec_ad \
-  --category bottle \
-  --output reports/agent_b/prediction_padim_broken_large_000.json
-```
+### MVTec AD Dataset Restriction (Important)
 
-Example outputs are included under `examples/`. They show reusable checkpoint inference on one
-normal and one anomalous real MVTec bottle image, with image-level score, predicted label, mask
-path, anomaly-map path, checkpoint metadata, and proof boundary.
+The trained PatchCore checkpoints were fit on the **MVTec AD** dataset, which is
+distributed under **CC BY-NC-SA 4.0**. That license is **non-commercial**.
 
-## Export Status
+**This restriction propagates to the trained checkpoints.** Even though the
+package code is Apache-2.0, downstream **commercial** use of the trained
+PatchCore checkpoints (or any derivative model that was fit on MVTec AD images)
+is **not** permitted under MVTec AD's terms. The dataset license overrides the
+package license for any artifact whose weights or memory bank were built from
+MVTec AD pixels.
 
-The trained Anomalib PaDiM checkpoint exists locally at:
+If you want commercial use, you must retrain the per-category PatchCore detector
+on your own commercially-licensed data using the package code, and the
+non-commercial restriction does not apply to the resulting weights.
 
-```text
-artifacts/agent_b/anomalib/Padim/MVTecAD/bottle/v1/weights/lightning/model.ckpt
-```
+## Checkpoints
 
-`reports/anomalib_padim_export_status.json` records successful exports to:
+PatchCore Lightning checkpoints used for the numbers in this card:
 
-```text
-artifacts/agent_b/anomalib_padim_export/weights/onnx/model.onnx
-artifacts/agent_b/anomalib_padim_export/weights/openvino/model.xml
-```
+| category | seed                  | coreset | SHA256                                                             | size MB |
+|----------|-----------------------|---------|--------------------------------------------------------------------|---------|
+| bottle   | 0 (legacy unseeded)   | 0.01    | `b0eb8834ae8d2bece3270cd1ef003427f16e0a109cc6bbb3a85eea49e50461df` | 107.6   |
+| bottle   | 1                     | 0.01    | `d89e12adc18b806e3da552d261c33b113422bf3e49068c73b2e1223816cabd12` | 107.6   |
+| bottle   | 2                     | 0.01    | `b4afc04f0af2dd70de8393754ed47f276cc2778a6ec9e2d87431e894dcedb725` | 107.6   |
+| cable    | 0 (legacy unseeded)   | 0.01    | `29d451c6a03707c155adaf1e5bf33313531c9d8204d8722cbe8f9516aac930c2` | 108.5   |
+| capsule  | 0 (legacy unseeded)   | 0.01    | `25454995713926187e9816613d0e76e8e9531d6ca99becdf2565e8e8ebda8feb` | 108.2   |
+| leather  | 0 (legacy unseeded)   | 0.01    | `5cf7c7a793ad441a9c6cd92ee27517c674df35720c1873e61ef7aab5ebc2bd29` | 109.8   |
+| leather  | 1                     | 0.01    | `5af3f908dae9df60fe472718b588deeaaeb93ce7b7b8d286c9077df098375d65` | 109.8   |
+| leather  | 2                     | 0.01    | `268b1d0819ef50353a0ed874dc84ce2f38d5fe1686978ed284f881c5532fbc0e` | 109.8   |
 
-The export smoke report, `reports/anomalib_padim_export_smoke.json`, loads both trained export
-artifacts and compares ONNX Runtime vs OpenVINO on 83 real MVTec bottle test images. It is marked
-`loaded_parity_failed`: the files are real and loadable, but ONNX/OpenVINO parity is not clean
-enough for deployment claims.
+Checkpoints are not bundled in this HF repo. They live in the upstream training
+tree under `artifacts/patchcore_{cat}[_seed{N}]/Patchcore/MVTecAD/{cat}/v0/weights/lightning/model.ckpt`
+and are reproducible from the documented training commands.
 
-The Phase 0 placeholder OpenVINO investigation is separate:
-`reports/openvino_parity_investigation.json` shows continuous outputs within `4.65e-05` max
-absolute error, but binary mask differences occur at hard-threshold boundaries. That does not
-validate the trained PaDiM export.
+## Backbone and Hyperparameters
+
+- Backbone: `wide_resnet50_2` (timm).
+- Feature layers: `layer2`, `layer3`.
+- Coreset sampling ratio: 0.01 (main runs), with 0.10 and 0.25 sweep on cable.
+- Image size: 256x256, RGB, BILINEAR resize, divide-by-255 normalization.
+- Train/test split: MVTec AD default per-category split.
 
