On 20 January 2027, the EU Machinery Regulation (Regulation (EU) 2023/1230), becomes the sole legal route for placing machinery on the European market. It replaces the Machinery Directive 2006/42/EC, the framework that governed machine safety in Europe for two decades, and it brings software that makes safety decisions, including AI that keeps revising its own behaviour, inside the scope of formal conformity assessment.
Every safety regime is an argument with time. A machine outlives the assumptions of its designers, and the law’s job is to fix responsibility at the moment those assumptions were declared sound. The new regulation exists because a class of machines has emerged whose behaviour keeps moving after that moment: machinery driven by artificial intelligence.
From Directive to Regulation: One Text, Everywhere, on the Same Day
The legal form matters. A directive instructs each member state to write its own national law, and 27 parliaments produce 27 slightly different texts. A regulation applies directly and identically across the entire European Union from its application date. On 20 January 2027 the old patchwork is repealed and a single text governs.
The substance matters more. Under the new regulation, a safety function, meaning any function of a machine whose failure raises the risk to people (keeping a robot arm slow while a human stands beside it, for example), falls within conformity assessment even when it is implemented purely in software. Conformity assessment is the documented process by which a manufacturer demonstrates that a product meets the law’s essential health and safety requirements before affixing the CE marking, the visible declaration that a product may circulate in the EU market.
The regulation goes further for machines that learn. Annex I lists the categories of machinery and safety components subject to the strictest assessment procedures, and Part A of that list includes safety components with fully or partially self-evolving behaviour using machine-learning approaches, together with machinery embedding such systems. Self-evolving means the behaviour continues to change after the machine leaves the factory. For this category the manufacturer must involve a notified body, an independent conformity-assessment organisation designated by an EU member state; the self-certification route available for most conventional machinery closes here.
Two further changes deserve notice. Digital documentation becomes acceptable: instructions for use and the declaration of conformity may be supplied in digital form. And the ladder of obligation is explicit: manufacturers carry the primary duties, importers and distributors carry verification duties for the machinery they move into and through the market, and any party that substantially modifies a machine, a systems integrator rebuilding a production line included, inherits the manufacturer’s obligations for the result.
The Clock Runs Backward From January 2027
Industrial machinery is designed on multi-year cycles. A robotic cell specified this year will reach its customers under the new regime, because conformity attaches at the moment a machine is placed on the market (offered or put into service in the EU), and from 20 January 2027 that moment belongs to Regulation 2023/1230 alone. Design decisions frozen in 2026 determine which conformity route is available in 2027.
The Machinery Regulation also arrives in company. The EU AI Act, the horizontal law governing artificial intelligence, reaches general application on 2 August 2026. Its high-risk regime, the tier carrying the strictest obligations, reaches AI systems embedded in sectorally regulated products, machinery among them, on 2 August 2028, a deadline the EU’s digital omnibus package moved from the original August 2027. The Parliament and the Council adopted that package in June 2026, and it enters into force on the third day after its publication in the Official Journal. Across adjacent regimes, from NIS2 in network security to DORA in financial services, the same demand recurs: operators of consequential systems must produce evidence that an outside party can examine and re-derive.
Why Learning-Based Machinery Has an Evidence Problem
Classical machine safety rests on determinism. A relay, an interlock, a fixed control program: given the same inputs, each produces the same output, every time. Evidence for such a system is an engineering artifact. You test the envelope, you document the result, and the file waits in a drawer until an auditor asks.
The regulation converts that file into a market-access requirement. For AI-based safety behaviour in the Annex I category, a notified body must be able to examine the safety case and reach the manufacturer’s conclusion by its own reading. The paperwork becomes load-bearing: European revenue rests on it the way a floor rests on a beam.
