Atomadic · Compliance Instrument Informational — not legal advice

EU AI Act Risk Screener

Describe an AI system and get an instant, deterministic read on which EU AI Act obligations bite — grounded in the amended text after the Digital Omnibus (Council final green light, 29 June 2026). The classifier logic mirrors a pipeline the engine invented and gated live — the receipts are at the bottom of this page.

Screen a system

Input — system profilesource → register

Try a preset

Plain-language description (optional)

Verdict — determinationregister → verdict

▍ awaiting input
select attributes or a preset,
then run the assessment

How this tool was born — live engine receipts

Nobody typed the pipeline below. The engine was handed one fuzzy intent and it composed a new, verified function out of two already-proven ones — then proved it clean before it existed on disk. Real output from this session:

SROS-V11 · shadow realm · invent throat

# 1 — is this tool novel? ask the resolver (strict bar) synthesize__coordinate__signal__io("assess__source__verdict__io") → coordinate_whitespace // novel. nearest neighbour 'assess__review__verdict__io' rejected # 2 — invent it: compose proven atoms by type (source→register→verdict) synthesize__intent__signal__sovereign("assess__source__verdict__io") → chain : ["assess__source__register__pure", "assess__register__verdict__pure"] gate : landed 1 · repaired 0 · violations_before_repair: [] sha256 : 829c5af0947c6f66b6946d562367007cf5f4dd95df3033bdb300fa8cb9109786 receipt : invent_served // gate PASS → written to shadow → hot-reloaded live # 3 — the code the engine wrote, untouched:

# assess__source__verdict__io  ·  block 0f1f5734…  ·  locator 0546639b0f1f
def assess__source__verdict__io(file_0):
    register_0 = assess__source__register__pure(file_0)
    register_0 = (register_0['grounded'] if (isinstance(register_0, dict) and ('grounded' in register_0)) else register_0)
    verdict_0 = assess__register__verdict__pure(register_0)
    return verdict_0

RESOLVE

Intent → coordinate

The engine names the tool and judges it genuinely new (whitespace), not a duplicate.

COMPOSE

Type-directed invention

Finds two proven atoms whose types chain — like composing lemmas into a theorem. Correct by construction.

GATE

Immune check

Zero violations, content-hash sealed, and parse(render(IR)) == IR enforced before any write.

SERVE

Hot-reload

The new atom became a callable tool in the same session — no restart, no redeploy.

Honest scope: this screener runs a grounded ruleset that mirrors the invented source → register → verdict shape. The engine composed and sealed the pipeline live (above); the legal rules are hand-verified against the cited Act, not generated. Multi-language output rendering isn't wired in the clean-room build yet — so it isn't claimed here.

The timeline actually in force

deadlines moved · June 2026

Obligation	Applies
Art 5 prohibited practices	2 Feb 2025	in force
GPAI model obligations	2 Aug 2025	in force
Art 50 transparency (label AI content, disclose AI)	2 Aug 2026	~ now
New Art 5 bans (nudify / CSAM)	2 Dec 2026
High-risk — Annex III (standalone)	2 Dec 2027	moved from Aug 2026
High-risk — Annex I (embedded)	2 Aug 2028	moved from Aug 2027

The Digital Omnibus simplification package pushed the high-risk deadlines back — Annex III to 2 Dec 2027. But Aug 2 2026 is not cancelled: Article 50 transparency still lands on schedule and touches nearly every consumer-facing AI product.

This is exactly why a deterministic, always-current screener matters: a stale blog — or a hallucinating chatbot — will still cite the dead "Aug 2026 high-risk" date. This one encodes the amended text and returns the same answer every time.