A durable provenance and consent layer for AI-era media. It records what is real, what changed, from what owned source, and under whose authority, in a form that survives the pipeline and that anyone can verify.
Trained on everyone's work. It produces plausible pixels from a prompt. It cannot tell you where a face came from, whether the performer agreed, or whether any of it was real. Impossible to prove, and impossible to defend.
Trained only on your production, its dailies, its lenses, its light. It repairs a shot from footage that actually exists. Every synthesized region names its sources, its approvals, and its consent. Provable frame by frame, and built to be defended.
Detect the error, propose from real footage first, gate on consent, collect human approvals, render locked to the approved bytes, then seal with durable provenance. If no honest repair exists, the answer is a clean CANNOT. There is no generate anyway.
Rights change over time. Something authorized in 2019 can be narrowed in 2022 and withdrawn in 2024. Continuum keeps the two questions apart and answers each on its own: was this authorized at the moment it was used, and is it permitted right now. Revocation changes the future, never the past. Scope can only narrow, never widen.
A use that was valid in 2021 stays a matter of record forever. Withdrawing consent later does not rewrite the past, and no one can quietly erase a use that happened.
The current answer reflects every narrowing and revocation to date. A withdrawn grant blocks new use immediately, without touching the history behind it.
Every event (grant, use, narrow, revoke) is written to an append-only hash chain. Each one carries both halves of a time interval, so it cannot be forged, reordered, or backdated by whoever holds the record.
The check is not a service we host. It is a standalone program that ships with the record and runs on your machine, offline, with no access to us and no private keys. It recomputes every hash, timestamp, beacon and proof, and reaches the same verdict we would. If anything was edited, deleted, reordered or backdated, it fails loud and gives no verdict. A tampered record gets no confident answer, by design.
$ python3 verify_ledger.py sample_genuine.json --authority AUTH-0007 --use FILM [PASS] hash_chain [PASS] timestamp [PASS] beacon [PASS] merkle_root [PASS] merkle_inclusion [PASS] archive_timestamp INTEGRITY: VERIFIED every hash, timestamp, beacon and proof recomputed verdict: was authorized in 2021 = True permitted now = False (revoked) exit 0 $ python3 verify_ledger.py sample_tampered.json [FAIL] entry_hash -> seq 0: body does not match its hash INTEGRITY: FAILED this ledger was altered VERDICT: WITHHELD a tampered record gets no confident answer exit 1
The offline verifier is open source. Read it, run it, and try to make a tampered record pass: github.com/Dr-PampersMuffin/continuum-verifier.
Proving an element was built honestly, from owned material, with consent, usually means tagging the file. That works until the footage hits a real pipeline. Transcode it, recompress it, change the format, and the tag is stripped. The proof is gone. Continuum takes the opposite approach: it keys the proof to the content itself, so it survives the pipeline and works across any tool a facility already uses.
Embedded credentials live inside the file. Transcoding, recompression, and format changes strip them. Once the tag is gone, you cannot prove a negative.
The record is keyed to the content itself. Re-hash the bytes and the full chain comes back: owned sources, approvals, consent, and whether AI was involved. It survives the pipeline and spans every tool, not just ours.
Take a sealed frame. Recompress it, resize it, rename it, strip every trace of metadata. Then hand it to anyone, offline, with no access to our system. The provenance link still comes back, recovered from the pixels themselves. When there is no reliable match, it says so rather than guessing. Embedded credentials die the moment a file is re-encoded. This does not.
We played the attacker against our own ledger: edit a sealed grant, delete a use, reorder events, strip a timestamp, backdate a revocation and re-seal the entire chain, and insert a phantom narrowing to rewrite history. An independent verifier caught every attempt.
The two-clock logic, the append-only authority ledger, its RFC 3161 timestamps, the Merkle batching, and the standalone verifier are operational now and independently checkable. In this build the timestamp authority and the external beacon are local development stand-ins; the protocol and the verification are identical to production, which points the same code at a public qualified timestamp authority and an external public beacon by configuration. The deep networks for photorealistic reconstruction are the next build, and they plug into gates that already exist, inheriting isolation, consent, and provenance from day one. This proves time, order, and integrity, not the truth of the underlying content. We would rather show you exactly where we stand than sell a mirage. That honesty is the whole point of the product.
Most tools bolt on safety after the fact. Here it is the foundation, enforced by the system, not promised by a policy.
A model learns a single film and can only ever serve that film. Checked at render time, fails closed.
Real footage first; generation only as a bounded last resort; an honest CANNOT when nothing works.
Nothing touches a performer's identity without a two-person consent scope. Synthesis is structurally ungrantable.
Every action in a tamper-evident chain. Approvals bind to the exact bytes approved.
The machine proposes; people approve. Higher risk demands more signatures, from different people.
The first AI post system a studio's legal team and a performer's union can both sign off on.
Fix a shot from footage that exists instead of reshooting, and carry proof of how every frame was made.
Every operation that touches an identity is recorded with its consent basis, maintained continuously, not assembled under legal pressure.
A portable, independently checkable record of who authorized what, for how long, and whether it still holds at the moment of use.
For pilot partnerships, studio evaluations, investment, or advisory conversations. The offline verifier is open source. Deeper technical materials, including the authority-anchor note and the engine codebase, are available under NDA to serious parties.