Authenticated
evaluation for AI.

The numbers AI companies publish are produced by instruments that fail in specific, fixable ways. We find the failures, file the fixes, and attest the instruments that survive, so your number survives the follow-up question.

Output, to date
Papers5The Evaluator Trust Boundary · Side-Channel Exfiltration · The Verbosity Premium · Grokking Has Finite Capacity · all Zenodo, DOI · one under review at USENIX Security ’27
Filed200+Pull requests and issues against the evaluation stack, across 100+ repositories and 45+ organizations
Landed30+Merged upstream, or fixed by maintainers after we filed · SWE-bench, UK AISI, Microsoft, NVIDIA
Datasets3The defect corpus · a sealed RL holdout, non-gameable · the ATT&CK/Alignment Rosetta Stone
01The work

One thing,
done at depth.

We audit the instruments the AI industry measures itself with: evaluation harnesses, benchmarks, judges, leaderboards, scoring pipelines. Not a red-team menu. Not a platform. A fixed-scope measurement-integrity audit that ends in an attestation with a scope, a standard version, and an expiry date.

The artifact is an attestation your buyers can file in procurement: the answer to the question their security teams are learning to ask.

01

Pin verification

Every score pinned to the exact dataset revision, judge model, and grading code that produced it. If the bytes moved, the number is not the number.

02

Grader and judge pipeline integrity

The scoring path read line by line: extraction, parsing, aggregation, caching. Every defect proven with a running reproduction, not asserted.

03

Oracle and dataset residency

Whether the evaluated system can reach the answer key, the judge prompt, or the ledger its score is written to. Residency is a property, not a hope.

04

Hardened re-grade of your published number

Your headline result re-run under an instrument that fails closed. If it survives, you have the only number in your category a third party stands behind.

grading.py · SWE-bench · the shape of #620
12def grade(output: str) -> bool:
-   return "PASSED" in output
 
+def grade(output: str, exit_code: int) -> bool:
+   return exit_code == 0 and "PASSED" in output

The old check trusted a string the agent under test could print itself. The fix binds the grade to something it can't write: the process exit code. Merged, #620.

02The question

The question is coming, and it doesn't schedule itself.

Independent of you, who verified this number? 

It arrives from four directions: your champion's security team filling out the AI addendum, your VC's diligence refresh, the EU questionnaire with high-risk clocks consolidating December 2027, and the competitor whose number gets verified first.

Attested companies answer in one line. Everyone else writes a paragraph and hopes the paragraph lands.

03The record

We've read the instruments your numbers come from.

The harnesses, the judges, the graders, the leaderboards: read line by line, findings filed, fixes merged. Three quarters of the defects we find are one class: the evaluator trusts an artifact the evaluated system controls.

SWE-benchlm-evaluation-harnessinspect_aiHELMgarakPyRITdeepevalpromptfoolightevalOpenCompassopikterminal-bench

We publish what we find, named, with reproductions and fixes. Disclosure is private-first, always, on the vendor's clock. The instruments don't break toward alarm. They break toward passing.

04Capacity

Founding cohort, this quarter.

A deliberately small roster of companies whose published numbers we audit, attest, and stand behind on a quarterly cycle. One verified number per category per cycle: when a category's seat is taken, it is taken for the cycle.

05Open source

Free, MIT, forever.

The scanner and the safety stack are open source because the audit has to be checkable. Install them, read them, run them against us.

terminal
$ pip install etb-scan
the offline scanner, no API keys
$ npm install @authensor/aegis
content-safety scanner
$ npm install @authensor/sentinel
behavioral monitor

Free claim check. 48 hours.

No charge, no strings. It is how most engagements start.

What happens
01

You send one public claim. A benchmark score, a safety rate, a leaderboard position.

02

We read the instrument, not the number. Harness, judge, grading code.

03

One page back: what holds, what is unverifiable, what is broken. Each with a reproduction.