Auditability as a Measurement

Modern European regulation reaches directly into the runtime decision loop. The EU Machinery Regulation (2023/1230), the EU AI Act (2024/1689), the NIS2 Directive (2022/2555), and DORA (2022/2554) all ask the same operational question: when an autonomous system acted or an intrusion unfolded, where was the safety boundary, what did the system decide, and on what basis?

Until now, compliance has tried to answer that question with documentation discipline: static technical files written at design time, post-mortem incident reports stitched together from SIEM logs, and narrative attestations. KAIROS replaces narrative reconstruction with an engineering measurement, giving every runtime safety verdict the same standing as a load-cell reading or a flight data recorder.

The Instrument Problem

Compliance evidence breaks down at the seam where two reviewers must agree. Two Notified Bodies reading the same Annex IV technical file reach different conclusions about whether a learning-based robot is adequately bounded. Two AI auditors reviewing the same system prompts and guardrail logs disagree on whether human oversight was effective. Two CSIRT analysts reconstructing the same breach from raw firewall and authentication logs produce two different 72-hour incident timelines.

The disagreement is structural. In each case, the artifact under review is a description written by a human, summarized by a human, and interpreted by a human. When the instrument that produces the evidence is prose, the evidence drifts between reviewers.

A physical measurement carries a different weight. A load-cell reading of 2.4 kN on a structural beam does not depend on the prose style of the engineer who ran the test. The reading is the artifact, and anyone with the recorded inputs and the calibrated instrument can reproduce the exact number.

What Every KAIROS Envelope Records

KAIROS Substrate sits in the decision loop of physical robots, autonomous AI agents, and security operations centers. On every evaluation, the engine computes the system’s stability margin against a signed deployment policy and emits a self-contained decision envelope.

Every envelope binds four verifiable elements at the moment of evaluation:

1. The Signed Baseline in Force: Before deployment, the domain calibration and deployment policy are locked and cryptographically signed. Every runtime envelope embeds those digests, proving which thresholds and rules governed the system at that exact moment.
2. The Normalized Inputs and Provenance: Whether the input is a joint trajectory, an agent tool-call trace, or a window of security telemetry, the envelope records the normalized pressure and capacity readings alongside the source event or trajectory identifiers that drove them.
3. The Preflight Action Preview: Before a robot actuator commits torque, an AI agent executes an external tool call, or a playbook isolates a critical network zone, KAIROS evaluates the candidate action against the safe set without advancing the operational clock.
4. The Cryptographic Human Oversight Receipt: When stability margin is exhausted and a human operator intervenes, any exception or override requires a cryptographically signed token bound to the specific evaluation request digest, policy version, operator identity, and expiration window.

Because the engine is deterministic, re-running the recorded inputs against the signed policy and calibration produces the identical numerical verdict across independent verification runs.

Why Margin Beats Binary Logs

Standard safety gates and SIEM rules emit binary verdicts: an action either passed or failed, an alert either fired or stayed silent. A binary log tells an assessor nothing about how close the system operated to failure until the boundary has already been crossed.

Every KAIROS evaluation records a continuous stability margin: the signed headroom between the system’s current control capacity and the policy floor, paired with a forward horizon showing how much room remains before boundary contact.

For an engineering team or an external regulator, continuous margin answers the questions a binary log cannot:

• Approach velocity: How rapidly a robot trajectory, an agent session, or a network zone lost headroom before intervention occurred.
• Causal drivers: Which pressure driver pushed the margin toward the floor, and which defensive or physical control was closest to giving way.
• Intervention timing: Whether early-warning directives gave the planner, agent, or analyst enough lead time to recover inside the safe interior before a hard stop was required.
• Policy sensitivity: How the same operational trace behaves when replayed under a stricter or looser policy floor during conformity testing or post-incident review.

One Measurement Standard Across Three Regimes

Because all three domain adapters run on the same deterministic substrate, the same architectural evidence model satisfies statutory mandates across physical machinery, high-risk AI, and critical infrastructure:

• EU Machinery Regulation (2023/1230): From 20 January 2027, machinery with self-evolving behavior requires formal conformity assessment and a one-year operational decision log under Annex III §1.2.1 and Annex IV. KAIROS provides pre-actuation safety envelopes at runtime and offline validation bundles across 1,800+ deterministic simulation scenarios.
• EU AI Act (2024/1689): High-risk AI obligations applying from 2 August 2026 mandate continuous risk management (Article 9), automatic event logging (Article 12), effective human oversight (Article 14), and documented robustness (Article 15). KAIROS binds every agent step into a continuous session posture record, previews every tool call before execution, and records human overrides as cryptographically verifiable tokens.
• NIS2 (2022/2555) and DORA (2022/2554): Essential entities and financial institutions face a 24-hour early-warning clock and a 72-hour incident notification deadline from the moment of awareness. Instead of spending weeks reconstructing a breach from scattered SIEM logs, KAIROS Cyber maintains a pre-indexed incident timeline where every zone escalation already carries its explanation record: affected zone, stability margin, dominant pressure drivers pinned to source event IDs, weakest control, and downstream blast radius.

What Changes in the Audit Room

When an operator hands a KAIROS replay bundle to a Notified Body, an AI Office auditor, or a CSIRT, the review shifts from grading a narrative to verifying a measurement.

Both parties load the same input trace, the same signed calibration manifest, and the same deployment policy into the verification runner. Both arrive at the same stability margins, the same gate verdicts, and the same cryptographic hash chain. Compliance ceases to be a retrospective documentation project and becomes a verifiable property of the runtime architecture.

Continue Reading by Regime

• Physical AI & Robotics: The Machinery Regulation and the Machines That Learn and The Technical File for a Machine That Learns
• Autonomous AI & Agents: Structural Conformance Under the EU AI Act
• Cybersecurity & Critical Infrastructure: Incident Reconstruction Under NIS2