# NULLRADIAL release benchmark — NULLRADIAL-BENCH-R1

> Generic object-detection robustness only. This report does not establish invisibility, identity protection, biometric evasion, plate evasion, or safety-system behavior.

## Outcome

**ember-halo** had the strongest measured radial-specific effect: disruption contrast `+0.0778` versus its exact-histogram spatial-permutation control; scene-cluster bootstrap 95% CI `-0.0409` to `+0.1856`.

Lock decision: **LOCKED_RESEARCH_REVISION** for `NR-EMBER-HALO-R1`. Here “locked” means only that the generator seed, SVG/PNG bytes, file hashes, and benchmark baseline are revision-controlled. It is not a security, privacy, detection-evasion, or road-safety certification.

## Compact metrics

Primary metric: mean confidence disruption including misses (`baseline confidence − matched variant confidence`; a miss receives the full baseline confidence). Positive radial-specific contrast means the radial layout disrupted detection more than its palette/histogram-matched control.

| Rank | Pattern | Radial disruption | Matched control | Radial-specific Δ | 95% scene-bootstrap CI | Radial continuity | Control continuity | YOLO Δ | DETR Δ |
|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|
| 1 | ember-halo | 0.312 | 0.235 | +0.078 | [-0.041, +0.186] | 55.9% | 69.0% | +0.074 | +0.080 |
| 2 | signal-moss | 0.143 | 0.200 | -0.057 | [-0.115, +0.006] | 82.5% | 77.8% | -0.079 | -0.040 |
| 3 | parallax-bloom | 0.202 | 0.263 | -0.060 | [-0.196, +0.069] | 71.7% | 67.5% | -0.027 | -0.086 |

## Detector builds and exact protocol

| Detector | Family/build | Weight SHA-256 | Threshold | Preprocessing |
|---|---|---|---:|---|
| yolov10n | YOLOv10n end-to-end ONNX export; Ultralytics metadata 8.1.34 | `a77dd863933f184a19e84361c64b788228a7c7dacc2c78939239a96ad3efca3b` | 0.25 | RGB; 640x640 aspect-preserving letterbox with value 114; float32 /255; NCHW |
| detr-resnet50-quantized | DETR transformer with ResNet-50 backbone; quantized Xenova ONNX port | `cae09a307ed9247da7e2ce8bcf81522a6817f1ea2e82b9c4dde59f5964b62b4f` | 0.5 | RGB; shortest edge 800, longest <=1333, bilinear; /255; mean [0.485,.456,.406], std [.229,.224,.225]; NCHW; fixed 64x64 int64 ones pixel_mask required by export |

YOLOv10n returns its exported 300×6 end-to-end detections. DETR uses softmax over 92 logits, excludes the final no-object class, decodes normalized center-width-height boxes, and intentionally applies no NMS because DETR is a set-prediction model.

## Paired design

- 4 scenes: two public neutral smoke-test scenes and two cropped user-supplied real patterned-product references.
- Three radial revisions and three exact RGB-histogram controls; `True` for all histogram equality checks.
- 21 eligible detector × scene × pattern paired cells (24 planned; 3 excluded because that detector had zero target detections in the unmodified baseline) and 56 total model/image inferences.
- Identical base image, hand-authored product-surface mask, repeat scale, blend strength, and lossless composite format within every radial/control pair.

## Tracking-oriented stability measurement

Association continuity, same-class matched IoU, and normalized centroid drift treat the baseline and patterned image as a two-frame pair. This tests detection-to-detection association stability; it is **not** a temporal tracker evaluation. Aggregate values are included in `metrics.csv` and the JSON.

## Uncertainty

Intervals use 10,000 resamples with seed `20260830`, clustering at the scene level so both detector measurements from a scene move together. With only four distinct scenes, the intervals are exploratory, discrete, and cannot establish population-level performance. A CI crossing zero means the measured radial advantage over its control is not stable in this sample.

## Supplied-reference baseline observations

The supplied files already depict patterned products and are screenshots, so they are not neutral controls. They are cropped to their photographic content and included as real-reference baselines only.

| Scene | Detector | Generic target detections |
|---|---|---|
| reference-vehicle | yolov10n | car: 1, person: 1, truck: 1 |
| reference-shirt | yolov10n | none |
| reference-vehicle | detr-resnet50-quantized | car: 2 |
| reference-shirt | detr-resnet50-quantized | person: 9 |

## Limitations

- Only four distinct scenes, two detectors, one threshold per detector, and synthetic surface recoloring were tested.
- The two supplied references already contain a pattern and screenshot framing; they are real-reference observations, not neutral before images.
- Hand-authored masks approximate printable surfaces and do not model folds, gloss, compound curvature, weather, distance, motion blur, print gamut, or installation seams.
- The block-permutation control exactly matches RGB histogram and retains local texture, but block boundaries and motif fragments mean it does not isolate radiality perfectly.
- DETR can emit pattern-induced false positives without NMS; these are retained because no NMS is the documented DETR set-prediction path.
- Detection association metrics are two-frame stability proxies, not a real temporal tracker benchmark.
- No face recognition, identity, biometric, license-plate, segmentation model, autonomous-driving system, or safety-critical stack was tested.
- Bootstrap intervals with four scene clusters are too small for population inference or marketing performance guarantees.

## Release files

- `release/report.json` — full machine-readable protocol, detections, pair metrics, statistics, decision.
- `release/metrics.csv` — compact pattern-level metrics.
- `release/lock.json` — exact revision-lock record.
- `release/provenance.json` and `release/PROVENANCE.md` — sources, hashes, and license boundaries.
- `release/artifact_manifest.json` — SHA-256 inventory.