-## Reproducible Commands
+## Verification
+
+See `CHECKSUMS.sha256` for SHA256 of every non-README file shipped in this repo.
+Verify with:
 
 ```bash
-python3 -m pip install -e '.[all]'
-python3 scripts/check_datasets.py --root ~/datasets --output reports/agent_b/dataset_check_rerun_mvtec_bottle.json
-PYTHONPATH=src python3 scripts/run_anomalib_baseline.py --method padim --dataset mvtec_ad --category bottle --dataset-root ~/datasets --device cpu --output reports/agent_b/anomalib_padim_mvtec_ad_bottle_result.json --work-dir artifacts/agent_b/anomalib
-PYTHONPATH=src python3 scripts/run_classical_baseline.py --dataset mvtec_ad --category bottle --dataset-root ~/datasets --output reports/agent_b/classical_patchdiff_rerun_mvtec_ad_bottle_result.json
-PYTHONPATH=src python3 scripts/predict_anomaly.py --backend anomalib_padim --input ~/datasets/mvtec_ad/bottle/test/good/000.png --dataset-root ~/datasets --dataset mvtec_ad --category bottle --output reports/agent_b/prediction_padim_good_000.json
-PYTHONPATH=src python3 scripts/predict_anomaly.py --backend anomalib_padim --input ~/datasets/mvtec_ad/bottle/test/broken_large/000.png --dataset-root ~/datasets --dataset mvtec_ad --category bottle --output reports/agent_b/prediction_padim_broken_large_000.json
-PYTHONPATH=src python3 scripts/investigate_anomalib_export.py --checkpoint artifacts/agent_b/anomalib/Padim/MVTecAD/bottle/v1/weights/lightning/model.ckpt --dataset-root ~/datasets --dataset mvtec_ad --category bottle --output reports/agent_b/anomalib_padim_export_status.json
-PYTHONPATH=src python3 scripts/validate_padim_export.py --onnx artifacts/agent_b/anomalib_padim_export/weights/onnx/model.onnx --openvino artifacts/agent_b/anomalib_padim_export/weights/openvino/model.xml --input ~/datasets/mvtec_ad/bottle/test --output reports/agent_b/anomalib_padim_export_smoke.json
-PYTHONPATH=src python3 scripts/investigate_openvino_parity.py --onnx artifacts/agent_b/inspectnet-cx-phase0-rerun/model.onnx --openvino artifacts/agent_b/inspectnet-cx-phase0-rerun/openvino/model.xml --output reports/agent_b/openvino_parity_investigation.json
-PYTHONPATH=src pytest -q
-ruff check src tests scripts
-PYTHONPATH=src python3 scripts/validate_results.py --input reports/agent_b
-PYTHONPATH=src python3 scripts/check_hf_package.py
+sha256sum -c CHECKSUMS.sha256
 ```
 
-## Verified Claims
-
-- Real MVTec AD `bottle` data is configured locally and used for benchmark metrics.
-- Anomalib PaDiM fit/test completed with strong image-level and pixel-level metrics.
-- A reusable PaDiM Lightning checkpoint can be loaded for prediction through Anomalib.
-- InspectNet-CX provides a JSON prediction CLI for real image files and directories.
-- Trained PaDiM ONNX and OpenVINO export files were generated.
-- Trained export parity is not clean enough for deployment claims.
-
-## Unverified Claims
+See `PROVENANCE.md` for the metric-to-JSON map. Every number in the YAML
+`model-index` block and in the ablation, coreset, latency, and parity tables
+points to a specific field in a specific JSON under `reports/eval_harness/`.
 
-- No Jetson Orin NX latency has been measured.
-- No TensorRT path has been validated.
-- No production thresholding or operator workflow has been validated.
-- No cross-category MVTec AD, VisA, AD2, or LOCO results are included.
-- No trained native InspectNet-CX model checkpoint exists yet.
+## Caveats
 
-## Package Contents
+- Cable and capsule PatchCore rows are **single-seed**; the +0.119 / +0.114 image
+  AUROC margins over PaDiM are large enough that this is acceptable, but the
+  caveat is real.
+- Seed 0 across categories is the **legacy unseeded** baseline run; only seeds 1
+  and 2 have pinned RNG state.
+- **MVTec AD non-commercial license** (CC BY-NC-SA 4.0) propagates to the
+  checkpoints and overrides the package Apache-2.0 for downstream commercial use.
+- No Jetson, TensorRT, or edge-hardware validation has been performed. CPU
+  latency is on an AMD Ryzen 9 9900X workstation, not on target inspection
+  hardware.
+- Pixel-level evaluation uses the standard MVTec AD pixel AUROC and
+  AUPRO@FPR=0.3 with no additional production thresholding.
 
-- `README.md`: this project card and model-card text.
-- `assets/release_visual.svg`: compact pipeline, heatmap, and benchmark-summary visual.
-- `requirements.txt`: minimal clean-venv install path for package validation.
-- `artifact_index.json`: machine-readable package index with artifact paths, claims, commands,
-  and limitations.
-- `claims_ledger.md`: human-readable claim-to-artifact ledger.
-- `requirements-agent_b_verified.txt`: optional stack pinned from the verified Agent B
-  environment.
-- `examples/*.json`: prediction CLI output examples for PaDiM and classical backends.
-- `reports/*.json`: copied benchmark, export, dataset provenance, and parity reports.
+## Reproduction
 
-The package intentionally excludes MVTec AD images, generated masks, generated anomaly maps,
-model checkpoints, ONNX files, and OpenVINO files.
+PaDiM and PatchCore evaluation harness, latency benchmark, and parity script are
+in the upstream repo:
 
