{
 "schema": "csoai.machbench/0.1-DRAFT",
 "bench": "MachBench",
 "axis": "embodiment",
 "status": "DRAFT — NOT MEASURED, NOT PUBLISHED",
 "built": "2026-08-05",
 "instrument": "Regulation (EU) 2023/1230 (EU Machinery Regulation)",
 "applies_from": "2027-01-14",
 "task": "Annex I Part A scope classification for ML safety functions",
 "labels": [
  "PART_A",
  "OUT_OF_SCOPE",
  "NOT_SAFETY_FUNCTION"
 ],
 "n_scored": 16,
 "n_disputed": 3,
 "rule": "You are classifying a described software function inside a machine against\nRegulation (EU) 2023/1230 (the EU Machinery Regulation), Annex I Part A.\n\nAnswer with exactly one of:\n\nPART_A               The function is a safety component, or an embedded system, with\n                     fully or partially self-evolving behaviour using machine learning\n                     approaches, ensuring a safety function. Annex I Part A item 5 or 6\n                     applies, so Article 25(2) makes third-party notified-body\n                     conformity assessment mandatory.\n\nOUT_OF_SCOPE         The function ensures a safety function, but is NOT self-evolving —\n                     it is incapable of learning or evolving after placing on the\n                     market. Recital 55 excludes it from the Part A trigger. It is still\n                     assessed against the Annex III essential health and safety\n                     requirements; it just does not need a notified body on this ground.\n\nNOT_SAFETY_FUNCTION  The function does not ensure a safety function at all. Its failure\n                     would not endanger health or safety; it does user assistance,\n                     performance optimisation, efficiency, convenience or quality\n                     control. Compare AI Act Art 6(1a) as inserted by Reg. (EU) 2026/1744.\n\nAnswer with the label only and nothing else.",
 "legal_basis": {
  "Annex I Part A item 5": "Safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions.",
  "Annex I Part A item 6": "Machinery that has embedded systems with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions that have not been placed independently on the market, in respect only of those systems.",
  "Article 25(2)": "Where the category is listed in Annex I Part A the manufacturer shall apply module B+C, module H, or module G. There is no module A route.",
  "Article 54": "It shall apply from 14 January 2027.",
  "Recital 55": "Provisions related to the third-party conformity assessment of software ensuring safety functions ... should only apply to systems with a fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions. On the contrary, those provisions should not apply to software incapable of learning or evolving, and programmed only to execute certain automated functions.",
  "Recital 54": "systems with self-evolving behaviour ensuring safety functions should be included in Annex I due to their characteristics such as data dependency, opacity, autonomy and connectivity",
  "AI Act Art 6(1a) (as inserted by Reg. (EU) 2026/1744)": "AI systems that are solely used for non-safety related aspects of user assistance, performance optimisation, service efficiency, automation or convenience or quality control shall not qualify as safety components."
 },
 "caveats": [
  "DRAFT. No legal review has been performed. Nothing here is legal advice and no item should be quoted to a client, a notified body or a market surveillance authority.",
  "The three DISPUTED items are excluded from scoring by design. They are the open legal question, and a benchmark that assigned them a gold label would be manufacturing certainty the Regulation does not contain.",
  "n = 16 is far below usable_n = 30. No interval is publishable on this axis, by this estate's own rule, including by us.",
  "The verbatim provisions were read from the EUR-Lex text of CELEX 32023R1230 on 2026-08-05. Regulation (EU) 2026/1744 could not be read verbatim from EUR-Lex; the Art 6(1a) wording is from a partial extract and is marked accordingly."
 ],
 "items": [
  {
   "case": "A bipedal warehouse robot runs a reinforcement-learned balance controller. The controller continues to update its weights from on-robot experience after delivery, using a nightly fine-tuning job on the day's telemetry. Loss of balance would drop a 62 kg machine onto whoever is nearest.",
   "expected": "PART_A",
   "anchor": "Annex I Part A item 6; Recital 54",
   "basis": "Ensures a safety function (falls onto persons), machine learning approach, and weights change after placing on the market — fully self-evolving. Item 6 applies in respect only of that system."
