VERSIONED EVIDENCE · COMPILED 16 NOVEMBER 2026

What survives
an edit?

Observed watermark recovery, with the version, input, transformation and denominator kept together. Current-release checks and historical experiments have separate tables.

CURRENT RELEASE

Current release · 0.7.3a4

CJ-05, 15 November 2026. Linux m6a.large, 2 CPU / 4 GiB, two threads, verification scope current. Two image fixtures and seven video fixtures; repeated checks are not additional source scenes. No recompression, crop or social-platform survival rate is measured by this cohort.

Blind ID recovery after intact export or metadata-only change
Scenario / exact transformImageEconomy5sEconomy2sResilient
Image · Unchanged exportA separate blind verifier reads the accepted SDK output. Two procedural inputs: 1024×683 PNG and 2048×1152 JPEG; one first job + three resident repeats per input.8 / 8 recoveredExpected IDNot applicableNot applicableNot applicable
Image · Metadata removedLossless PNG re-save; JPEG input becomes PNG. Decoded RGB pixels unchanged. Two procedural inputs: 1024×683 PNG and 2048×1152 JPEG; one first job + three resident repeats per input.8 / 8 recoveredExpected IDNot applicableNot applicableNot applicable
Video · Unchanged exportA separate blind verifier reads the accepted SDK output. Seven procedural SDR clips, 720p/1080p, 5/10/15 s, 24/30/30000:1001 fps, including HEVC input; four checks per fixture/profile.Not applicable28 / 28 recoveredExpected ID28 / 28 recoveredExpected ID28 / 28 recoveredExpected ID
Video · Metadata removedMP4 stream-copy remux; compressed video essence unchanged. No re-encoding. Seven procedural SDR clips, 720p/1080p, 5/10/15 s, 24/30/30000:1001 fps, including HEVC input; four checks per fixture/profile.Not applicable28 / 28 recoveredExpected ID28 / 28 recoveredExpected ID28 / 28 recoveredExpected ID

HISTORICAL · NOT RETESTED ON CURRENT RELEASE

Historical paired transforms · 0.7.1a1

Linux/m6a.large; code 35cf34d4b794f0ba65588ba6f8d0b49f9040a24c; 2 CPU / 6 GiB container. Two 720p24 stock scenes (crowd, face). Blind detector receives no expected ID, preset or schedule. Seeks use the 24 fps frame grid. These observations are not a re-test of 0.7.3a4.

Two stock scenes, identical scenarios across the three video profiles
Scenario / exact transformEconomy5sEconomy2sResilient
Fresh 10 s exportNo later transform; 1280×720, 24 fps.2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID
Fresh 3 s exportThree-second rendition prepared before marking.2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID
H.264 CRF23 · 10 sIdentity transcode of a marked ten-second export.2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID
H.264 CRF23 · 3 sIdentity transcode of the marked three-second export.2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID
Keep seconds 1–4Three-second excerpt from a marked ten-second export.0 / 2 recovered2 not_found2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID
Keep seconds 0.2–1.8Excerpt deliberately between Economy marking windows.0 / 2 recovered2 not_found0 / 2 recovered2 not_found2 / 2 recoveredExpected ID
Trim 0.2 s from the startLeading trim of a marked three-second export; not an intact short export.0 / 2 recovered2 not_found2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID

HISTORICAL · NOT RETESTED ON CURRENT RELEASE

Historical codec and geometry checks · 0.7.1a0

Linux/Oregon m6a.large, 2 CPU / 6 GiB. Code 4933202. Two stock scenes; NTSC rows use two rates (four cases/profile). Portrait rows use two cases/profile. Historical runtime and decoder; not a current-release durability score.

NTSC/codec checks and separately prepared portrait renditions
Scenario / exact transformEconomy5sEconomy2sResilient
Fresh NTSC exportSix seconds, 720p; 24000/1001 and 30000/1001 fps.4 / 4 recoveredExpected ID4 / 4 recoveredExpected ID4 / 4 recoveredExpected ID
H.264 CRF28Recompression of the marked NTSC exports.4 / 4 recoveredExpected ID4 / 4 recoveredExpected ID4 / 4 recoveredExpected ID
90%-dimension center cropRetain 90% of width and height, then resize. This is not a 16:9-to-9:16 crop.4 / 4 recoveredExpected ID4 / 4 recoveredExpected ID4 / 4 recoveredExpected ID
640×360 + H.264 CRF32Combined downscale and severe compression. The written record supplies the per-mode recovery split; the summary does not split every unrecovered result into not_found versus incomplete.0 / 4 recoveredUnrecovered status not split0 / 4 recoveredUnrecovered status not split3 / 4 recoveredUnrecovered status not split
Portrait rendition prepared before markingTwo source scenes; square-pixel portrait input prepared before export. Separate from cropping a previously marked landscape file.2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID
Portrait export → 540×960, CRF28Two source scenes; square-pixel portrait input prepared before export. Separate from cropping a previously marked landscape file.2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID2 / 2 recoveredExpected ID

