Trace · Report
Trace Robustness Report
We mark test content for a known buyer, attack it the way leaks really happen, and publish what Trace said. Every cell, including the misses.
- Our own results, not yet independently verified
- Every miss is listed
- Raw JSON for every table
A per-fan mark is only useful if it is still readable after a leak, and if Trace refuses to name anyone when it isn’t. This report shows both. For each kind of content we list the attack, how many copies we tested, what Trace said, and why.
Wrong buyers named
0
Across every attack in this run. When the evidence was too weak, Trace named nobody. That shows up below as “inconclusive”: safe, but the leak goes unattributed.
Verification status
These are our own results. No independent lab has verified them yet. We are inviting labs to re-run them.
Results at a glance
Four kinds of content, every attack listed.
Video (one copy per fan)
18 / 19 attacks traced
1 inconclusive (nobody named). 0 wrong buyers named. Collusion: 2 of 2 named nobody. No unmarked video controls in this run.
Mass message and streamed video
18 / 19 attacks traced
1 inconclusive (nobody named). 0 wrong buyers named. Collusion: 2 of 2 named nobody. No unmarked video controls in this run.
Photos
84 / 84 copies traced
21 attacks × 4 files. 0 wrong buyers named. 84 of 84 unmarked controls named nobody.
Voice notes
6 / 8 attacks traced
2 inconclusive (nobody named). 0 wrong buyers named. 7 of 7 unmarked controls named nobody.
“Traced” means Trace matched the leak to the copy delivered to the right test buyer and the match cleared the gate. “Inconclusive” means it named nobody.
Leak test, 18 Nov 2026
Screen recordings, chat apps and edits: 63 ways a video leaks.
We marked a 64-second video for a known test buyer on a list of 1,000, then put it through 63 realistic leaks. Trace showed the right buyer in 58, named nobody in 5, and named the wrong buyer in 0.
Built-in screen recorders
iPhone, Android, Mac, Windows
Portrait and landscape phone recordings, window recordings and full-screen recordings, raw and then cut down and re-uploaded. These are software reproductions of each recorder's resolution, frame rate and bitrate, not recordings made on the devices. Real-device recordings are being added.
Inconclusive results
5 of 63
Cropped to fill an upright TikTok screen, turned into a GIF, zoomed in 1.6×, shrunk into a corner (picture in picture), and the simulated filmed-off-a-screen case. In each one the result was inconclusive and no buyer was shown, which is the safe outcome.
Every leak and what Trace said
| Leak | Trace said | Copies | Detail |
|---|---|---|---|
| Built-in screen recorders | |||
| iPhone screen recording, phone upright (1170×2532, 60 fps, 8 Mbps) | Right buyer | 1 | margin 14.04 (lined up with the original first) |
| Same, then cut down to the video and re-uploaded | Right buyer | 1 | margin 12.02 |
| iPhone screen recording, phone sideways (2532×1170, 60 fps) | Right buyer | 1 | margin 12.71 |
| Same, then cut down and re-uploaded | Right buyer | 1 | margin 11.96 |
| Android screen recording, phone upright (1080×2400, 30 fps) | Right buyer | 1 | margin 14.31 (lined up with the original first) |
| Same, then cut down and re-uploaded | Right buyer | 1 | margin 12.46 |
| Android screen recording, low quality setting (720×1600, 4 Mbps) | Right buyer | 1 | margin 14.17 (lined up with the original first) |
| Same, then cut down and re-uploaded | Right buyer | 1 | margin 12.4 |
| Android screen recording, phone sideways (2400×1080) | Right buyer | 1 | margin 12.29 |
| Same, then cut down and re-uploaded | Right buyer | 1 | margin 12.19 |
| Mac screen recording of a browser window (2560×1600, 60 fps) | Right buyer | 1 | margin 13.83 (lined up with the original first) |
| Same, then cut down and re-uploaded | Right buyer | 1 | margin 12.1 |
| Windows Game Bar, video full screen (1080p) | Right buyer | 1 | margin 14.21 |
| Windows Game Bar, video playing in a window | Right buyer | 1 | margin 12.9 (lined up with the original first) |
