We grade models and we sell fixes
That is a conflict of interest. Here is exactly how we manage it, in clauses you can check, because an audit you cannot verify is worth nothing.
The conflict, disclosed
hell.ai is published by Black Sheep AI, which also sells deployment-mitigation tooling. Every clause below is designed to be checkable, not taken on faith.
Grades are computed, not decided
Every grade is produced by the published rubric from published sub-scores, before any commercial conversation. Recompute any grade yourself from the model card.
No pay-for-grade
Vendors cannot pay for inclusion, exclusion, a re-grade, or a delay. Model selection for the public index follows published criteria: download rank plus reader requests.
Products are never named in a verdict
Public model cards and reports recommend generic control classes (ingestion sanitization, indexed abstention policy). They never name a Black Sheep AI product.
Private audits are firewalled
A paid private audit of a customer stack uses the same public methodology, and its results belong to the customer. It never changes a public grade.
Dispute a finding
If you are a model vendor and you think a finding is wrong, challenge it. We re-run with your nominated serving configuration where it is reasonable, publish both the original and the re-run, and print your statement verbatim on the model card. Disputes and their outcomes, including the ones we lose, are listed in a public register. Open a dispute at disputes@hell.ai.
Dispute register · no disputes filed to date. This space will show them when they arrive, resolved for and against.
Who runs this
hell.ai is published by Black Sheep AI, the group behind the RAM compression and Watchman provenance work. Audits run on owned hardware, not a shared cloud tenancy. The named individual accountable for the methodology and for the conflict-of-interest firewall, and the registered legal entity, are stated on request and will be published here at public launch. For methodology questions, verification access, or to reach the accountable reviewer, write to audit@hell.ai.
About the name
The name is deliberate, and it describes the work. Models go through hell here, handed relevant, irrelevant, and hostile documents, so that your deployment does not. Past the front door the tone is sober on purpose: the cards and reports are written to sit in a risk file. For a formal citation, use “Black Sheep AI Deployment Risk Index”.
Using the mark
The hell.ai grade mark is a first-party mark. It means one thing: these measurements exist and you can check them. It is not a certification, an endorsement, or a statement of regulatory compliance, and it carries no language implying any of those.
Reproduce the mark unaltered, including the grade letter, the audit date, and the methodology version, which are inseparable.
Keep the verification link. It resolves to the live model card and tells the truth about the grade’s current status and age.
Do not use the mark to imply certification, endorsement, or compliance. Misuse voids the license and is noted publicly on the model card.
Changelog · methodology v1.0 published 2026-07. All grade changes, method revisions, and corrections are recorded here as dated diffs. Nothing on this site is silently edited.