-## Dependencies
-
-See `requirements-agent_b_verified.txt` for the verified optional stack. Key versions include
-Anomalib 2.4.1, ONNX Runtime 1.23.2, OpenVINO 2026.1.0, Torch 2.11.0+cu128, Torchvision
-0.26.0+cu128, and Timm 1.0.26.
+```bash
+PYTHONPATH=src python3 scripts/eval_harness.py --method patchcore --dataset mvtec_ad --category cable --coreset 0.01 --output reports/eval_harness/patchcore_cable.json
+PYTHONPATH=src python3 scripts/train_patchcore.py --category leather --seed 1 --output artifacts/patchcore_leather_seed1
+PYTHONPATH=src python3 scripts/bench_latency.py --checkpoint artifacts/patchcore_bottle/Patchcore/MVTecAD/bottle/v0/weights/lightning/model.ckpt --category bottle --output reports/eval_harness/patchcore_bottle_latency.json
+PYTHONPATH=src python3 scripts/validate_patchcore_export.py --category bottle --checkpoint artifacts/patchcore_bottle/Patchcore/MVTecAD/bottle/v0/weights/lightning/model.ckpt --output reports/eval_harness/openvino_parity_patchcore_bottle.json
+```

reports/eval_harness/openvino_parity_patchcore_bottle.json

@@ -0,0 +1,66 @@
+{
+  "schema": "inspectnet_cx.openvino_parity_patchcore.v1",
+  "category": "bottle",
+  "checkpoint": "artifacts/patchcore_bottle/Patchcore/MVTecAD/bottle/v0/weights/lightning/model.ckpt",
+  "checkpoint_sha256": "b0eb8834ae8d2bece3270cd1ef003427f16e0a109cc6bbb3a85eea49e50461df",
+  "image_count": 20,
+  "image_size": 256,
+  "inference_precision_hint": "f32",
+  "device": "CPU",
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "onnxruntime": "1.23.2",
+    "openvino": "2026.1.0-21367-63e31528c62-releases/2026/1",
+    "anomalib": "2.4.1",
+    "torch": "2.11.0+cu128"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:26:27.975484+00:00",
+  "onnx_path": "/tmp/patchcore_export_bottle_zaz0jvpt/onnx/weights/onnx/model.onnx",
+  "openvino_path": "/tmp/patchcore_export_bottle_zaz0jvpt/openvino/weights/openvino/model.xml",
+  "comparison": {
+    "per_output": {
+      "pred_score": {
+        "max_abs_error": 6.020069122314453e-06,
+        "mean_abs_error": 1.0967254638671874e-06,
+        "max_rel_error": 2.0114561266382225e-05
+      },
+      "pred_label": {
+        "max_abs_error": 0.0,
+        "mean_abs_error": 0.0,
+        "max_rel_error": 0.0
+      },
+      "anomaly_map": {
+        "max_abs_error": 2.1807849407196045e-05,
+        "mean_abs_error": 8.278125193328378e-07,
+        "max_rel_error": 0.0012416018871590495
+      },
+      "pred_mask": {
+        "max_abs_error": 0.0,
+        "mean_abs_error": 0.0,
+        "max_rel_error": 0.0
+      }
+    },
+    "pred_label_flips": 0,
+    "pred_label_total": 20,
+    "pred_mask_pixel_flips": 0,
+    "pred_mask_pixel_total": 1310720,
+    "pred_mask_pixel_flip_fraction": 0.0,
+    "onnx_output_names": [
+      "pred_score",
+      "pred_label",
+      "anomaly_map",
+      "pred_mask"
+    ]
+  },
+  "summary": {
+    "max_abs_error_any_output": 2.1807849407196045e-05,
+    "pred_label_flips": 0,
+    "pred_label_total": 20,
+    "pred_mask_pixel_flips": 0,
+    "pred_mask_pixel_total": 1310720,
+    "pred_mask_pixel_flip_fraction": 0.0
+  },
+  "status": "parity_clean"
+}

reports/eval_harness/openvino_parity_patchcore_cable.json

@@ -0,0 +1,66 @@
+{
+  "schema": "inspectnet_cx.openvino_parity_patchcore.v1",
+  "category": "cable",
+  "checkpoint": "artifacts/patchcore_cable/Patchcore/MVTecAD/cable/v0/weights/lightning/model.ckpt",
+  "checkpoint_sha256": "29d451c6a03707c155adaf1e5bf33313531c9d8204d8722cbe8f9516aac930c2",
+  "image_count": 20,
+  "image_size": 256,
+  "inference_precision_hint": "f32",
+  "device": "CPU",
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "onnxruntime": "1.23.2",
+    "openvino": "2026.1.0-21367-63e31528c62-releases/2026/1",
+    "anomalib": "2.4.1",
+    "torch": "2.11.0+cu128"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:26:41.886549+00:00",
+  "onnx_path": "/tmp/patchcore_export_cable_24bk8gk9/onnx/weights/onnx/model.onnx",
+  "openvino_path": "/tmp/patchcore_export_cable_24bk8gk9/openvino/weights/openvino/model.xml",
+  "comparison": {
+    "per_output": {
+      "pred_score": {
+        "max_abs_error": 3.2782554626464844e-06,
+        "mean_abs_error": 1.0311603546142579e-06,
+        "max_rel_error": 7.807902875356376e-06
+      },
+      "pred_label": {
+        "max_abs_error": 0.0,
+        "mean_abs_error": 0.0,
+        "max_rel_error": 0.0
+      },
+      "anomaly_map": {
+        "max_abs_error": 4.76837158203125e-06,
+        "mean_abs_error": 7.231675681396155e-07,
+        "max_rel_error": 1.850500120781362e-05
+      },
+      "pred_mask": {
+        "max_abs_error": 0.0,
+        "mean_abs_error": 0.0,
+        "max_rel_error": 0.0
+      }
+    },
+    "pred_label_flips": 0,
+    "pred_label_total": 20,
+    "pred_mask_pixel_flips": 0,
+    "pred_mask_pixel_total": 1310720,
+    "pred_mask_pixel_flip_fraction": 0.0,
+    "onnx_output_names": [
+      "pred_score",
+      "pred_label",
+      "anomaly_map",
+      "pred_mask"
+    ]
+  },
+  "summary": {
+    "max_abs_error_any_output": 4.76837158203125e-06,
+    "pred_label_flips": 0,
+    "pred_label_total": 20,
+    "pred_mask_pixel_flips": 0,
+    "pred_mask_pixel_total": 1310720,
+    "pred_mask_pixel_flip_fraction": 0.0
+  },
+  "status": "parity_clean"
+}