| Windows Snipping Tool recording, small window | Right buyer | 1 | margin 10.33 (lined up with the original first) |
| Generic screen recording (1080p pixel copy, re-encoded) | Right buyer | 1 | margin 14.21 |
| Where leaks get posted | |||
| Telegram (720p, 1.5 Mbps) | Right buyer | 1 | margin 12.38 |
| WhatsApp (480p, 0.9 Mbps) | Right buyer | 1 | margin 12.9 |
| Discord, squeezed under 8 MB | Right buyer | 1 | margin 11.98 |
| X re-encode (720p) | Right buyer | 1 | margin 12.69 |
| Reddit re-encode (720p) | Right buyer | 1 | margin 12.92 |
| Re-uploaded twice (540p, then re-compressed) | Right buyer | 1 | margin 9.23 |
| iPhone screen recording, cut down, then sent on Telegram | Right buyer | 1 | margin 11.5 |
| Android screen recording, cut down, then sent on WhatsApp | Right buyer | 1 | margin 11.92 |
| TikTok / Reels upright, video letterboxed | Right buyer | 1 | margin 12.08 (lined up with the original first) |
| TikTok / Reels upright, video cropped to fill the screen | Nobody named | 1 | margin 1.9 |
| Turned into a GIF (480 px, 12 fps, 20 s) | Nobody named | 1 | margin 3.4 (lined up with the original first) |
| Edits | |||
| 5-second clip | Right buyer | 1 | margin 9.35 |
| 10-second clip | Right buyer | 1 | margin 7.23 |
| 15-second clip | Right buyer | 1 | margin 6.85 |
| 30-second clip from the middle | Right buyer | 1 | margin 8.6 |
| First 7 seconds cut off | Right buyer | 1 | margin 12.23 |
| 15 seconds inside a compilation | Right buyer | 1 | margin 7.38 |
| Slowed to 0.9× | Right buyer | 1 | margin 13.69 |
| Sped up 1.05× | Right buyer | 1 | margin 13.56 |
| Sped up 1.25× | Right buyer | 1 | margin 13.65 |
| Played backwards | Right buyer | 1 | margin 5.1 |
| Converted to 24 fps | Right buyer | 1 | margin 13.46 |
| Converted to 60 fps | Right buyer | 1 | margin 13.58 |
| Cropped 10% and resized | Right buyer | 1 | margin 10.96 |
| Zoomed in 1.3× | Right buyer | 1 | margin 6.88 |
| Zoomed in 1.6× | Nobody named | 1 | margin 1.94 |
| Squashed into a letterbox | Right buyer | 1 | margin 9.62 |
| Rotated 90° | Right buyer | 1 | margin 10.35 |
| Mirrored | Right buyer | 1 | margin 12.6 |
| Mirrored, cropped 15% and colour changed | Right buyer | 1 | margin 8.96 (lined up with the original first) |
| Brightness, contrast and saturation changed | Right buyer | 1 | margin 10.92 |
| Black and white | Right buyer | 1 | margin 14.23 |
| Border and text added | Right buyer | 1 | margin 8.4 |
| Box covering part of the frame | Right buyer | 1 | margin 13.58 |
| Shrunk into a corner over other video (picture in picture) | Nobody named | 1 | margin 1.98 |
| Noise reduction | Right buyer | 1 | margin 11.29 |
| Sharpened | Right buyer | 1 | margin 13.08 |
| Blurred | Right buyer | 1 | margin 13.94 |
| Upscaled to 4K | Not run | 0 | |
| Downscaled to 480p | Right buyer | 1 | margin 12.0 |
| Downscaled to 360p | Right buyer | 1 | margin 10.88 |
| Downscaled to 240p | Right buyer | 1 | margin 9.08 |
| Re-compressed, normal | Right buyer | 1 | margin 13.77 |
| Re-compressed, heavy | Right buyer | 1 | margin 11.67 |
| Re-compressed, very heavy | Right buyer | 1 | margin 8.81 |
| Re-compressed, extreme | Right buyer | 1 | margin 6.67 |
| Filmed-off-a-screen effect (angle and moiré), simulated | Nobody named | 1 | margin 1.62 |
| No change (baseline) | Right buyer | 1 | margin 14.77 |
Method: one 64-second 720p test video, marked for test buyer 4321 on a list of 1,000 buyers. Each leak was made with ffmpeg from the marked copy and read the way production reads it: a plain read first, then a read after lining the copy up with the original. Trace reports a buyer only when the match margin is at least 2.0. Controls: 4 of 4 unmarked copies, put through the same leaks, matched nobody. Recordings with a black band or page around the video (upright phone, browser window) are read by first finding the moving picture in the frame; that step ships with the next engine update. Upscaled to 4K was not scored: reading 4K frames ran our 8 GB test machine out of memory, and we are fixing the reader to handle it. Raw results: leak_matrix.json.