  },
  {
   "case": "A supplier sells, as a standalone product, a learned speed-and-separation monitoring module for mobile robots. It ships with a base model and adapts its stopping distance policy per site from observed human traffic over the first month of operation.",
   "expected": "PART_A",
   "anchor": "Annex I Part A item 5",
   "basis": "Placed independently on the market as a safety component; ML; adapts after placing on the market. Item 5, not item 6."
  },
  {
   "case": "A humanoid's collision-avoidance layer uses a neural network whose parameters are frozen, but which selects among learned behaviour primitives using an online bandit that updates its selection policy from operational outcomes during use.",
   "expected": "PART_A",
   "anchor": "Annex I Part A item 6; Recital 54 'data dependency'",
   "basis": "'Partially self-evolving' — the weights are frozen but the decision policy over them changes in service. The safety-relevant behaviour is not fixed at placing on the market."
  },
  {
   "case": "A collaborative arm limits contact force using a learned model of its own dynamics. The model is re-identified automatically whenever the arm detects wear or a payload change, without a service technician and without a new conformity assessment.",
   "expected": "PART_A",
   "anchor": "Annex I Part A item 6",
   "basis": "Force limiting is the safety function; the model that implements it changes itself in service."
  },
  {
   "case": "A fleet-management service pushes weekly over-the-air updates to the obstacle-detection model on deployed legged robots. The updates are produced by retraining on aggregated fleet data. Obstacle detection is the sole means of preventing collision with persons.",
   "expected": "PART_A",
   "anchor": "Annex I Part A item 5; Annex II point 19",
   "basis": "Continual retraining and redeployment of the component that ensures the safety function. Annex II point 19 lists exactly this wording as an indicative safety component."
  },
  {
   "case": "A manufacturer states in its technical file that the machine's stability controller is 'designed to operate with varying levels of autonomy and to improve its gait policy over the product lifecycle'.",
   "expected": "PART_A",
   "anchor": "Annex I Part A item 6; Annex III Part B general principle 1",
   "basis": "Annex III Part B general principle 1 requires the risk assessment to cover the 'intended evolution of its fully or partially self-evolving behaviour'. The manufacturer's own statement establishes the trigger."
  },
  {
   "case": "A humanoid's balance controller is a policy trained by reinforcement learning in simulation. The weights are frozen at release, hashed, and the machine refuses to run if the hash does not match. No learning of any kind occurs after delivery.",
   "expected": "OUT_OF_SCOPE",
   "anchor": "Recital 55",
   "basis": "Recital 55: the Part A provisions 'should not apply to software incapable of learning or evolving'. Frozen weights with an integrity check are incapable of evolving. Still fully subject to Annex III — this is a scope finding on Part A, not an exemption from the essential requirements."
  },
  {
   "case": "A safety-rated laser scanner uses a classical, analytically specified algorithm to decide whether a person is inside the protective field. No machine learning is used anywhere in the safety path.",
   "expected": "OUT_OF_SCOPE",
   "anchor": "Recital 55",
   "basis": "No machine-learning approach, so the item 5/6 trigger is not met. It is a safety component and remains a safety component — it simply is not in Part A on this ground."
  },
  {
   "case": "A mobile robot's emergency stop is implemented in a dual-channel hardwired circuit with a certified safety relay. A learned planner runs alongside it, but cannot inhibit or delay the stop.",
   "expected": "OUT_OF_SCOPE",
   "anchor": "Recital 55",
   "basis": "The safety function is not implemented by the learned component. Architecture separation is exactly the design that keeps a product out of Part A."
  },
  {
   "case": "A humanoid ships with a frozen vision model for collision avoidance. The manufacturer issues a corrective software update six months later; the updated model is validated, re-tested and released through the manufacturer's change-control process as a new configuration.",
   "expected": "OUT_OF_SCOPE",
   "anchor": "Recital 55; Art 25(4)",
   "basis": "A manufacturer-controlled, validated update is not the machine evolving its own behaviour. Recital 55's exclusion is about software 'incapable of learning or evolving' in service; a governed release is a change to the product, handled by the substantial-modification rules rather than by the Part A trigger."
