Conformity is the preservation of identity over time. The technical file constitutes the structural evidence that the deployed asset remains the assessed asset. From January 2027, Annex IV of the Machinery Regulation (Regulation (EU) 2023/1230) transforms this documentation from a static design archive into an operational telemetry feed. For systems with self-evolving behavior, the machine authors its own compliance record during execution. The documentation architecture shifts from memory to active structural logging.
The Baseline Documentation Mandate
Annex IV, Part A dictates the baseline contents of the technical documentation. The mandate includes the system description, the procedural risk assessment, protective measures, and residual risk quantifications. The file contains design schematics, applied harmonized standards, and calculation reports that verify physical capability. Production measures bind series units to the assessed design baseline.
The data remains available for market surveillance authorities for ten years post-deployment. The regulatory framework accepts digital formats for the entire archive. For conventional deterministic machinery, this compilation concludes at shipment. The file acts as a closed artifact.
The Operational Directives for Adaptive Systems
Three specific provisions force the file to remain open during operation.
Annex IV, point (m). The file must contain the source code or programming logic of the safety-related software. The code carrying the safety function is a mandatory inspection target. For learned policies, the programming logic manifests as the model weights and architecture.
Annex IV, point (n). Autonomous and sensor-driven machinery must document the general characteristics, capabilities, and limitations of the system. Training data, development processes, testing regimes, and validation methods become mandatory file contents. The tested boundary of reliable operation forms a stated structural limit.
Annex III, section 1.2.1. Control systems exhibiting self-evolving behavior trigger a dedicated requirement set. The system must operate strictly within its defined task and movement space, retaining the capacity for external safety correction. The mandate specifies the continuous recording of the safety-related decision-making process. The system must retain this localized data for one year to demonstrate conformity upon authority request.
The framework mandates a five-year tracing log for interventions and safety-software updates. The regulation prohibits modifications to machine-generated rules that introduce hazard potential. These provisions fuse the compliance file with the machine’s runtime execution trace.
The Measurement Problem in Probabilistic Stacks
The text mandates a retained, producible record of the safety-related decision-making process. The regulation omits the structural definition of this record for neural architectures. A full state capture of a neural policy generates terabytes of tensor activations, burying the compliance evidence under raw volume. A minimal execution log captures only the actuation result, lacking the safety rationale.
Probabilistic behavior introduces a separate decay mechanism. Re-evaluating a past scenario through the same stack yields divergent trajectories. The retention clause assumes the decision process is a discrete, interpretable object. The standard robot-learning stack distributes this process across uninterpretable vector spaces.
Structural Isolation as the Conformity Anchor
Structural isolation resolves the compliance load. Routing every safety-relevant judgment through a singular, deterministic evaluation layer limits the recording obligation to that specific conduit. The abstract regulatory requirements collapse into distinct physical artifacts.
- The 1.2.1 decision record materializes as the layer’s ledger. Each proposed motion generates a verdict and a margin reading within a tamper-evident chain.
- Point (g) test reports gain perfect reproducibility. A deterministic evaluation against a committed scenario cohort yields identical verdicts on any execution hardware.
- Point (n) limitations surface as a measured envelope. The evaluation boundary defines the statistical and physical limits of the system.
- Point (m) exposure narrows to the evaluation layer. The primary behavior policy remains an opaque execution engine.
The architecture isolates the safety logic from the task complexity. A localized, deterministic gate governs the probabilistic load.
The Deterministic Substrate
KAIROS is the deterministic substrate for physical AI. It intercepts the operational stream, evaluating proposed motion envelopes against structural recovery limits before actuation. The kernel acts as a physical firewall, stripping probabilistic decay from the safety decision. The execution yields a replayable, hash-bound record. The identical committed inputs guarantee the identical verdict output across all compute environments.
The empirical baseline is measured. On a frozen, pre-registered cohort of 1,800 simulated MuJoCo scenarios, the velocity-aware gate accepted zero of the 1,200 collision hazards passed by a velocity-blind baseline. The layer autonomously recovered 81% of the safe-when-slowed motion blocked by the baseline, generating zero falsely-safe verdicts. The certified results re-run from committed bytes in under five minutes on commodity hardware, producing a bit-identical match.
The methodology is pre-registered, run once, and reproducible from committed bytes. The engine functions as a certification-supporting evidence instrument for the Machinery Regulation and ISO 10218-2 frameworks. The manufacturer holds the conformity mandate; KAIROS supplies the deterministic proof structure.
This document constitutes structural information. Conformity assessment remains a localized professional discipline.
Read the Research
Three computational studies and two calibration debriefs, complete with reproducibility repositories, are public at anankelabs.io/spindle/. Review the documentation. Re-execute the audits. Send inquiries to the engineering team.