AI observatory.

Explore measured reliability, token usage and decisions behind our advisory AI, with evidence you can verify.

Loading verified records…

Public repositories are discovered each collection, including new ones. Confidential repository identities and payloads remain private. Shared CI measurements and controller-reported agent events appear separately below. Historical uninstrumented tasks are absent.

Verified records

—Awaiting data

Pipeline valid output

—Excludes credential skips

Tokens consumed

—Measured usage only

Mean analysis time

—Measured attempts only

EXECUTION

Across the fleet

Signed records only

Measurements by repository and mode
Repository / mode	Runs / skips	Valid rate	Attempts / run	Input tokens	Output tokens	Reasoning	Findings proposed / published	Patches accepted / rejected

RELIABILITY

Failure categories

Attempt failures and validation types are separate counts.

EVIDENCE, WITH LIMITS

What these numbers mean

A valid response passed the output contract. It does not prove the findings are correct. Published findings have a GitHub publisher receipt.

Token totals include cached input and reasoning output. Missing usage stays unknown. Historical runs before collection are absent.

Review quality outcomes
Not yet collected. No estimate is shown.

Read the schema and verification guide ↗

ADVERSARIAL TESTING

Latest signed red-team run

Awaiting verified red-team evidence.

These are measured attacks against the shared runner. A passing suite does not establish general security or complete OWASP coverage.

AGENT WORK

Controller attempt evidence

Awaiting verified agent evidence.

Signatures verify hosted ingestion of locally reported records. They do not independently prove local execution or correct decisions. Payload retention includes hashes and bounded excerpts, rather than full prompts and outputs.

Controller-reported events by repository, stage and provider
Repository / stage	Provider / models	Events	Reported outcomes	Input / output tokens	Usage coverage	Mean measured time

Download signed agent dataset ↗