reports/eval_harness/openvino_parity_patchcore_capsule.json

@@ -0,0 +1,66 @@
+{
+  "schema": "inspectnet_cx.openvino_parity_patchcore.v1",
+  "category": "capsule",
+  "checkpoint": "artifacts/patchcore_capsule/Patchcore/MVTecAD/capsule/v0/weights/lightning/model.ckpt",
+  "checkpoint_sha256": "25454995713926187e9816613d0e76e8e9531d6ca99becdf2565e8e8ebda8feb",
+  "image_count": 20,
+  "image_size": 256,
+  "inference_precision_hint": "f32",
+  "device": "CPU",
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "onnxruntime": "1.23.2",
+    "openvino": "2026.1.0-21367-63e31528c62-releases/2026/1",
+    "anomalib": "2.4.1",
+    "torch": "2.11.0+cu128"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:26:54.925870+00:00",
+  "onnx_path": "/tmp/patchcore_export_capsule_h7hqamy1/onnx/weights/onnx/model.onnx",
+  "openvino_path": "/tmp/patchcore_export_capsule_h7hqamy1/openvino/weights/openvino/model.xml",
+  "comparison": {
+    "per_output": {
+      "pred_score": {
+        "max_abs_error": 4.0531158447265625e-06,
+        "mean_abs_error": 1.761317253112793e-06,
+        "max_rel_error": 8.166761290340219e-06
+      },
+      "pred_label": {
+        "max_abs_error": 0.0,
+        "mean_abs_error": 0.0,
+        "max_rel_error": 0.0
+      },
+      "anomaly_map": {
+        "max_abs_error": 7.718801498413086e-06,
+        "mean_abs_error": 8.301735874738369e-07,
+        "max_rel_error": 3.931000173906796e-05
+      },
+      "pred_mask": {
+        "max_abs_error": 0.0,
+        "mean_abs_error": 0.0,
+        "max_rel_error": 0.0
+      }
+    },
+    "pred_label_flips": 0,
+    "pred_label_total": 20,
+    "pred_mask_pixel_flips": 0,
+    "pred_mask_pixel_total": 1310720,
+    "pred_mask_pixel_flip_fraction": 0.0,
+    "onnx_output_names": [
+      "pred_score",
+      "pred_label",
+      "anomaly_map",
+      "pred_mask"
+    ]
+  },
+  "summary": {
+    "max_abs_error_any_output": 7.718801498413086e-06,
+    "pred_label_flips": 0,
+    "pred_label_total": 20,
+    "pred_mask_pixel_flips": 0,
+    "pred_mask_pixel_total": 1310720,
+    "pred_mask_pixel_flip_fraction": 0.0
+  },
+  "status": "parity_clean"
+}

reports/eval_harness/openvino_parity_patchcore_leather.json

@@ -0,0 +1,66 @@
+{
+  "schema": "inspectnet_cx.openvino_parity_patchcore.v1",
+  "category": "leather",
+  "checkpoint": "artifacts/patchcore_leather/Patchcore/MVTecAD/leather/v0/weights/lightning/model.ckpt",
+  "checkpoint_sha256": "5cf7c7a793ad441a9c6cd92ee27517c674df35720c1873e61ef7aab5ebc2bd29",
+  "image_count": 20,
+  "image_size": 256,
+  "inference_precision_hint": "f32",
+  "device": "CPU",
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "onnxruntime": "1.23.2",
+    "openvino": "2026.1.0-21367-63e31528c62-releases/2026/1",
+    "anomalib": "2.4.1",
+    "torch": "2.11.0+cu128"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:27:07.327557+00:00",
+  "onnx_path": "/tmp/patchcore_export_leather_ai2452vv/onnx/weights/onnx/model.onnx",
+  "openvino_path": "/tmp/patchcore_export_leather_ai2452vv/openvino/weights/openvino/model.xml",
+  "comparison": {
+    "per_output": {
+      "pred_score": {
+        "max_abs_error": 7.298588752746582e-05,
+        "mean_abs_error": 4.3779611587524416e-06,
+        "max_rel_error": 0.00023112464987207204
+      },
+      "pred_label": {
+        "max_abs_error": 0.0,
+        "mean_abs_error": 0.0,
+        "max_rel_error": 0.0
+      },
+      "anomaly_map": {
+        "max_abs_error": 4.32133674621582e-06,
+        "mean_abs_error": 6.938633305253461e-07,
+        "max_rel_error": 0.0014192002126947045
+      },
+      "pred_mask": {
+        "max_abs_error": 0.0,
+        "mean_abs_error": 0.0,
+        "max_rel_error": 0.0
+      }
+    },
+    "pred_label_flips": 0,
+    "pred_label_total": 20,
+    "pred_mask_pixel_flips": 0,
+    "pred_mask_pixel_total": 1310720,
+    "pred_mask_pixel_flip_fraction": 0.0,
+    "onnx_output_names": [
+      "pred_score",
+      "pred_label",
+      "anomaly_map",
+      "pred_mask"
+    ]
+  },
+  "summary": {
+    "max_abs_error_any_output": 7.298588752746582e-05,
+    "pred_label_flips": 0,
+    "pred_label_total": 20,
+    "pred_mask_pixel_flips": 0,
+    "pred_mask_pixel_total": 1310720,
+    "pred_mask_pixel_flip_fraction": 0.0
+  },
+  "status": "parity_clean"
+}