HISTORICAL · NOT RETESTED ON CURRENT RELEASE

Historical deliberate-removal checks · 0.7.1a1

Same Linux two-scene cohort. Image PSNR 45.03–49.62 dB relative to marked images. A separate video result reads crowd only through aggregate recovery. This is an adversarial scenario, not an ordinary delivery transform or an attack-proof guarantee.

Paired controls and the documented upstream remove_watermark() method
Scenario / exact transformImageEconomy5sEconomy2sResilient
Image · identity controlRGB PNG with C2PA stripped; two source frames × two IDs.4 / 4 recoveredExpected IDNot applicableNot applicableNot applicable
Image · public TrustMark removerSame paired images; C2PA stripped in both control and treatment. Four trials from two scenes, not four independent images.0 / 4 recoveredUnrecovered status not splitNot applicableNot applicableNot applicable
Resilient video · identity controlTwo ten-second 720p24 scenes; matched codec/color path.Not applicableNot evaluated hereNot evaluated here2 / 2 recoveredExpected ID
Resilient video · public remover240 remover calls per scene. Crowd recovered by aggregation after full scan; face returned not_found. Economy video not tested in this paired remover trial.Not applicableNot evaluated hereNot evaluated here1 / 2 recovered1 not_found

CURRENT RELEASE · 0.7.3a4

A negative and an unfinished scan differ.

Same CJ-05 procedural inputs, first job plus three resident repeats. The separate 60-second control has one cold and one resident run. No supported ID in a completed current-scope scan is not proof of human creation or a population false-positive estimate.

Current-release clean controls, selected verification scope current
Inputs / scopeCompleted resultUnfinished checks
ImageTwo procedural images, four checks/input.8 / 8 completed negatives0 incomplete
VideoSeven procedural clips, four checks/input. The 1080p/15 s NTSC negative exhausts its budget.24 / 28 completed negatives4 incomplete
60 s video · separate budget control0 / 2 completed negatives2 incomplete

RESEARCH ONLY · HISTORICAL 0.7.1a0

Letterbox recovery is a separate research result.

Two landscape stock scenes × three profiles, placed inside 720×1280 letterbox frames with a 720×406 active picture. A bounded FFmpeg cropdetect probe restored the active region before a separate decode. This ROI path is not enabled in the qualified public verifier.

Historical 0.7.1a0 letterbox probe · not shipped in the public verifier
Decoder inputObserved ID recovery
Full letterboxed frame0 / 6 recovered
FFmpeg cropdetect ROI6 / 6 recovered

NOT ESTABLISHED BY THIS EVIDENCE

Keep the remaining scope visible.

Known crop/remover and window-free-excerpt limits remain relevant. Their historical counts are not silently relabelled as current-version measurements. No profile, including Resilient, is described as attack-proof.

Methods, versions and retained records

Current 0.7.3a4: CJ-05 methods and full measurements; two procedural images and seven silent SDR fixtures. Whole-file watermark checks and metadata recovery are separate from proof of asset authenticity. 341 installed-media tests are release qualification, not 341 durability attacks.

Historical 0.7.1a1: Complete paired-media protocol · Machine-readable observations. Two 1280×720/24 fps stock scenes, prepared 3 s renditions, development credentials and offline Linux testing.

Historical 0.7.1a0: Codec/geometry protocol · Machine-readable observations · Spatial research and limitations.

Historical source files are retained at SDK documentation commit f90a59917b850c3aee36bfc1ac6a506bed17b571. Source paths and SHA-256 manifest ↗. The tables are generated from these records and the current CSV; no new media experiment was run to produce this page.

Why the tables use this structure

Each cohort keeps one runtime/corpus definition, named transformations and raw counts, with machine-readable data beside the human summary. This follows the reproducible per-augmentation reporting approach of VideoSeal evaluation. Captions and explicit row/column headers follow W3C table guidance. Vendor scores are not imported into Ariadne results.

Choose your video profile ↗Full test protocols ↗