# NULLRADIAL release-candidate screen S2

> Exploratory digital screening only. This is not proof of AI invisibility, a physical-product result, a privacy guarantee, or a certification.

## What ran

- 152 inferences across 2 pinned detector families, 4 scenes, 3 candidates, and 3 repeat scales.
- Every candidate is compared with a smooth spectral-rank surrogate that preserves the source master’s complete RGB pixel multiset and exact repeat edges.
- The same four R1 scenes are reused; results select physical-test priorities and cannot establish population performance.

## Scale comparison

| Scale | Repeat factor | Mean candidate-minus-surrogate contrast | YOLO mean | DETR mean |
|---|---:|---:|---:|---:|
| macro | 1.80× | +0.050 | +0.067 | +0.038 |
| micro | 0.55× | +0.037 | +0.025 | +0.046 |
| standard | 1.00× | +0.034 | -0.038 | +0.089 |

## Candidate × scale results

| Rank | Candidate | Scale | Cells | Candidate-minus-surrogate contrast | 95% scene CI | YOLO | DETR |
|---:|---|---|---:|---:|---|---:|---:|
| 1 | Broken Halo | micro | 7 | +0.133 | [+0.038, +0.338] | +0.092 | +0.164 |
| 2 | Prism Canopy | macro | 7 | +0.066 | [-0.042, +0.210] | +0.075 | +0.059 |
| 3 | Broken Halo | standard | 7 | +0.059 | [+0.015, +0.160] | -0.003 | +0.105 |
| 4 | Lattice Bloom | macro | 7 | +0.043 | [-0.082, +0.210] | -0.061 | +0.121 |
| 5 | Broken Halo | macro | 7 | +0.042 | [-0.049, +0.132] | +0.186 | -0.067 |
| 6 | Lattice Bloom | micro | 7 | +0.033 | [-0.027, +0.101] | +0.050 | +0.020 |
| 7 | Lattice Bloom | standard | 7 | +0.028 | [-0.048, +0.170] | -0.021 | +0.065 |
| 8 | Prism Canopy | standard | 7 | +0.016 | [-0.090, +0.096] | -0.092 | +0.097 |
| 9 | Prism Canopy | micro | 7 | -0.055 | [-0.102, +0.008] | -0.066 | -0.046 |

## Decision boundary

S2 is a locked exploratory digital screen. It can prioritize physical swatches; it cannot qualify real-world performance, certify AI invisibility, or justify a privacy guarantee.

All three locked artworks remain candidate textures. Any post-screen artwork change requires a new ID. The next valid gate is preregistered physical cloth testing with new scenes, cameras, participants or mannequins, print lots, folds, seams, motion, and independent replication.

## Limitations

- Only four scenes and two detector families were used; four scene clusters are too few for population inference.
- The two reference screenshots already contain patterns and are not neutral before-images.
- Synthetic recoloring does not reproduce cloth drape, print gamut, folds, seams, motion, weather, wash, or camera recapture.
- A mirrored exact-repeat transform introduces bilateral structure and does not prove that seamlessness causes any detector change.
- Candidate and control master pixels match exactly as multisets, but resizing, tiling, masking, and shading mean final composite histograms are not identical.
- The surrogate reduces block-grid confounding; it still does not isolate ornament, radiality, motif semantics, or any one visual property causally.
- Detector predictions are used as a pseudo-baseline rather than independent human ground truth; false positives, baseline misses, recall, precision, and AP therefore cannot be resolved.
- Scale comparisons are exploratory and were not powered or preregistered for a confirmatory claim.
- No identity, face, biometric, tracking, segmentation, pose, infrared, thermal, RF, or safety-critical system was tested.
- No candidate is certified, validated for real-world use, or guaranteed to alter a new model or camera.