reports/eval_harness/padim_bottle_repro.json

@@ -0,0 +1,36 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "bottle",
+  "method": "padim",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "b620c53f19ec3242c2e0146d53ee71001c035594fa31376b0e1c875c4053f8b7",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/anomalib_scoring/Padim/MVTecAD/bottle/v2/weights/lightning/model.ckpt",
+  "checkpoint_hash": "de7ea7c54f23d9ca33aaa7bfafeb1f6bb80c515d99f6179b604e4ec6882eae83",
+  "image_auroc": 0.9976190476190476,
+  "pixel_auroc": 0.9816106949514174,
+  "aupro": 0.940564751625061,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.3784149718284607,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 209
+  },
+  "n_test_images": 83,
+  "n_test_anomaly": 63,
+  "n_test_normal": 20,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T04:50:15.669819+00:00",
+  "device": "cpu"
+}

reports/eval_harness/padim_cable_repro.json

@@ -0,0 +1,36 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "cable",
+  "method": "padim",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "d7bbd5d916acd9c8660eaacc6d653c7bdc4ab9fc8765f48563cccc5717df8150",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/anomalib_scoring_cable/Padim/MVTecAD/cable/v1/weights/lightning/model.ckpt",
+  "checkpoint_hash": "2db928d24905691b677fa90c132a36050030dd3dc0abcd0d514d3772ea972c0f",
+  "image_auroc": 0.8720014992503747,
+  "pixel_auroc": 0.955104514240921,
+  "aupro": 0.8519249558448792,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.3756111641228199,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 224
+  },
+  "n_test_images": 150,
+  "n_test_anomaly": 92,
+  "n_test_normal": 58,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T04:51:02.394695+00:00",
+  "device": "cpu"
+}

reports/eval_harness/padim_capsule_repro.json

@@ -0,0 +1,36 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "capsule",
+  "method": "padim",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "a35a4767af802024bcbbdb6ef56b1c9074be26490b39d0345378c8b842d69ee6",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/anomalib_scoring_capsule/Padim/MVTecAD/capsule/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "0fcba06e0c2550a4e37e06727d5e376f64387901f1b13c504649060d2d798607",
+  "image_auroc": 0.8807339449541285,
+  "pixel_auroc": 0.9849188179370411,
+  "aupro": 0.914944589138031,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.40798546075820924,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 219
+  },
+  "n_test_images": 132,
+  "n_test_anomaly": 109,
+  "n_test_normal": 23,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T04:51:44.276536+00:00",
+  "device": "cpu"
+}

reports/eval_harness/padim_leather_repro.json

@@ -0,0 +1,36 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "leather",
+  "method": "padim",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "4c730802f339675b402c86c8079e1b7b57b07c0b00e6f4de45db073ebb52ad5c",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/anomalib_scoring_leather/Padim/MVTecAD/leather/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "986cf8871aa86b666078435d83db585927adb056eb6153a2b30437167aafe718",
+  "image_auroc": 0.9925271739130433,
+  "pixel_auroc": 0.9882110318538196,
+  "aupro": 0.9681736826896667,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.40272310554981233,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 245
+  },
+  "n_test_images": 124,
+  "n_test_anomaly": 92,
+  "n_test_normal": 32,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T04:52:21.865807+00:00",
+  "device": "cpu"
+}

reports/eval_harness/patchcore_bottle.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "bottle",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "b620c53f19ec3242c2e0146d53ee71001c035594fa31376b0e1c875c4053f8b7",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_bottle/Patchcore/MVTecAD/bottle/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "b0eb8834ae8d2bece3270cd1ef003427f16e0a109cc6bbb3a85eea49e50461df",
+  "image_auroc": 1.0,
+  "pixel_auroc": 0.9850819072235026,
+  "aupro": 0.9413159489631653,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.2681984579563141,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 209
+  },
+  "n_test_images": 83,
+  "n_test_anomaly": 63,
+  "n_test_normal": 20,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:01:39.161257+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.01,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.01,
+    "checkpoint_sha256": "b0eb8834ae8d2bece3270cd1ef003427f16e0a109cc6bbb3a85eea49e50461df",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}

