For AI labs & enterprise fine-tuners · Defensive Exposure Audit
The same probe suite rights holders use against models, run for the other side: a neutral third party tests your model against specified corpora (a plaintiff's catalogue, a licensing counterparty's archive, your own pre-release checklist) and documents what it does and doesn't emit. Hash-audited, reproducible, and honest about limits: the report a litigator can put weight on precisely because it never overclaims.
Why a third party
An internal "we checked, it's fine" carries little weight, and a vendor who sells certainty carries less. BotWitness is the neutral recorder both sides can cite: the same archive that produces publisher-side evidence produces your crawl-compliance attestations, and the same probe methodology we publish for rights holders runs unchanged against your model. Neutrality is the product; we report what the probes show, both directions, with the limits stated.
The engagement
Probe suite
Reproduction (verbatim continuation), recognition (DE-COP multiple-choice), and (where your endpoint exposes token probabilities) calibrated membership statistics, run against the corpora that matter: a claimant's works, licensed archives, or your pre-release checklist. Any OpenAI-compatible endpoint works, including private deployments.
Paired artifact
From the same tamper-evident archive: what robots.txt permitted your crawler across the relevant domains and dates, the attestation that corroborates your collection story with a record you don't control.
Discipline
Every transcript SHA-256 hashed and retained; probes at temperature 0; methods documented in the report itself. If the matter goes to experts, they can re-run it.
Honesty
A null result is reported as a null result, not a clean bill of health. Signal is reported as signal, not a probability. That is exactly what makes the document usable in front of a regulator or a court.
When labs call us
Probe the release candidate against high-risk corpora before launch, catch the memorized passage before a journalist does.
A licensing counterparty claims your model recites their catalogue. Get the neutral measurement before you price the deal.
Respond to a claim with a documented, reproducible probe record and a crawl-compliance attestation, instead of assertions.
Engagements
Defensive audits are operator-run engagements, from $5,000 per model/corpus scope, delivered as a signed report package. Tell us the model surface, the corpora, and the deadline. Our concierge scopes it in minutes, and a human approves every quote.