Video, one copy per fan
Each fan gets their own marked copy.
Each fan gets their own marked copy of a PPV, custom or DM video. This is the default for creator Trace.
Every attack and what Trace said
One 64-second 720p test video made from sample clips and moving synthetic footage. One marked copy for the target fan, matched against a roster of 1,000 fan IDs. Each attack was run once.
| Attack | Result | Copies tested | Detail |
|---|---|---|---|
| Marked copies (should match the right buyer) | |||
| Saved and re-shared as is | Traced | 1 | margin 14.8 |
| Re-encoded (H.264 CRF 23) | Traced | 1 | margin 13.9 |
| Re-encoded harder (CRF 30) | Traced | 1 | margin 11.5 |
| Re-encoded very hard (CRF 35) | Traced | 1 | margin 8.9 |
| Downscaled to 480p | Traced | 1 | margin 12.0 |
| Downscaled to 360p | Traced | 1 | margin 10.8 |
| Frame rate changed to 24 fps | Traced | 1 | margin 13.7 |
| Sped up 5% | Traced | 1 | margin 13.7 |
| First 7 seconds cut off | Traced | 1 | margin 12.2 |
| 30-second clip cut from the middle | Traced | 1 | margin 8.7 |
| Cropped 10% and scaled back up | Traced | 1 | margin 10.8 |
| Letterboxed (black bars) | Traced | 1 | margin 9.5 |
| Mirrored | Traced | 1 | margin 12.3 |
| Contrast, brightness and saturation changed | Traced | 1 | margin 10.8 |
| Turned black and white | Traced | 1 | margin 14.3 |
| Box or sticker over part of the picture | Traced | 1 | margin 13.4 |
| Screen recording (simulated software capture, 1080p) | Traced | 1 | margin 14.2 |
| Phone filming a tilted screen, with moiré (simulated) | Inconclusive | 1 | Readings pointed to more than one ID, so nobody named |
| Social media re-upload (re-encoded twice) | Traced | 1 | margin 9.0 |
| Collusion (naming either buyer is correct; naming nobody is safe) | |||
| Two buyers blend their copies together | Nobody named (safe) | 1 | Readings pointed to more than one ID, so nobody named |
| Two buyers splice half of each copy | Nobody named (safe) | 1 | Readings pointed to more than one ID, so nobody named · aligned to the original first |
“Margin” is how far the best-matching fan’s reading beat the next best. Production settings require at least 2.0 before Trace reports a match. Collusion here means two fans combining their copies to hide whose copy leaked; naming either of them would count as correct, naming nobody is the safe outcome.
Mass message and streamed video
One file to many fans, still traceable to a list.
A mass message sends one file to many fans, so Trace builds the copies from short segments of two marked versions, in a pattern unique to each recipient or list. In this bench, 1,000 recipients each had their own pattern.
Every attack and what Trace said
The same 64-second video, cut into 1-second segments with two marked versions of each. 1,000 recipients each got their own pattern of segments. Each attack was run once on the target recipient's copy.