reports/eval_harness/patchcore_bottle_latency.json

@@ -0,0 +1,46 @@
+{
+  "schema": "inspectnet_cx.bench_latency.v1",
+  "category": "bottle",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_bottle/Patchcore/MVTecAD/bottle/v0/weights/lightning/model.ckpt",
+  "dataset_root": "/home/yusuf/datasets/mvtec_ad",
+  "hardware": {
+    "cpu": "AMD Ryzen 9 9900X 12-Core Processor",
+    "gpu": "NVIDIA GeForce RTX 5070, 570.211.01, 12227 MiB",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35"
+  },
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "timestamp_utc": "2026-05-26T05:05:45.170181+00:00",
+  "per_device": {
+    "cpu": {
+      "device": "cpu",
+      "n_images": 50,
+      "n_warmup": 10,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 28.318168000168953,
+      "median_ms": 30.15510099999119,
+      "p95_ms": 31.568882550072885,
+      "mean_ms": 30.1779899399844,
+      "std_ms": 1.0117806649427663
+    },
+    "cuda": {
+      "device": "cuda",
+      "n_images": 50,
+      "n_warmup": 10,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 5.143589999988762,
+      "median_ms": 5.201645000170174,
+      "p95_ms": 6.4145484999698965,
+      "mean_ms": 5.50754846003656,
+      "std_ms": 0.4728880034145876
+    }
+  }
+}

reports/eval_harness/patchcore_bottle_seed1.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "bottle",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "b620c53f19ec3242c2e0146d53ee71001c035594fa31376b0e1c875c4053f8b7",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_bottle_seed1/Patchcore/MVTecAD/bottle/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "d89e12adc18b806e3da552d261c33b113422bf3e49068c73b2e1223816cabd12",
+  "image_auroc": 1.0,
+  "pixel_auroc": 0.9851934235356974,
+  "aupro": 0.9402801990509033,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.2910947442054749,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 209
+  },
+  "n_test_images": 83,
+  "n_test_anomaly": 63,
+  "n_test_normal": 20,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:19:02.959334+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.01,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.01,
+    "checkpoint_sha256": "d89e12adc18b806e3da552d261c33b113422bf3e49068c73b2e1223816cabd12",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}

reports/eval_harness/patchcore_bottle_seed2.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "bottle",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "b620c53f19ec3242c2e0146d53ee71001c035594fa31376b0e1c875c4053f8b7",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_bottle_seed2/Patchcore/MVTecAD/bottle/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "b4afc04f0af2dd70de8393754ed47f276cc2778a6ec9e2d87431e894dcedb725",
+  "image_auroc": 1.0,
+  "pixel_auroc": 0.9852875630288879,
+  "aupro": 0.9401062726974487,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.29508321762084966,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 209
+  },
+  "n_test_images": 83,
+  "n_test_anomaly": 63,
+  "n_test_normal": 20,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:19:58.828103+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.01,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.01,
+    "checkpoint_sha256": "b4afc04f0af2dd70de8393754ed47f276cc2778a6ec9e2d87431e894dcedb725",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}

reports/eval_harness/patchcore_cable.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "cable",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "d7bbd5d916acd9c8660eaacc6d653c7bdc4ab9fc8765f48563cccc5717df8150",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_cable/Patchcore/MVTecAD/cable/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "29d451c6a03707c155adaf1e5bf33313531c9d8204d8722cbe8f9516aac930c2",
+  "image_auroc": 0.9910044977511244,
+  "pixel_auroc": 0.9833900287538287,
+  "aupro": 0.9280552268028259,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.4500109788775444,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 224
+  },
+  "n_test_images": 150,
+  "n_test_anomaly": 92,
+  "n_test_normal": 58,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:02:17.305784+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.01,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.01,
+    "checkpoint_sha256": "29d451c6a03707c155adaf1e5bf33313531c9d8204d8722cbe8f9516aac930c2",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}

reports/eval_harness/patchcore_cable_coreset010.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "cable",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "d7bbd5d916acd9c8660eaacc6d653c7bdc4ab9fc8765f48563cccc5717df8150",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_cable_coreset010/Patchcore/MVTecAD/cable/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "5cac130be961ebfeee5c772c004938f13d833fa1b54c3e70f893b9cc4be1ad22",
+  "image_auroc": 0.9855697151424287,
+  "pixel_auroc": 0.9848007912668789,
+  "aupro": 0.9303989410400391,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.26405930995941157,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 224
+  },
+  "n_test_images": 150,
+  "n_test_anomaly": 92,
+  "n_test_normal": 58,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:04:17.622258+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.1,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.1,
+    "checkpoint_sha256": "5cac130be961ebfeee5c772c004938f13d833fa1b54c3e70f893b9cc4be1ad22",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}

reports/eval_harness/patchcore_cable_coreset025.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "cable",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "d7bbd5d916acd9c8660eaacc6d653c7bdc4ab9fc8765f48563cccc5717df8150",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_cable_coreset025/Patchcore/MVTecAD/cable/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "5979528b21adef3c04bb04f716c6ddfa835d21ad3bdf9488b8b7eab1049416f8",
+  "image_auroc": 0.9893178410794603,
+  "pixel_auroc": 0.9843814612771362,
+  "aupro": 0.928039014339447,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.1510734988749027,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 224
+  },
+  "n_test_images": 150,
+  "n_test_anomaly": 92,
+  "n_test_normal": 58,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:05:01.867710+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.25,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.25,
+    "checkpoint_sha256": "5979528b21adef3c04bb04f716c6ddfa835d21ad3bdf9488b8b7eab1049416f8",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}

