Fortress binds crop-hardened provenance and per-session watermarks into content, attributes it back to its origin, and guards against false attribution. Sovereign, on-device, no cloud round-trip. Available for licensing or acquisition.
One engine family, two deliberately opposite behaviours. The SDK hard-wires the profile per facade, so a use case cannot pick the wrong one.
The mark is meant to still be there after the content has travelled: compression, re-encode, resize, re-share. Each copy can carry a per-recipient mark, so a leak decodes to its source.
The opposite goal: a fragility map that shatters on edits, so a changed region becomes visible and an unaltered capture is provable. Verified by raw sign flips, never by error correction.
All figures: Research & evidence · reproducible harness under NDA · the fragile verdict is “captured live & unaltered”, never “not AI-generated”.
A tiled, crop-hardened embed carries the proof in every region, so attribution survives the transforms that destroy a naive single-region watermark: JPEG q40–q85 and 50–75% downscale at 100%, noise 98%, recolor 90%, crops 10–50% at 86–92% (source: Fortress benchmark, re-run 2026-09-02, 44 real photos). Origin is bound at the moment of capture, on-device; an adaptive profile self-verifies at embed time and stays fully invisible where the content allows.
crops 10–50%: 86–92% recovery (source: Fortress benchmark, re-run 2026-09-02, 44 real photos) — blind centered-crop recovery is a disclosed open gap
per-viewer mark for leak tracing; video attribution uses an evidence floor (≥3 frames, ≥60% share)
null-control measured: 0/49 false attributions on unmarked assets, 0 hits in 200 unmarked video frames
Fortress is evaluated the way our LLM-trust work is — demonstrated, not asserted. Robustness is evaluated per attack (compression, crop, resize, recapture), imperceptibility measured (PSNR), and — critically — false / phantom attribution is measured against a null control, not assumed away.
The numbers above come from a real, founder-supplied 44-image corpus with real codecs. What does not work, in the same breath: blind centered-crop recovery in the shipped SDK is an open gap — an earlier "crop solved" figure came from a harness that was given the crop factor and key, and we retracted it; a full generative AI redraw destroys any blind watermark (information theory — that boundary is why the fragile tamper layer exists, which flagged AI edits and screenshots in our own Gemini img2img tests); collusion (averaging two copies) defeats per-session tracing — no traitor-tracing codes yet; audio marks survive MP3/AAC/Opus but are audible-ish (SNR 14–17 dB). The reproducible harness ships under NDA. We do not claim "military-grade" or "unbreakable."
Claims photos, lab images and agent decisions that must be provably unaltered — and whose alteration must be visible, with locality.
Screeners, creator media, high-security and confidential files: every copy carries a per-recipient mark that survives re-share and decodes to the source.
Proof-of-real is bound inside our own camera at the shutter — never reconstructed later, never granted to an upload.
AI can now fabricate micrographs, histopathology and electron-microscope images that reviewers cannot tell from authentic captures. Today provenance is asserted administratively, after capture. There is no cryptographic proof that an image came from a specific instrument, unmodified, at a specific time.
We want to test whether a keyed, fragile integrity map can be bound as early as possible in an instrument’s image path. Today it runs in software on the finished image. A later verifier shows where an image was changed after sealing, for edits of roughly 64 px and larger, including local AI inpainting. Verification can run zero-retention.
The watermark core that proves the concept is working software: in a lab test on 3 images the fragility map raised 0 false alarms and flagged re-save, screenshot, video replay and local inpainting; it misses edits of about 32 px and smaller. It is software only. Anything inside an instrument or sensor is research, with no partner and no hardware yet — that is the co-development scope, not a finished product. We do not claim a mark survives a full generative redraw; for scientific integrity that is the point: fragile by design, so any manipulation is provable. Core methods patent pending (provisionals 03–04/2026). Not a KYC or consumer-identity pitch.
Every number in the films is measured; each film has a scene that says what we do not claim.
All films, both styles, both languages: /films/
For camera & capture vendors, DAM and media platforms, creator platforms, and content-authenticity providers who need robust, sovereign provenance with attribution integrity.
| Model | What it is | Fit |
|---|---|---|
| License + royalty | Embed Fortress provenance/attribution as an SDK; per-device or per-asset royalty. | Capture / media / creator / DAM platforms |
| Exclusive field license | Upfront + royalty, exclusive in your field of use. | A strategic content-authenticity partner |
| Acquisition | One-time asset sale — the provenance IP, SDK, and evaluation harness. | Strategic acquirer / corp-dev |
Data room under NDA: technical method, patent status, reproducible robustness & attribution benchmarks. SDK evaluation access on request (IP@destill.ai) — Node SDK with 12 per-use-case facades, browser bundle, and the test suite that produced the numbers on this page.
Licensing, royalty, or acquisition — start with a 20-minute conversation and the evidence pack.
IP@destill.ai