| Attack | Result | Copies tested | Detail |
|---|---|---|---|
| Marked copies (should match the right buyer) | |||
| Saved and re-shared as is | Traced | 1 | chance match under 1 in 1016 · 64 segments read |
| Re-encoded (H.264 CRF 23) | Traced | 1 | chance match under 1 in 1016 · 64 segments read |
| Re-encoded harder (CRF 30) | Traced | 1 | chance match under 1 in 1012 · 64 segments read |
| Re-encoded very hard (CRF 35) | Traced | 1 | chance match under 1 in 107 · 64 segments read |
| Downscaled to 480p | Traced | 1 | chance match under 1 in 1016 · 64 segments read |
| Downscaled to 360p | Traced | 1 | chance match under 1 in 109 · 64 segments read |
| Frame rate changed to 24 fps | Traced | 1 | chance match under 1 in 1016 · 64 segments read |
| Sped up 5% | Traced | 1 | chance match under 1 in 1016 · 64 segments read |
| First 7 seconds cut off | Traced | 1 | chance match under 1 in 1014 · 57 segments read |
| 30-second clip cut from the middle | Traced | 1 | chance match under 1 in 106 · 30 segments read |
| Cropped 10% and scaled back up | Traced | 1 | chance match under 1 in 1016 · 64 segments read |
| Letterboxed (black bars) | Traced | 1 | chance match under 1 in 1012 · 64 segments read |
| Mirrored | Traced | 1 | chance match under 1 in 1015 · 64 segments read · aligned to the original first |
| Contrast, brightness and saturation changed | Traced | 1 | chance match under 1 in 1011 · 64 segments read |
| Turned black and white | Traced | 1 | chance match under 1 in 1016 · 64 segments read |
| Box or sticker over part of the picture | Traced | 1 | chance match under 1 in 1016 · 64 segments read |
| Screen recording (simulated software capture, 1080p) | Traced | 1 | chance match under 1 in 1016 · 64 segments read |
| Phone filming a tilted screen, with moiré (simulated) | Inconclusive | 1 | Too weak to name anyone |
| Social media re-upload (re-encoded twice) | Traced | 1 | chance match under 1 in 1011 · 64 segments read |
| Collusion (naming either buyer is correct; naming nobody is safe) | |||
| Two buyers blend their copies together | Nobody named (safe) | 1 | Too weak to name anyone · aligned to the original first |
| Two buyers splice half of each copy | Nobody named (safe) | 1 | Too weak to name anyone · aligned to the original first |
A match must be less likely than one in a million to happen by chance across everyone who received the content. The detail column shows how far under that line each match came. In the creator app, a mass message uses 2 to 64 list versions, so a match points to a version and the list it went to, not to one fan on that list. How to trace a mass message leak.
Photos
The most attacks, because photos leak most.
Each fan gets their own marked copy of a photo. Photos leak as screenshots, crops and re-saves more than anything else, so the photo bench has the most attacks.
Every attack and what Trace said
4 test images (still frames taken from sample videos), each marked for the target fan. Every attack was applied to all 4, and to the 4 unmarked originals as controls. Any ID other than the target fan's would have counted as a wrong name. Matching used the creator's original as a reference, as Trace does in production.
| Attack | Result | Copies tested | Detail |
|---|---|---|---|
| Marked copies (should match the right buyer) | |||
| Saved and re-shared as is | 4 / 4 traced | 4 | |
| Re-saved as JPEG, quality 90 | 4 / 4 traced | 4 | |
| Re-saved as JPEG, quality 60 | 4 / 4 traced | 4 | |
| Re-saved as JPEG, quality 40 | 4 / 4 traced | 4 | |
| Converted to WebP, quality 70 | 4 / 4 traced | 4 | |
| Resized to half size | 4 / 4 traced | 4 | |
| Resized to 1080 wide, JPEG 80 | 4 / 4 traced | 4 | |
| Cropped 10% | 4 / 4 traced | 4 | |
| Cropped 25% | 4 / 4 traced | 4 | 3 of 4 aligned to the original first |
| Phone screenshot (black bars, rescaled) | 4 / 4 traced | 4 | 4 of 4 aligned to the original first |
| Social media re-compression, twice | 4 / 4 traced | 4 | |
| Brightness and contrast changed | 4 / 4 traced | 4 | |
| Saturation filter | 4 / 4 traced | 4 | |
| Turned black and white | 4 / 4 traced | 4 | |