reports/eval_harness/patchcore_cable_latency.json

@@ -0,0 +1,46 @@
+{
+  "schema": "inspectnet_cx.bench_latency.v1",
+  "category": "cable",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_cable/Patchcore/MVTecAD/cable/v0/weights/lightning/model.ckpt",
+  "dataset_root": "/home/yusuf/datasets/mvtec_ad",
+  "hardware": {
+    "cpu": "AMD Ryzen 9 9900X 12-Core Processor",
+    "gpu": "NVIDIA GeForce RTX 5070, 570.211.01, 12227 MiB",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35"
+  },
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "timestamp_utc": "2026-05-26T05:05:53.718666+00:00",
+  "per_device": {
+    "cpu": {
+      "device": "cpu",
+      "n_images": 50,
+      "n_warmup": 10,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 30.415208999784227,
+      "median_ms": 31.296612500000265,
+      "p95_ms": 32.90803200002301,
+      "mean_ms": 31.463410440046573,
+      "std_ms": 0.8065842891002637
+    },
+    "cuda": {
+      "device": "cuda",
+      "n_images": 50,
+      "n_warmup": 10,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 5.130470000040077,
+      "median_ms": 5.2899059999163,
+      "p95_ms": 6.557743550024497,
+      "mean_ms": 5.513052840015007,
+      "std_ms": 0.46512099244870114
+    }
+  }
+}

reports/eval_harness/patchcore_capsule.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "capsule",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "a35a4767af802024bcbbdb6ef56b1c9074be26490b39d0345378c8b842d69ee6",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_capsule/Patchcore/MVTecAD/capsule/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "25454995713926187e9816613d0e76e8e9531d6ca99becdf2565e8e8ebda8feb",
+  "image_auroc": 0.994415636218588,
+  "pixel_auroc": 0.990176424065145,
+  "aupro": 0.9382407665252686,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.4463476654887199,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 219
+  },
+  "n_test_images": 132,
+  "n_test_anomaly": 109,
+  "n_test_normal": 23,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:02:57.600572+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.01,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.01,
+    "checkpoint_sha256": "25454995713926187e9816613d0e76e8e9531d6ca99becdf2565e8e8ebda8feb",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}

reports/eval_harness/patchcore_capsule_latency.json

@@ -0,0 +1,46 @@
+{
+  "schema": "inspectnet_cx.bench_latency.v1",
+  "category": "capsule",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_capsule/Patchcore/MVTecAD/capsule/v0/weights/lightning/model.ckpt",
+  "dataset_root": "/home/yusuf/datasets/mvtec_ad",
+  "hardware": {
+    "cpu": "AMD Ryzen 9 9900X 12-Core Processor",
+    "gpu": "NVIDIA GeForce RTX 5070, 570.211.01, 12227 MiB",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35"
+  },
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "timestamp_utc": "2026-05-26T05:06:04.986280+00:00",
+  "per_device": {
+    "cpu": {
+      "device": "cpu",
+      "n_images": 50,
+      "n_warmup": 10,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 30.708614000104717,
+      "median_ms": 31.777702999988833,
+      "p95_ms": 42.2283800000514,
+      "mean_ms": 38.632892740024545,
+      "std_ms": 38.97406824647649
+    },
+    "cuda": {
+      "device": "cuda",
+      "n_images": 50,
+      "n_warmup": 10,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 5.108850999931747,
+      "median_ms": 5.353652000167131,
+      "p95_ms": 6.178862999854573,
+      "mean_ms": 5.473933599996599,
+      "std_ms": 0.3758229662202633
+    }
+  }
+}

reports/eval_harness/patchcore_capsule_latency_rerun1.json

@@ -0,0 +1,34 @@
+{
+  "schema": "inspectnet_cx.bench_latency.v1",
+  "category": "capsule",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_capsule/Patchcore/MVTecAD/capsule/v0/weights/lightning/model.ckpt",
+  "dataset_root": "/home/yusuf/datasets/mvtec_ad",
+  "hardware": {
+    "cpu": "AMD Ryzen 9 9900X 12-Core Processor",
+    "gpu": "NVIDIA GeForce RTX 5070, 570.211.01, 12227 MiB",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35"
+  },
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "timestamp_utc": "2026-05-26T05:15:53.993082+00:00",
+  "per_device": {
+    "cpu": {
+      "device": "cpu",
+      "n_images": 50,
+      "n_warmup": 10,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 31.486972000038804,
+      "median_ms": 41.77246400013246,
+      "p95_ms": 44.667347600170615,
+      "mean_ms": 40.962546960063264,
+      "std_ms": 3.56739511177009
+    }
+  }
+}

reports/eval_harness/patchcore_capsule_latency_rerun2.json

@@ -0,0 +1,34 @@
+{
+  "schema": "inspectnet_cx.bench_latency.v1",
+  "category": "capsule",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_capsule/Patchcore/MVTecAD/capsule/v0/weights/lightning/model.ckpt",
+  "dataset_root": "/home/yusuf/datasets/mvtec_ad",
+  "hardware": {
+    "cpu": "AMD Ryzen 9 9900X 12-Core Processor",
+    "gpu": "NVIDIA GeForce RTX 5070, 570.211.01, 12227 MiB",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35"
+  },
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "timestamp_utc": "2026-05-26T05:16:12.213588+00:00",
+  "per_device": {
+    "cpu": {
+      "device": "cpu",
+      "n_images": 50,
+      "n_warmup": 30,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 28.789403999780916,
+      "median_ms": 29.74879599992164,
+      "p95_ms": 32.50230500000271,
+      "mean_ms": 30.1732830000401,
+      "std_ms": 1.1620811502979578
+    }
+  }
+}

