Inspect the source
Scan a bundled sample or a public GitHub repository. Annex identifies source patterns and maps relevant controls to the selected markets.
Source, not self-reportBUILT FOR ACCOUNTABLE AI
Show the code.
Back the claim.
Your AI policy says people can override a decision. Does the code agree?
Annex connects AI governance obligations to the files and lines that support them. Find gaps, plan the fix, and give reviewers evidence they can check.
No signup. No API key. Your own demo workspace.
52executable controls
5rule packs
82source detectors
ANNEX / EVIDENCE REVIEWBundled sample
HireFlowProhibited practice detected
A hiring signal
that crosses a line.
Emotion inference in a recruitment workflow
01 / THE OBLIGATIONEU AI Act · Article 5(1)(f)
Emotion inference at work or in education is prohibited, subject to the medical and safety exceptions.
02 / THE EVIDENCE
src/interview/signal.ts
17 export async function
detectEmotion(frames, transcript)
{ /* source cited in the scan */ }
Rule-based finding File · Line · Hash
FROM QUESTION TO EVIDENCE
Less chasing documents.
More understanding the system.
Made for builders, founders and assessors who need to work from the same facts.
Review the finding
Read the obligation beside the cited code. Separate missing evidence from partial implementation. Ask for a plain language explanation.
Every answer has contextAct and verify
Download a technical dossier or remediation patch. Compare scans, inspect the signed ledger, and publish a summary without exposing code.
A reviewable path forwardA DEMO WITH A BEFORE AND AFTER
The fix should change
more than the paperwork.
Compare HireFlow with HireFlow v3. Follow the prohibited practice back to its source, then inspect the revised system, remaining gaps and scan history.
Follow the evidence →HIREFlowProhibited
Emotion inference in hiring
↓ Implementation changes
HIREFlow v3High risk
Prohibited path removed. Obligations remain.
A better scan is evidence of change, not a certificate of compliance.
TRUST STARTS WITH CLEAR LIMITS
Checkable findings.
Human judgment.
Static analysis finds evidence in code. It cannot prove that a control works in production, determine every legal exception, or replace a qualified assessment.
AI is optional and only explains an existing finding. It cannot edit the scanner’s verdict or the evidence ledger.
Classification and scoring run without a model. Repeat the same scan inputs and rules to reproduce the evidence.
Hash chains and signatures make edits detectable. A signature identifies a key; it does not prove legal compliance.
The 50-case in-house benchmark tests regressions. It is not a real-world accuracy guarantee.
The public demo uses temporary, isolated workspaces. Use self-hosting for private code and durable storage.
LEXHACK 2026 · AI SAFETY & GOVERNANCE