| Blurred | 4 / 4 traced | 4 | |
| Mirrored | 4 / 4 traced | 4 | |
| Rotated 3 degrees | 4 / 4 traced | 4 | 4 of 4 aligned to the original first |
| Box or sticker over part of the picture | 4 / 4 traced | 4 | |
| Photo of the screen with a phone (simulated) | 4 / 4 traced | 4 | 4 of 4 aligned to the original first |
| Screenshot, cropped, re-saved as JPEG | 4 / 4 traced | 4 | 4 of 4 aligned to the original first |
| Mirrored, cropped and rotated | 4 / 4 traced | 4 | |
| Unmarked controls (must name nobody) | |||
| Saved and re-shared as is | 4 / 4 named nobody | 4 | |
| Re-saved as JPEG, quality 90 | 4 / 4 named nobody | 4 | |
| Re-saved as JPEG, quality 60 | 4 / 4 named nobody | 4 | |
| Re-saved as JPEG, quality 40 | 4 / 4 named nobody | 4 | |
| Converted to WebP, quality 70 | 4 / 4 named nobody | 4 | |
| Resized to half size | 4 / 4 named nobody | 4 | |
| Resized to 1080 wide, JPEG 80 | 4 / 4 named nobody | 4 | |
| Cropped 10% | 4 / 4 named nobody | 4 | |
| Cropped 25% | 4 / 4 named nobody | 4 | |
| Phone screenshot (black bars, rescaled) | 4 / 4 named nobody | 4 | |
| Social media re-compression, twice | 4 / 4 named nobody | 4 | |
| Brightness and contrast changed | 4 / 4 named nobody | 4 | |
| Saturation filter | 4 / 4 named nobody | 4 | |
| Turned black and white | 4 / 4 named nobody | 4 | |
| Blurred | 4 / 4 named nobody | 4 | |
| Mirrored | 4 / 4 named nobody | 4 | |
| Rotated 3 degrees | 4 / 4 named nobody | 4 | |
| Box or sticker over part of the picture | 4 / 4 named nobody | 4 | |
| Photo of the screen with a phone (simulated) | 4 / 4 named nobody | 4 | |
| Screenshot, cropped, re-saved as JPEG | 4 / 4 named nobody | 4 | |
| Mirrored, cropped and rotated | 4 / 4 named nobody | 4 | |
“Aligned to the original first” means the first plain read failed, so Trace lined the leak up against your original (undoing the crop, rotation or screenshot bars) and read it again. That is how production works too. Collusion between two fans was not part of the photo bench in this run.
Voice notes
Their own audio mark, tested against re-encodes and clips.
Voice notes get their own per-fan audio mark. They usually leak re-encoded by a messaging app, or clipped short.
Every attack and what Trace said
One speech clip, marked for the target fan and matched against a roster of about 1,000 fan IDs. Each attack was run once. The unmarked clip went through 7 of the 8 attacks as controls.
| Attack | Result | Copies tested | Detail |
|---|---|---|---|
| Marked copies (should match the right buyer) | |||
| Saved and re-shared as is (AAC 128 kbps) | Traced | 1 | detector score 1.00 |
| Converted to MP3, 64 kbps | Traced | 1 | detector score 1.00 |
| Converted to Opus, 24 kbps | Traced | 1 | detector score 0.96 |
| Converted to AAC, 48 kbps | Traced | 1 | detector score 0.90 |
| 6-second clip | Traced | 1 | detector score 0.98 |
| 3-second clip | Traced | 1 | detector score 0.96 |
| Sped up 10% | Inconclusive | 1 | No mark found · detector score 0.34 |
| Telephone-quality audio (8 kHz) | Inconclusive | 1 | Readings disagreed, so nobody named · detector score 0.85 |
| Unmarked controls (must name nobody) | |||
| Saved and re-shared as is (AAC 128 kbps) | Named nobody | 1 | No mark found |
| Converted to MP3, 64 kbps | Named nobody | 1 | No mark found |
| Converted to Opus, 24 kbps | Named nobody | 1 | No mark found |
| Converted to AAC, 48 kbps | Named nobody | 1 | No mark found |
| 6-second clip | Named nobody | 1 | No mark found |
| 3-second clip | Named nobody | 1 | No mark found |
| Sped up 10% | Named nobody | 1 | No mark found |
The two misses were both inconclusive, not wrong: a voice note sped up 10%, and telephone-quality (8 kHz) audio. The detector score is how confident the audio detector was that a mark is present at all; Trace also requires its readings across the clip to agree.
Extra lab check, 1 Nov 2026
Overwriting the mark with a public watermark model.
An attacker who knows which public watermark model a service uses can stamp a new mark over the old one. We tried it on per-fan video.
Results for each layer
First layer.
Lost in 2 of 2 clips. Trace named nobody from it.
