Proof
The numbers behind the claim.
Independently reproducible. Every test run is logged with a Run ID. No cherry-picking — failure modes documented alongside successes.
NEEDLE-IN-HAYSTACK · 100,000 DECOYS
One real watermark.
A hundred thousand fakes.
We seeded a single real watermark into a haystack of 99,999 random decoy IDs. The detector recovered the real one with a 5.38× margin over the strongest decoy, and zero false positives across the rest.
REAL
32/32 bits
TOP DECOY
0.78 acc
MARGIN
5.38×
100,000 DECOYS + 1 REAL
IDLE
Visualization shows a 10K-cell sample of the actual 100K-decoy haystack used in production gates.
DETECTION RUNS
Score-ratio & bit-accuracy across runs.
Baseline (no re-encode) · 100K decoys
✓ DETECTED
Score
2.88
Ratio
5.38×
Real bits
32/32
Confidence
HIGH
H.264 CRF 23 light · 100K decoys
✓ DETECTED
Score
1.63
Ratio
3.37×
Real bits
32/32
Confidence
HIGH
H.264 CRF 28 medium · 100K decoys
✓ DETECTED
Score
1.00
Ratio
2.20×
Real bits
31/32
Confidence
HIGH
H.264 CRF 32 heavy · 100K decoys
✓ DETECTED
Score
0.73
Ratio
1.77×
Real bits
29/32
Confidence
MED
MULTI-CODEC ROBUSTNESS
Survives every transcode we've thrown at it.
Bit accuracy across single-token detection. Anything ≥ 0.81 with a 1.5× score-ratio passes our gate.
SCENARIO
CODEC
BIT ACC
DISTRIBUTION
STATUS
─── threshold at 0.81 bit accuracy
!
ENGINEERING NOTE
The legacy bit-accuracy gate (gap ≥ 0.09) would have failed on the heavy re-encode runs —
top decoy hit 0.813, leaving zero margin. The continuous score-ratio gate (≥ 1.5×) is what saved them.
We kept both in production.