Afterwave

Research

How to tell whether a model update made a machine less safe, and what regulators will ask for when it does.

predicted at-risk region (prone)confirmed regression (0.21 lux, 27.6 m)0.1110100100010000510203040light (lux, log scale)initial range (m)0.100.80prior P(violation)
Technical report AW-TR-2026-01 · October 2026 · 25 min

Release gating for learned perception in closed-loop autonomy

Change-guided falsification with paired confirmation: a statistical test for whether a model update made a machine less safe, and a case study in which a better model is blocked.

Read
19 Jul 2023Regulation enters into forceArt. 5420 Jan 2024Provisions on notifying authorities and notified bodies applyArts 26–4220 Jul 2024Commission's delegated powers, including amending Annex I, applyArt. 6(2)–(6), Art. 4727 Jul 2026Digital Omnibus on AI in force: machinery moves to AI Act Annex I, Section BReg. (EU) 2026/174420 Oct 2026Member States' penalty rules applyArt. 5020 Jan 2027Regulation applies; Directive 2006/42/EC repealedArts 51, 542 Aug 2028AI-specific requirements for machinery, via Machinery Regulation delegated actsMR Art. 8 as amended
Regulatory analysis · October 2026 · 30 min

The EU Machinery Regulation and machine learning: what applies from 20 January 2027

A close reading of Regulation (EU) 2023/1230 for teams shipping learned safety functions: scope, conformity assessment, the new requirements on self-evolving systems, logging, software updates, and how the 2026 Digital Omnibus rewired its relationship with the AI Act.

Read