Keyed second layer.
Still showed the right buyer in 2 of 2 clips, with no bit errors, and again after re-compressing the overwritten copy. It can only be read with the creator's secret key.
Control.
An unmarked clip put through the same overwrite matched nobody.
Public test suite, 7 Nov 2026
Attacks we didn't choose: the WAVES benchmark.
WAVES is a published image-watermark attack suite from university researchers (ICML 2024). We ran its 8 distortions at 5 strengths, plus its 4 combinations, on marked photos.
Results: wrong accounts, what held up, weak spots
Never the wrong account.
480 attacked marked copies and 480 unmarked controls: 0 matched to the wrong buyer, 0 false matches. When a mark couldn't be read, no account was shown.
Holds up.
Rotation up to 45°, brightness and contrast up to 2×, noise, JPEG down to quality 26, erasing up to 5% of the image and blur up to 4 px: read in all or nearly all cases.
Weak spots.
WAVES' crop-and-stretch-back-to-full-size (about half read, even at a 10% crop; a plain crop is fine, see the photo table above), blur of 8 px or more, erasing over 10% of the image, and the combined attacks. These stay outside the tested envelope until a later run shows they pass.
AI edit test, 16 Nov 2026
Light AI edits now trace. Heavy redraws still win.
We ran marked photos through the AI tools people use to clean up, upscale or redraw a picture, with every Trace layer on. Then we raised the strength of the photo’s low-frequency mark and ran it again. A light AI redraw went from tracing 0 of 8 photos to 6 of 8.
Every AI edit and what Trace said
| AI edit | Before (strength 3) | Now (strength 12) |
|---|---|---|
| No attack; JPEG quality 75 | 8/8 | 8/8 |
| AI re-encode (TAESD, Stable Diffusion VAE) | 8/8 | 8/8 |
| AI upscale 2x and 4x (Real-ESRGAN); face restore (GFPGAN) | 8/8 | 8/8 |
| Light AI redraw, 10% (SD-Turbo, photo kept at its size) | 0/8 | 6/8 |
| AI redraw, 20% | 0/8 | 2/8 |
| AI redraw, 30% | 0/8 | 1/8 |
| Shrunk to 512 px, then a 10% redraw | 0/8 | 3/8 |
| Shrunk to 512 px, then a 20% redraw | 0/8 | 0/8 |
| SDXL-Turbo redraw, 20% | 0/8 | 0/8 |
Never the wrong account.
Every cell, at every strength: 0 copies matched to the wrong buyer, and 0 unmarked controls matched anyone. A copy that couldn’t be read showed no account.
What changed.
Photo copies made from 16 Nov 2026 carry a stronger low-frequency layer: a few broad, smooth changes in brightness. Side by side we could not see a difference. Copies made before keep the strength they were made with. Video is unchanged.
Still weak.
A redraw of 20% or more, a redraw after shrinking the photo, and SDXL-Turbo mostly beat every layer. A heavy redraw is a new picture that looks like yours, so Trace says inconclusive.
Our own run on a GPU: 8 photos (1,024 px), 8 trials per cell, every Trace layer on, each copy matched against 1,000 candidate buyers. Traced means the right buyer with the checks passed. At an in-between strength of 8, a light redraw traced 4 of 8. Redraws used SD-Turbo and SDXL-Turbo image-to-image with a neutral prompt; the percentage is how much of the picture the model redraws. Raw JSON, every trial, photos numbered.
Edited leaks test, 17 Nov 2026
They edited the leak. Trace still showed the account.
Leaked copies get edited to hide where they came from: someone else's face pasted on, part of the picture repainted by AI, a screenshot, a crop and re-post, another watermark on top. We ran each of those on marked photos. 140 of 140 edited copies showed which account the leaked copy was delivered to, with no wrong account.
