Who made this — and has it changed since?
Assayer is an independent assay office for AI-era content. Creators and models sign at the source; anyone downstream gets one calibrated origin-and-integrity verdict — fused from every signal that survives real editing, and honest enough to answer “Inconclusive” before it ever falsely accuses.
Watermarks wash out. We built for that.
An AI-text watermark is a statistical bias in token choice, not hidden data. It survives a copy-paste — and little else. So Assayer never leans on a single mark; it weighs every signal by how much of it could have survived the edits it actually finds.
Sign at the source. Fuse what survives. Abstain before accusing.
Sign at the source
At creation, a maker or model mints a portable content credential that binds who, when, and which tool to the artifact’s content hash. Signing is client-side — the key never leaves the creator.
Gather what survived
Downstream, Assayer collects whatever remains — the credential, any surviving watermark, a semantic fingerprint, metadata — and recomputes the hash to see what changed since signing.
Grade, then fuse
Each signal is weighted by how reliable it is and how much could have survived the edits detected, then fused into one calibrated band — Verified, Likely, or Inconclusive — never a lone number.
Abstain before accusing
Below the confidence floor, or when signals conflict, the verdict is a calm grey “Inconclusive” — logged, reproducible, and inside an intended false-accusation budget. We go grey, not red.
A hallmark, stamped at creation.
The strongest signal doesn’t depend on a fragile mark surviving later edits — it’s a cryptographic credential minted when the content is made, bound to its hash and registered to a maker’s mark. One SDK call, client-side, and the chain of custody begins.
assayer.sign(artifact, key)
A confidence that means what it says.
A stated “90%” should be right about nine times in ten — so Assayer is designed to measure and publish its own calibration, and to abstain below a floor rather than invent certainty. The costliest error in this field is a false accusation, so the UI has no “guilty” state — only confidence, and honest grey.
When evidence is thin or conflicting, Assayer goes grey, not red. “We can’t be sure” is a first-class, logged answer — never an accusation.
Settle disputes with history, not a snap judgement.
Every verdict writes to a reproducible chain of custody — each hop annotated with which signals survived it, and exactly where integrity last provably held. Anyone can re-run the check.
The network is the moat.
A portable, open content-credential standard plus an issuer registry of makers’ marks — worth more with every creator that signs and every surface that verifies. Institutions wire one endpoint into an LMS, CMS, newsroom or intake system and verify at scale.
The mark is a commodity. The calibrated verdict and the Register are not.
Any single watermark can be stripped or paraphrased away, so Assayer doesn’t sell one. It owns two things that compound instead: a survival-weighted fusion engine that turns whatever evidence is left into one honest, calibrated verdict — modelling the asymmetric-key reality, where a key-holder’s credential is near-certain and a keyless guess is heavily discounted — and the Register, a portable credential network that grows more valuable with every maker who signs. The defensible part is the verdict you can trust and the network others plug into, not the model and not the mark.
The honest answers.
Isn’t a watermark enough?
No — and pretending otherwise is the mistake. An AI-text watermark is a statistical bias in token choice, and a single paraphrase through an unkeyed model wipes it completely; file and metadata marks strip even more easily. Assayer treats marks as disposable corroborators and leans on the cryptographic credential and the Register, which survive copy-paste, re-export and translation. The moat is the calibrated verdict plus the network, not the mark.
Could Assayer falsely accuse someone of using AI?
The whole design resists it. When evidence is thin or conflicting, Assayer returns a calm grey “Inconclusive” instead of an alarm-red accusation, and it is built around an intended, independently auditable false-accusation budget. We would rather say “we can’t be sure” than risk ruining a student or author — abstaining is the feature, not a failure. The UI has no “guilty” state, only confidence and honest abstention.
Do you claim to detect every AI-generated text, or which model made it?
No. We make no claim of an unbreakable watermark or perfect detection. A specific tool is named only when a credential declares it; without a credential or key, and after a clean unkeyed paraphrase, there may be nothing left to attribute — so Assayer reports a calibrated AI-range-vs-human-range likelihood, or abstains, and never pins a specific model on unsigned text. Honesty about the ceiling is what makes the verdicts you do get trustworthy.
Is this a walled garden, or does it work with open standards?
It builds on open, C2PA-style content credentials rather than a proprietary silo, and the Register is portable so issued credentials resolve anywhere. What Assayer adds on top is the survival-weighted fusion engine and the issuer network — the calibrated verdict and the compounding coverage — not another standard to be locked into.
Does verifying content expose it, or its author?
Signing is client-side and binds to a content hash, not the content itself, so the private artifact and the secret key never leave the creator. Verification is designed to be privacy-respecting — resolving a credential and grading signals without hoarding the underlying content — because a provenance layer that leaked what it checked would defeat its own purpose.
Assay the origin. Abstain before you accuse.
Assayer is an early concept from The Adjacent Possible. We’re building the survival-weighted fusion engine and the credential Register, and we’d like to talk to newsrooms, universities, courts, and enterprises who need provenance they can defend — and who refuse to falsely accuse.
Or write to [email protected]