Here the architecture of modern AI grinds against the architecture of assessment. A learning-based stack is probabilistic by construction: its outputs carry intentional randomness, and they shift with random seeds, floating-point ordering, hardware, and model updates, so rerunning yesterday’s scenario can produce a different trajectory today. The evidence such a stack yields naturally is statistical: distributions of outcomes, aggregate failure rates across millions of runs.
An assessor’s question has a different shape. It is singular and forensic: on this input, on this day, why was this motion judged safe. The gap between what the stack emits and what the assessor requires deserves a precise name: the gap is reproducibility of the safety decision, not capability of the robot. Assessment needs the decision to be repeatable; the stack offers it as a probability.
Four Readers, Four Exposures
For an investor, the regulation turns conformity into a revenue gate. A robotics company’s date of European market access now runs through notified-body assessment, and the diligence question becomes concrete: show the evidence pipeline that survives third-party re-derivation, and show its cost and duration.
For a decision maker at a robotics company, the pressure lands in the design freeze. The documentation architecture chosen this year, what is recorded, how it is committed, whether an outsider can rerun it, determines the assessment route open when the machine ships into the 2027 regime. Evidence has joined payload and cycle time as a product requirement.
For a regulator or notified body, the assessable unit is shifting underneath the profession. Guards, interlocks, and stop circuits are inspectable by hand; self-evolving behaviour is inspectable through its records alone. The decision record, complete and replayable, becomes the object of assessment.
For an insurer, underwriting AI-driven machinery means pricing a machine whose behaviour at the moment of loss must be reconstructed after the fact. A replayable record of what the machine decided, and why, converts claims investigation from expert conjecture into replay. Actuarial confidence wants exactly the evidence the regulation now demands.
A Reflex Arc for Physical AI
Every animal that survives contact with the world carries a reflex arc: a short, fast circuit that evaluates a motion before the body commits to it. The spinal cord answers before the brain finishes composing an opinion. Physical AI, meaning AI that moves machines through the physical world, already has the brain. AnankeLabs builds the arc.
KAIROS is a deterministic evaluation layer for physical AI: deterministic in the strict sense that the same committed inputs produce the same verdict, bit for bit, on any machine. It evaluates a robot’s proposed motions before actuation and leaves behind tamper-evident, replayable evidence of every decision, a record in which any alteration is visible.
The judgment it renders is the reflex judgment: whether a proposed motion preserves a recoverable future, a state from which the machine can still return to safety. Where policy allows, the response is graded: “slow down” rather than only “stop.” A reflex that modulates is a reflex the body can work with.
The anatomy around it stays whole. Planners, controllers, simulators, and emergency stops all remain in place; KAIROS evaluates what they are about to do and proves afterwards what was decided and why. The nervous system gains a reflex and keeps its brain.
The measured result, stated with its scope: on a frozen, pre-registered cohort of 1,800 simulated scenarios (in MuJoCo, a physics simulator; simulation evidence, a single robot embodiment), the velocity-aware gate accepted zero of the 1,200 collision hazards that a velocity-blind baseline gate accepted, while autonomously recovering 81% of the safe-when-slowed motion the baseline blocked, at zero falsely-safe verdicts.
Where the Evidence Stands Today
AnankeLabs is not third-party certified and does not claim to be. The methodology is engineered to be certifiable: pre-registered, run once, held out, reproducible from committed bytes. Today’s offering is certification-supporting evidence for a customer’s own conformity case, under ISO 10218-2 (the safety standard for industrial robot applications and their integration), CE marking, and the Machinery Regulation. The manufacturer owns the safety case; KAIROS supplies the reproducible evidence.
The evidence carries its own proof of replay. The certified results re-run from committed bytes in under five minutes on commodity hardware, ending in a bit-identical match, by any credentialed party.
This article is information, and conformity is a decision for professionals. Readers weighing their obligations under the regulation should consult their own conformity-assessment advisors.
Read the Research
Three computational studies, each with a reproducibility repository, and two calibration debriefs are public at anankelabs.io/spindle/. Read them, rerun them, and write to us with what you find.