Write a kernel once. Run it on CPU or GPU. Ship a cryptographic receipt anyone, on any machine, can re-execute to confirm the result bit-for-bit. Don't trust it. Re-run it.
GPU floating-point is non-deterministic by default - the same code gives different bits on different hardware. Obsign runs its kernels in a fixed-point integer mode where every substrate computes byte-for-byte the same answer. Here is that mode, running the framework's own coherence-field PDE, right now, on your machine.
Every Obsign output ships a signed receipt. Paste one into the verifier and your browser re-hashes the claim and checks the Ed25519 signature locally - the same way a Content Credentials viewer would. Nothing is uploaded, ever.
The world is racing toward generative output that hallucinates, can't be reproduced, and can't be undone. Obsign inverts every one of those: a deterministic integer operation whose output is bit-reproducible, whose every changed pixel is receipted, and whose original is always recoverable.
Obsign preserves the original, hashes it, and receipts the exact pixel diff - precisely what the operation added. A structure-tensor coherence map measures trust; large changes in low-structure regions light up the invention map as suspicious.
The enhancement is provably reversible. You can always get the source image back, exactly, and prove that the receipt belongs to it. Generative quality becomes transparent, reversible, and usable in trust-critical work.
From a bit-exact integer model to a frontier diffusion network, Obsign wraps any architecture in the strongest guarantee it can honestly support - and degrades gracefully, never to zero.
Quantized integer inference. Identical to the last bit on CPU, GPU, and native C++. Re-run it anywhere and it reproduces. This is the flagship.
Any neural network, run with fixed seeds and deterministic ops, wrapped in an accountable receipt. Reproduce it on the same setup; the diff is still receipted and reversible.
Even a black-box generative model gets the preserved original, the signed diff, and the coherence and invention maps. You always know exactly what changed, and can undo it.
A receipt gives certainty. But most files in the world arrive with no provenance at all. For those, Obsign runs a forensic panel that localizes traces and reports calibrated evidence - and, uniquely, the analysis itself is deterministic, signed, and re-runnable: another examiner re-runs the exact pipeline and gets a bit-identical report. The way forensic science is supposed to work.
Copy-move cloning (localized at source and destination) and editor-signature byte facts (Photoshop, Topaz, GIMP, Office) - exact and reliable, on images and PDFs.
Noise, CFA/demosaic and resampling detectors flag traces with explicit confidence and limits - never a verdict. Absence of a finding never establishes authenticity.
Most forensic tools are black boxes. Obsign's report reproduces to the bit and is signed - the reproducibility Daubert and NIST actually ask for.
Nobody productizes provably-deterministic, auditable, programmable image and compute transforms. Capture-provenance is commoditizing - cameras and labs already sign at the source. Edit-provenance, the part regulation is forcing and no one has solved, is wide open.
Write a per-pixel kernel in Coherence Language. Obsign's compiler turns it into a GPU shader inside your browser and runs it live on your webcam - with a receipt that proves the CPU and GPU paths produced the same pixels. The same engine that powers the SDK, running on glass.
getUserMedia → GPU shader → canvas, compiled from your .cl source in real time. Nothing leaves the page.
The same kernel lowers to WGSL on the GPU and to native integer code on the CPU - and they agree bit-for-bit.
A live CPU=GPU readback comparison proves the transform is reproducible before you ever trust the output.
Deterministic, auditable enhancement with a chain-of-custody report shaped for Daubert / FRE 702. The examiner re-runs it; the result is identical.
A real Content Credentials edit step that adds something the standard can't: the operation re-runs and reverses. Unbroken provenance chain.
An OFX plugin that ships the model into Resolve and Natron, with tile-invariant output - no seams, bit-identical regardless of how the host tiles the frame.
Wrap any model - even your best closed one - in a provable, reversible, signed layer. Sell trust over commodity quality.
If the trust pitch isn't honest, the whole thing collapses. So here is the calibrated truth.
We are looking for design partners in forensics, journalism, and post-production - the places where reproducible, reversible, signed compute changes what is possible. The bottleneck is validation, not code. If that's you, let's talk.