NULLRADIAL LAB / S2 + S3 / 30 AUG 2026
AI invisibility is the research brief. Evidence is the product. Three S2 fields have bounded digital screens; twelve new S3 visual candidates are registered for controlled comparison. None has earned a real-world claim.
S3 / registered visual intake
These image-generated masters expand recursive, quasiperiodic, reaction-diffusion, multiscale, interference, network, and ornate boundary hypotheses. Their hashes and prompts are frozen. Beauty voting is open; matched-control detector screening is pending.
NR–S3–10 S3.V001 / registeredRecursive boundary garden / midnight teal · emerald · coral · saffron. Complexity without a center.
NR–S3–11 S3.V001 / registeredNonrepeating flow basins / oxblood · ultraviolet · cyan · ember. Flow returns. It never repeats.
NR–S3–12 S3.V001 / registeredIrregular jewel tessellation / cobalt · ruby · malachite · amber. Architecture, grown irregular.
NR–S3–13 S3.V001 / registeredBroken-phase floral waves / peacock · magenta · acid · ivory. Several centers. No single target.
NR–S3–14 S3.V001 / registeredQuasiperiodic foliage / emerald · turquoise · rose · violet. Order without ordinary repetition.
NR–S3–15 S3.V001 / registeredMorphogenetic labyrinth / charcoal · bone · teal · coral. Biology writes in negative space.
NR–S3–16 S3.V001 / registeredRecursive couture branching / imperial purple · malachite · copper. Rules grow into ornament.
NR–S3–17 S3.V001 / registeredPhase-shifted line fields / navy · lime · cyclamen · ice. Motion begins before the cloth moves.
NR–S3–18 S3.V001 / registeredMultiscale organic clusters / peat · forest · chartreuse · magenta. The spectrum learns to grow.
NR–S3–19 S3.V001 / registeredInterrupted rings and occluding blooms / cobalt · vermilion · saffron. The halo evolves past symmetry.
NR–S3–20 S3.V001 / registeredNested chromatic networks / cyan · emerald · violet · amber. Every branch carries another world.
NR–S3–21 S3.V001 / registeredBaroque boundary territories / lapis · crimson · peacock · celadon. Ornament meets the edge of chaos.
The prompts requested seamless-style continuity, but measured opposing-edge RMSE was nonzero for all twelve generated sources. NULLRADIAL therefore labels them seam-variable candidates—not exact repeats. S3 preregisters continuous, phase-shifted, and panel-composed variants so seam behavior is tested instead of assumed.
Preview all twelve on product + voteThe commercial thesis
The literature shows that large, optimized physical patterns can change outcomes on particular detectors under particular conditions. It also shows brittle transfer, side-view failures, range limits, metric reversals, and a gulf between a digital render and cloth in motion. NULLRADIAL is building the versioned candidate, manufacturing, and qualification layer between those facts and a sellable surface system.
S2 / release candidates
Each 2,048-pixel master has exact opposing edges and a SHA-256 identity. The screen compares candidate layouts with smooth, pixel-multiset-matched spectral surrogates. That comparison does not isolate radiality causally.
NR–RC–07 S2.0 / exact repeatLarge radial florals and feather lattices carrying dense couture-scale detail.
NR–RC–08 S2.0 / exact repeatIrregular concentric rosettes with fractured rings and a deliberately uneven scale hierarchy.
NR–RC–09 S2.0 / exact repeatA botanical canopy interlocked with fractured radial geometry and etched filigree.
S2 used four scenes and two detector families. Two scenes were already-patterned screenshots. Detector outputs—not independent human annotations—served as the pseudo-baseline. The statistics prioritize swatches; they do not estimate product recall, precision, AP, tracking, identity, or safety performance.
Question / does bigger win?
Across all three candidates, macro repeats had the largest mean point estimate. The single strongest candidate-scale cell was Broken Halo at micro scale. Standard scale split the detector-family means. The honest answer is candidate-specific—and physical pixel footprint must be held equal before treating size as causal.
| Scale | Factor | Mean layout contrast | YOLO | DETR | Agreement |
|---|---|---|---|---|---|
| Macro | 1.80× | +0.050 | +0.067 | +0.038 | Both family means positive |
| Micro | 0.55× | +0.037 | +0.025 | +0.046 | Both family means positive |
| Standard | 1.00× | +0.034 | −0.038 | +0.089 | Detector means disagree |
They make roll goods, scale comparisons, cut maps, and version checks cleaner. Every S2 master has zero opposite-edge pixel error.
No core apparel study establishes that an invisible image seam survives garment seams, feed error, stretch, folds, abrasion, or laundering.
Ornate candidates produced measurable but inconsistent digital differences. Ornament itself is not the cause, and physical transfer remains open.
Evidence ladder
No percentage travels without its candidate, material, model, camera, conditions, controls, sample unit, and uncertainty. A changed dependency creates a new attempt—not a better past.
A literature-backed candidate with locked art. No NULLRADIAL performance result.
Synthetic comparison on named frozen builds. Useful for prioritization, not a product claim.
Held-out scenes, multiple controls, tasks, transformations, and at least three model families.
Printed lot tested across fit, folds, seams, light, motion, distance, wash, and weather.
Preregistered study with independent ground truth, clustered statistics, and all failures.
New data and blinded samples tested by a team that did not optimize the candidate.
Exact SKU, BOM, lot, process, tolerances, retest triggers, and bounded marketing claim.
The label we can earn next
There is no universal AI-invisibility certification. The next defensible stage is narrower: “Protocol-ready for independent, controlled real-world validation under a named study.” It means the sample and protocol are ready to test—not that the product works or carries an industry certification.
Propose a replicationOne candidate, scale, substrate, construction, print lot, primary endpoint, and analysis.
New consenting participants or mannequins; new scenes; three cameras; light, view, motion, folds, seams, and distance.
Independent ground truth, four frozen builds across at least three model families, clustered statistics, and every failure.
A separate team receives coded samples from a new lot and runs the sealed protocol without outcome-driven changes.
Independent tester map
No organization listed has agreed to test or endorse NULLRADIAL. The exact protocol, accreditation scope, sample custody, raw-data rights, and publication of null results must be confirmed in writing.
Priority inquiry for blinded physical AI performance and robustness evaluation.
OPTICAL / MARYLANDSponsored independent replication route with new data and publishable failures.
RF / GLOBALPlanar shielding route; exact laboratory, method scope, lot plan, and uncertainty required.
TEXTILE / BELGIUMDurability track across laundering, stretch, abrasion, flex, moisture, and lot variation.
Preliminary IP direction / details held for counsel
A non-exhaustive landscape found close prior art around adversarial clothing, tiled motifs, seam-spanning graphics, multi-model testing, silver textiles, and provenance ledgers. Four narrower manufacturing and qualification invention-disclosure tracks are now on counsel hold. These are research hypotheses—not patentability opinions. Enabling details, alternatives, and claim structures are intentionally absent from this public page pending licensed patent counsel and a filing decision.
Primary sources / inspect the frontier
S2 + S3 / public artifacts