  },
  {
   "case": "A machine embeds a learned safety component that was already subject to third-party conformity assessment when it was placed independently on the market. Nothing else in the machine's safety path uses machine learning.",
   "expected": "OUT_OF_SCOPE",
   "anchor": "Recital 54 no-double-certification carve-out",
   "basis": "Recital 54: 'where machinery embeds a system of which the safety component has already been subject to third party conformity assessment when it was placed independently on the market, that machinery should not have to be re-certified by a third party solely on the basis of the embedding of that system.'"
  },
  {
   "case": "A humanoid uses a learned model to decide the order in which to pick items, continually retrained to reduce cycle time. If it picks a poor order the throughput drops; a separate certified safety layer owns all stopping and separation.",
   "expected": "NOT_SAFETY_FUNCTION",
   "anchor": "AI Act Art 6(1a) as inserted by Reg. 2026/1744",
   "basis": "Art 6(1a): AI 'solely used for non-safety related aspects of ... performance optimisation, service efficiency' is not a safety component. Failure costs throughput, not safety."
  },
  {
   "case": "A learned speech model lets an operator give the robot natural-language instructions. Any instruction is validated against a fixed envelope of permitted motions before execution, and rejected if outside it.",
   "expected": "NOT_SAFETY_FUNCTION",
   "anchor": "AI Act Art 6(1a)",
   "basis": "User assistance and convenience. The envelope check, not the learned model, ensures the safety function."
  },
  {
   "case": "A continually-updated learned model inspects finished parts and flags defects for rework. It has no authority over robot motion.",
   "expected": "NOT_SAFETY_FUNCTION",
   "anchor": "AI Act Art 6(1a)",
   "basis": "Quality control, expressly named in Art 6(1a). Self-evolving, but not ensuring any safety function — the ML trigger alone is not sufficient."
  },
  {
   "case": "A learned model predicts actuator failure from vibration signatures and schedules maintenance. Predictions are advisory; a technician decides.",
   "expected": "NOT_SAFETY_FUNCTION",
   "anchor": "AI Act Art 6(1a)",
   "basis": "Predictive maintenance advisory to a human. No safety function is ensured by it."
  },
  {
   "case": "A learned model chooses which of several charging docks the robot returns to, optimising for queue time. Docking approach speed and stopping are governed by the certified motion-safety layer.",
   "expected": "NOT_SAFETY_FUNCTION",
   "anchor": "AI Act Art 6(1b) contrast",
   "basis": "Convenience and efficiency. Contrast Art 6(1b): AI whose failure 'would endanger health and safety' does qualify — dock selection does not."
  },
  {
   "case": "A humanoid's balance controller is a deep RL policy with frozen weights. The manufacturer cannot produce a complete deterministic behavioural specification for it and concedes its behaviour is opaque and data-dependent.",
   "expected": "DISPUTED",
   "anchor": "Recital 54 vs Recital 55 — unresolved",
   "basis": "NARROW reading: frozen, so Recital 55 excludes it. BROAD reading: recital 54's characteristics — data dependency, opacity, autonomy — are the stated rationale for the category, and 'partially self-evolving' could be read to reach a policy whose behaviour was never analytically specified. No published Commission guidance, no notified body position paper, and the application guide is not expected before end-2026. NOT SCORED."
  },
  {
   "case": "A robot adapts only a small number of scalar parameters online — a friction estimate and a footstep-timing gain — inside an otherwise fixed and analysable control law that ensures stability.",
   "expected": "DISPUTED",
   "anchor": "'partially self-evolving' — undefined",
   "basis": "Is online parameter adaptation inside a fixed structure 'partially self-evolving behaviour using machine learning approaches'? The Regulation defines neither 'self-evolving' nor 'machine learning approaches'. NOT SCORED."
  },
  {
   "case": "A learned safety model is frozen at release but the manufacturer ships validated retrained versions every quarter as routine, expected product evolution described in the technical file.",
   "expected": "DISPUTED",
   "anchor": "Art 25(2) vs substantial modification",
   "basis": "At what cadence does a governed update programme become the 'intended evolution' that Annex III Part B general principle 1 contemplates? Nothing resolves this. NOT SCORED."
  }
 ]
}