The EU AI Act Article 12
logging layer for high-risk AI.
Article 12 requires every high-risk AI system to log its events automatically and keep them for at least six months. Praxa records each decision your AI makes (input, score, rationale, downstream action) into a per-agent SHA-256 hash chain, so the record is tamper-evident and auditor-verifiable. The high-risk obligations apply from December 2, 2027, but the log is architectural: you can only capture a decision as it happens, not reconstruct it later. And the deployer stays liable even for AI systems they bought.
Wrap your model call
A 3-line change around the model call your high-risk AI already makes. Works with the Vercel AI SDK, OpenAI, Anthropic, or LangChain — whatever powers the decision.
Every decision is logged
Each decision appends to a per-agent SHA-256 chain — input, score, rationale, downstream action. Forward-only, so an altered or deleted record breaks continuity and shows.
Hand the auditor evidence
Generate the EU AI Act Annex IV evidence pack on demand. Auditor-ready markdown or JSON — chain continuity is verifiable without a Praxa account.
Article 12 is the requirement you cannot backfill.
Most AI regulation asks for a policy or an annual report. Article 12 of the EU AI Act asks for something a slide deck cannot fake: an automatic, per-event log of what your high-risk AI actually did, retained for at least six months. High-risk obligations apply from December 2, 2027 (the date moved under the 2026 Digital Omnibus), but a log only exists if it was written at the moment of the decision. Start it the day before the audit and you have nothing to show. A deployer stays on the hook even when the AI system was bought from a vendor. Praxa gives you that record from day one.
Automatic event logging
Article 12 requires high-risk AI systems to record events over their lifetime, automatically. Praxa wraps the model call so every decision is logged the moment it happens — no manual journaling, no gaps a regulator can point at.
Six-month minimum retention
The logs must be kept for at least six months, longer where other law applies. Praxa’s retention tiers cover the Article 12 window and beyond, with the dated, ordered chain intact for the whole period.
Any high-risk category
Hiring, lending, insurance, biometric identification, education, access to essential services — any Annex III high-risk system carries the Article 12 duty. Praxa is the logging layer regardless of which category you deploy in.
Tamper-evident audit chain
Per-agent SHA-256 hash chain. Any altered, inserted, or deleted decision breaks continuity from that point — verifiable by your auditor via the public REST API or a standalone 25-line Node verifier. Praxa never has to be in the room.
Hiring is the Annex III category being tested first.
AI used to screen and rank candidates is a named Annex III high-risk use, so it carries the full Article 12 logging duty. It is also the category already in litigation: in Mobley v. Workday (class certified in 2026), an AI hiring vendor was held to answer as an “agent” of the employer. A per-decision log stopped being a nice-to-have. It is what plaintiff’s counsel asks for in discovery.
If you deploy AI for hiring, Praxa is the per-decision tamper-evident log underneath your bias audit. Bias-audit vendors such as Holistic AI and Warden AI produce the point-in-time report on whether a tool discriminates. When a regulator or counsel asks “show me how this candidate was scored on this date,” an annual report cannot answer. A hash-chained per-decision record can. Different artifact, complementary to the audit, and where NYC Local Law 144 applies, the real data the bias audit should be computed from.
Free to evaluate. Audit-ready from your first decision.
10,000 decisions per month on the free tier. Upgrade when procurement, an auditor, or counsel asks for the evidence pack.