Architecture decisions, benchmark results, and compliance engineering from the ComplyEdge team.
The AI Act says why to log and how long to keep it, not which fields to write. What ISO/IEC 24970 and prEN 18229-1 are expected to settle, what AESIA Guía 12 already lists, and the one field none of them asks for: the rule and the law behind each decision.
Read articleNo logging rule asks for the rule and the legal provision behind each decision. Explanation duties under AI Act Article 86, GDPR Articles 15 and 22, the Court of Justice in C-203/22 and the Consumer Credit Directive are hard to meet without it.
Read articleA record of each AI decision that holds no prompt and no answer, and one call that tells a reviewer whether your own copy is the exact text a decision was about. The steps, the calls and the limits.
Read articleSince 2 August 2026, anyone can start a case under the EU AI Act: a complaint, a whistleblower report or a downstream provider's claim, alongside the authorities' own checks. The question is what you produce when one is opened. The three channels, correctly scoped, and what a hollow log looks like next to a complete, exportable decision record.
Read articleFrom 2 August 2026, Article 50 binds. When an AI agent interacts with a person, the system must disclose what it is and on whose behalf it acts. Here is the obligation, the Commission's guidance on agents, and what the runtime enforcement audit trail must carry: hash, timestamp and chain position on every decision, the legal citation when a phrase rule or the injection check fires. A wider-review block does not cite an article. and an export a reviewer can verify offline. That kind of record is now specified as ADCR, the open Rulebook.
Read articleThree checks against the EU AI Act rules that anyone can run from the public repository, with no account: a rule ID and the article it applies, benchmark scores that re-derive offline, and an audit export that fails verification when one field is edited.
Read articleGoogle and Meta lead at 83.3%; xAI trails at 16.7%. Industry average 53.1% on public documentation across six GPAI obligation categories: open leaderboard, reproducible from provider evidence files. Every score is a reading of public documents, still marked needs_review.
Read articleMost AI guardrails use an LLM-as-judge pattern. That works as a content filter: it does not work as compliance. We put OPA/Rego on the hot path. After a clean pass on a paid plan, an injection meaning check runs when the month's model budget covers it. If that budget is used up, the phrase rules decide. The wider legal review runs only when you turn it on and that budget covers it. Here is why the order matters.
Read article