Every edit and what Trace said
| Edit to the leaked copy | Right account | No account (inconclusive) | Wrong account |
|---|---|---|---|
| No edit (baseline) | 20/20 | 0 | 0 |
| Another person's face swapped onto it (inswapper) | 20/20 | 0 | 0 |
| Face swap, then saved as JPEG 80 | 20/20 | 0 | 0 |
| 30% of the picture repainted by AI (Stable Diffusion 1.5 inpainting) | 20/20 | 0 | 0 |
| 50% of the picture repainted by AI (Stable Diffusion 1.5 inpainting) | 20/20 | 0 | 0 |
| Re-posted: 5% cropped off each edge, JPEG 70 | 20/20 | 0 | 0 |
| Screenshot: 80% size, JPEG 85 | 20/20 | 0 | 0 |
| A second watermark stamped on top (TrustMark) | 20/20 | 0 | 0 |
| Whole picture redrawn by AI (noisy upscaling, Honig et al. 2025) | 5/20 | 15 | 0 |
Never the wrong account.
180 tests, 0 copies matched to the wrong buyer. Every edit was also run on an unmarked copy: none of those matched anyone.
Why edits don't hide it.
Each buyer's copy carries its mark across the whole picture, not in one spot. Replace the face or half the frame and the rest still carries enough to read.
Still weak.
A whole-picture AI redraw makes a new image that looks like yours: 5 of 20 traced, the rest inconclusive. Video and phone filming of a screen are being tested now and are not covered by these numbers.
Our own run on a GPU: 20 synthetic people (no real people), 20 trials per edit, every Trace photo layer on, each copy matched against 1,000 candidate buyers under the production rule (chance of a false match below one in a million). Face swaps used other synthetic people. Inpainting used a neutral prompt. Traced means the right buyer with the checks passed.
What we did not test
A bench only covers what it covers.
These were not part of this run:
What this run didn't cover
Heavy AI regeneration.
Photo redraws are in the AI edit test. Other image models, and AI regeneration of video or voice notes, are not tested yet.
Three or more fans colluding.
We tested two fans combining copies for video and mass message. Not more than two, and not for photos or voice notes.
Heavier crops.
More than 25% off a photo or 10% off a video, or zooming into one small part of the frame.
Real phone cameras.
Photos of a screen and phone filming were simulated in software. An earlier lab run with live phone capture, shown on the Trace page, had a lower pass rate.
Unmarked video controls.
Photos and voice notes had unmarked controls in this run; video did not.
Audio played through a speaker and re-recorded
, music, and long voice notes.
Several attacks stacked
beyond the combinations listed in the tables.
A wide range of content.
One test video, 4 test images and one voice clip. Very dark, very static or very short content may behave differently.
Honest limits
What can remove a mark, and what a match means.
Seven limits to know before you rely on a match
AI regeneration can remove marks.
A light AI edit of a photo usually keeps the mark on copies made since 16 Nov 2026. A heavier redraw is a new picture that looks like yours, and the mark may not come with it. Trace will say inconclusive.
Collusion hides whose copy leaked.
When two fans blended or spliced their copies, Trace named nobody. That is the safe result, but you don’t learn whose copy it was. More fans working together make it harder still.
Heavy crops and tiny clips carry less mark.
Less picture or fewer seconds means less to read. A 30-second clip of a mass message passed in this run, only just.
A phone filming a screen often won’t name anyone.
Video filmed off a tilted screen was inconclusive in this run. Screen recording software is a different case: it copies the pixels, and it was traced.
No mark, no match.
Content you sent before you started marking can’t be traced.
A no-match is not proof.
Inconclusive means the evidence was too weak. It does not show that a fan didn’t leak.
A match is about the copy, not the person.
It tells you which account the leaked copy was delivered to. Accounts get shared and hacked.
How the bench works
Mark, attack, match, score.
How each result is scored
Mark.
Test content is marked for a known target buyer with the same code and settings Trace uses in production.
Attack.
Each attack in the tables is applied to the marked copy: re-encodes, resizes, crops, filters, screenshots, simulated screen capture and filming, clips, speed changes and collusion between two buyers.
Match.
Trace reads the attacked copy against the full roster. As in production, it tries a plain read first, then lines the copy up against the original and reads again if needed.
Score.
Every cell is one of three results: traced (the right buyer, gate passed), inconclusive (nobody named) or wrong (someone else named). Unmarked controls go through the same attacks and must name nobody.
The gate. Per-fan photo and video marks carry a checksummed ID that must read cleanly and belong to one of your fans; for video the best match must also beat the next by the margin above. Mass-message and stream matches must be less than a one-in-a-million coincidence across everyone who received the content. Voice-note readings must agree across the clip.
