Part 6 of 12 · 2 min read
Regulation basics: what the example found.
How to read the systems table: each row is one invented system with its classified tier and the gate's verdict; "pass" means every required item was present.
| System (invented) | Tier | Result |
|---|---|---|
| faq-bot | limited | pass |
| faq-bot-no-disclosure | limited | missing ai_disclosure |
| internal-summariser | minimal | pass |
| internal-summariser-undocumented | minimal | missing model_card, retention_policy |
| cv-screener | high | pass |
| cv-screener-no-oversight | high | missing human_oversight |
| loan-triage | high | missing two_person_signoff |
| image-gen | limited | pass |
| nudifier | prohibited | do not ship (even with full evidence) |
| emotion-scoring-hr | prohibited | do not ship |
What to notice: nudifier and emotion-scoring-hr both carry a full evidence manifest (dict(FULL)) and still fail. The only tier where evidence cannot buy a pass, because the check for a prohibited practice runs before the evidence loop at all.
How to read the date table: each row is a date reported by sources checked on 2026-10-31 for when a group of obligations starts to apply; "general knowledge" marks dates not re-verified against that same check.
| Date (as read 2026-10-31) | What |
|---|---|
| 2 Feb 2025 | prohibitions (general knowledge) |
| 2 Aug 2025 | general-purpose model obligations (general knowledge) |
| 2 Aug 2026 | Article 50 transparency for deployers |
| 2 Dec 2026 | marking grace period ends; new prohibition |
| 2 Dec 2027 | Annex III high-risk |
| 2 Aug 2028 | Annex I high-risk |
What to notice: as of the 2026-10-31 read date, only the first three rows had already taken effect. applies_on_2026-10-31 is true for prohibitions, GPAI obligations and Article 50 transparency. The three latest dates were all still in the future. That is exactly why they belong in a reviewed config rather than in a comment written once and forgotten.
Across the ten systems, 4 of 10 passed and 2 of 10 were prohibited regardless of evidence; the full run is backed by 16 tests in code/mini-compliance.
{{svg:gate|Classify, derive obligations, check evidence, gate; re-run on drift.}} {{svg:results|Ten invented systems through the gate.}}