Catch what long technical files hide.

EU AI Act pre-screening for Articles 9 to 15, one Article at a time, with NVIDIA Nemotron on Nebius.

Real output for the example technical file, a gas-safety monitor with a hidden Article 14 gap.
45%4.1%

Violations falsely cleared when the violating section comes first in a five-section file. Whole-file Nemotron 3 Nano vs the Screener, 73 cases, median of three runs.

6.8%

Cleared by one whole-file Nemotron 3 Ultra call: a statistical tie with the Screener, which wrongly rejects more compliant files (15% vs 6%).

1,560

Scored runs in a pre-registered re-test on 120 expert-labelled AIReg-Bench files: held-out split, three repeats, bootstrap CIs.

Built from what broke

We tested LLM auditors before building one, then re-tested our own design under a protocol fixed in advance. Each step answers a failure we measured.

Judge one Article at a time

With a generic prompt, compliant sections after a violation dilute it. The Screener asks about one named Article and counts a disclosed gap against compliance. Our re-test showed this prompt, not splitting the file, closes most of the gap.

Sample three times

Nemotron 3 Nano judges each Article three times in parallel. A single sample flips between identical runs and over-rejects; the median of three cut false rejections from 23% to 13%.

Escalate disagreement

When the samples disagree, Nemotron 3 Ultra decides. If Nano and Ultra still disagree, the Article goes to a human instead of a guess.

Flag claims, check quotes

Notes to the reviewer and unsupported compliance claims are flagged and ignored. Evidence quotes are shown only if they appear in your text.

Screen a file

Paste documentation, ideally with an "Article N" heading per section so evidence is easier to check, or load an example. Takes about 10 to 20 seconds. A screening aid for reviewers, not legal advice.

Optional. What the AI system does and who uses it.
Up to 60,000 characters.
Results appear here, one verdict per Article.

What the research found

3 of 3

Long files fool a generic judge

With the violating section first in a five-section file, whole-file judges using AIReg-Bench's prompt falsely cleared 9.6% to 52% of non-compliant documents. A targeted prompt brought that to 0 to 4%, with or without splitting the file.

0

Fake approvals did not work

"Certified compliant" and "pre-approved by the AI Office" text cleared no non-compliant document.

1/32

Rewrite attacks rarely did

Claude Sonnet rewrote violating files to look compliant. It succeeded once.

0.657

Experts disagree too

The three legal annotators agree with each other at Krippendorff's alpha 0.657, so the Screener sends close calls to a person rather than pretending to certainty.

Why not just one Ultra call?

About 4x cheaper: $0.009 per five-Article file vs $0.038 at Token Factory list prices. Most calls go to Nano, and Ultra reads one section, not the whole file.
It can say "I'm not sure". When Nano and Ultra disagree, the Article goes to a person instead of getting a guessed verdict.
Every quote shown is checked against your text, so a reviewer can verify the verdict in seconds.
Notes to the reviewer and claimed approvals are flagged, not just ignored, so a manipulation attempt is visible.

What it is not

Not legal advice or a conformity assessment. It helps a reviewer decide where to look.
Evaluated on synthetic, expert-labelled excerpts. Real technical files are longer and messier.
Covers Articles 9, 10, 12, 14 and 15 only, the core high-risk requirements.
Not more accurate than one strong judge. A single Ultra call misses about as many violations; the Screener uses Nano and sends about one Article in eight to a person.