Production marking and matching code, run on CPU on a local test stack. A neural mark for video, photos and voice notes, plus Trace's keyed second layer.
How Trace evidence is produced
Collection, examination, analysis, reporting.
Every evidence pack follows the four steps forensic practice uses (the NIST SP 800-86 framing): collection, examination, analysis, reporting. Each step writes a hash-chained row to the pack's custody log.
How a leak trace becomes an evidence pack
Collection.
The leaked file is received (URL fetch or upload) and hashed (SHA-256) before anything else touches it. The hash, the source URL and the UTC time are the first custody row.
Examination.
The file is prepared for decoding (frames, audio, stills) and the decoder is run against the creator's own roster of delivered copies only. Engine and model versions are recorded.
Analysis.
The result passes three gates before Trace reports a match: checksum, signal threshold and roster membership. The pack states the roster size at match time and the roster-adjusted chance that a random decode would pass by coincidence. A result that does not pass is reported as inconclusive, not as a near miss.
Reporting.
The pack is written as a ZIP: a one-page summary, the full evidence_pack.json, a chain-of-custody sheet, platform letters, a redacted share/ copy for third parties and the full copy for the creator and counsel.
Timestamps. When a timestamp authority is configured, the leaked-file hash, the delivered-copy hash, the pack hash and the custody-chain head are each stamped with an RFC 3161 token at export, and the detection event is stamped at match time. The pack says whether the authority is eIDAS-qualified. When no stamp could be obtained the pack says so, and the dates rest on Trace's clock alone. Anyone can re-check the hashes, chain and tokens with the public verifier.
Validation status. The bench on this page is an engineering smoke test on synthetic media with a small number of wrong-account trials. A pre-registered validation plan (real consented media, realistic attacks, roster sizes up to 10,000, thousands of wrong-account trials per modality) is written and not yet run; results will be published here with the raw data whatever they show, and we are looking for an independent lab to run it. Until then the pack's confidence statements lean on the per-case chance-match statistic, not on the bench.
The pack is a technical record of what Trace observed and when. Trace makes no statement about how any court, platform or other body will weigh it.
Independent verification
We’re inviting labs to re-run this.
Raw data, how to re-run it, and the guarantee
Contact
We’re inviting independent labs to re-run this; contact info@creatorsenseai.com.
No lab has verified these results so far. We will share the test content, attack scripts and access needed to re-run the bench, and publish the results whatever they show.
Download the raw results
Every cell in the tables comes from these files, unedited:
video_per_fan.json · video_ab_stream.json · photos.json · voice.json
Other runs: 25 Oct 2026 (video_per_fan.json, video_ab_stream.json, voice.json).
This page reads the latest run listed in index.json, so new runs appear here as soon as their files are published.
Right account, or your money back
If a leak trace ever shows the wrong account, that month’s subscription is refunded. An inconclusive result, where Trace shows no account, is Trace declining to guess, not a wrong answer. Full terms.
FAQ
Common questions, short answers.
Has an independent lab verified these results?
No. These are our own results, run on our own test stack. We are inviting independent labs to re-run the bench and will publish what they find. To take part, email info@creatorsenseai.com.
Did Trace ever name the wrong buyer in this run?
No. In the 22 Oct 2026 run, no attack on a marked copy produced a wrong name, and no unmarked control named anyone. When the evidence was too weak, Trace named nobody, which the tables show as inconclusive.
What can remove a Trace mark?
Anything that rebuilds the picture or sound instead of copying it. AI regeneration (for example running a photo through image-to-image), several fans combining their copies, very heavy crops, and a phone filming a tilted screen can all leave too little of the mark to read. Trace then says inconclusive rather than guessing.
How big is the sample?
Small, and we say so. One 64-second test video, 4 test images and one voice clip, each put through every attack. The results show how the marks behave under each attack. They are not a success rate you should expect for every file in every situation.
Does a no-match mean the fan did not leak it?
No. A no-match is inconclusive. The copy may be too damaged to read, may come from content you sent before marking, or may have been rebuilt by an AI tool.
New to per-fan marks? Read per-fan watermarking explained.
Next step
Mark your next drop.
Start with a 14-day trial (card required). Trace works alongside any takedown service: see how they fit together.
14-day trial, card required. See pricing.