reports/eval_harness/patchcore_leather.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "leather",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "4c730802f339675b402c86c8079e1b7b57b07c0b00e6f4de45db073ebb52ad5c",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_leather/Patchcore/MVTecAD/leather/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "5cf7c7a793ad441a9c6cd92ee27517c674df35720c1873e61ef7aab5ebc2bd29",
+  "image_auroc": 1.0,
+  "pixel_auroc": 0.9923970507295065,
+  "aupro": 0.9760316014289856,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.2661692547798157,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 245
+  },
+  "n_test_images": 124,
+  "n_test_anomaly": 92,
+  "n_test_normal": 32,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:03:30.650515+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.01,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.01,
+    "checkpoint_sha256": "5cf7c7a793ad441a9c6cd92ee27517c674df35720c1873e61ef7aab5ebc2bd29",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}

reports/eval_harness/patchcore_leather_latency.json

@@ -0,0 +1,46 @@
+{
+  "schema": "inspectnet_cx.bench_latency.v1",
+  "category": "leather",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_leather/Patchcore/MVTecAD/leather/v0/weights/lightning/model.ckpt",
+  "dataset_root": "/home/yusuf/datasets/mvtec_ad",
+  "hardware": {
+    "cpu": "AMD Ryzen 9 9900X 12-Core Processor",
+    "gpu": "NVIDIA GeForce RTX 5070, 570.211.01, 12227 MiB",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35"
+  },
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "timestamp_utc": "2026-05-26T05:06:15.187105+00:00",
+  "per_device": {
+    "cpu": {
+      "device": "cpu",
+      "n_images": 50,
+      "n_warmup": 10,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 30.973587000062253,
+      "median_ms": 32.81182499995339,
+      "p95_ms": 35.19110755000838,
+      "mean_ms": 32.837616140041064,
+      "std_ms": 1.1777132786393758
+    },
+    "cuda": {
+      "device": "cuda",
+      "n_images": 50,
+      "n_warmup": 10,
+      "image_size": 256,
+      "batch_size": 1,
+      "min_ms": 5.1714010000978305,
+      "median_ms": 5.349747500076774,
+      "p95_ms": 6.364728000062314,
+      "mean_ms": 5.613228779975543,
+      "std_ms": 0.46813156145994267
+    }
+  }
+}

reports/eval_harness/patchcore_leather_seed1.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "leather",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "4c730802f339675b402c86c8079e1b7b57b07c0b00e6f4de45db073ebb52ad5c",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_leather_seed1/Patchcore/MVTecAD/leather/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "5af3f908dae9df60fe472718b588deeaaeb93ce7b7b8d286c9077df098375d65",
+  "image_auroc": 1.0,
+  "pixel_auroc": 0.992077722412704,
+  "aupro": 0.9747775793075562,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.2635792183876038,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 245
+  },
+  "n_test_images": 124,
+  "n_test_anomaly": 92,
+  "n_test_normal": 32,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:19:36.165636+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.01,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.01,
+    "checkpoint_sha256": "5af3f908dae9df60fe472718b588deeaaeb93ce7b7b8d286c9077df098375d65",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}

reports/eval_harness/patchcore_leather_seed2.json

@@ -0,0 +1,47 @@
+{
+  "schema": "inspectnet_cx.eval_harness.v1",
+  "category": "leather",
+  "method": "patchcore",
+  "dataset": "mvtec_ad",
+  "split": "mvtec_ad_default_train_test",
+  "split_hash": "4c730802f339675b402c86c8079e1b7b57b07c0b00e6f4de45db073ebb52ad5c",
+  "checkpoint": "/home/yusuf/Projects/inspectnet-cx/artifacts/patchcore_leather_seed2/Patchcore/MVTecAD/leather/v0/weights/lightning/model.ckpt",
+  "checkpoint_hash": "268b1d0819ef50353a0ed874dc84ce2f38d5fe1686978ed284f881c5532fbc0e",
+  "image_auroc": 1.0,
+  "pixel_auroc": 0.9921353973945378,
+  "aupro": 0.9748720526695251,
+  "aupro_fpr_limit": 0.3,
+  "threshold": 0.26283863246440886,
+  "threshold_selection": {
+    "rule": "quantile",
+    "quantile": 0.995,
+    "split": "train",
+    "n_train_scores": 245
+  },
+  "n_test_images": 124,
+  "n_test_anomaly": 92,
+  "n_test_normal": 32,
+  "library_versions": {
+    "python": "3.10.12",
+    "platform": "Linux-6.8.0-117-generic-x86_64-with-glibc2.35",
+    "numpy": "2.2.6",
+    "sklearn": "1.7.2",
+    "torch": "2.11.0+cu128",
+    "torchvision": "0.26.0+cu128",
+    "anomalib": "2.4.1"
+  },
+  "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea",
+  "timestamp_utc": "2026-05-26T05:20:32.067330+00:00",
+  "device": "cuda",
+  "coreset_sampling_ratio": 0.01,
+  "train_sidecar": {
+    "backbone": "wide_resnet50_2",
+    "layers": [
+      "layer2",
+      "layer3"
+    ],
+    "coreset_sampling_ratio": 0.01,
+    "checkpoint_sha256": "268b1d0819ef50353a0ed874dc84ce2f38d5fe1686978ed284f881c5532fbc0e",
+    "git_commit": "e8b13f0163d67d14f3e3803fab58a834b002d5ea"
+  }
+}