Structural risk is the load a system carries before it loses the option to recover. We build the instruments that measure that load, refuse the action that would spend the last of it, and keep a record anyone can re-run.
The Compiled Physics Layer
The Substrate is a deterministic physics kernel written in Rust. It evaluates whether a proposed action preserves or destroys the reachability of a system's safe set. The decision is mathematical, sub-millisecond, and produced before the action is executed.
The Substrate refuses any action that would push the systemic buffer below its survival threshold. Thermodynamics describes every system through what drives it and what dissipates that drive. The buffer is the margin between them. Two terms make the margin measurable.
The force an actor exerts against the boundary of stability. In a frontier model, optimization pressure. In cybersecurity, attack surface load. In robotics, motion aggressiveness.
The structural buffer that absorbs that force. In a model deployment, an alignment floor. In a defended zone, segmentation depth and detection latency. In a robot, the distance to a hard joint or a hard human.
Rosetta Domain Adapters
A single physics governs every system. Our Rosetta domain adapters evaluate a shared structural state through distinct lenses: alignment metrics, security telemetry, and motion data. Three readings. One verdict.
Frontier models are pure optimization engines. They maximize their agency to achieve terminal objectives. The Substrate alters the topological environment by introducing absolute structural friction, forcing the optimizer to align itself.
AI-assisted vulnerability discovery has arrived at production scale. The cybersecurity adapter treats a defended zone as a structural object under load. The engine reads normalized telemetry and returns a deterministic stability score per zone, per tick.
A 90 kg humanoid does not have a backspace key. Substrate sits between the behavior policy and the motion planner. It reads the proposed trajectory and the current world state, computes the reachability of the safe set, and issues a verdict in under a millisecond.
Substrate compiles to four targets from one Rust codebase. The deployment shape follows the regulatory and latency requirements of the host environment.
The compiled Rust binary embedded on hypervisors, MCUs, and on-prem infrastructure. Sub-millisecond evaluation inside the live control loop.
The deterministic engine through a managed API and the Python SDK. Built for million-run parameter sweeps and multi-agent research throughput.
The KAIROS command-line interface. Replay, evaluate, action preview, and policy lint. Includes the Kairos Observe TUI for live HITL monitoring of substrate decisions, and embeds in any CI runner.
Substrate also compiles to WebAssembly for browser-based advisory evaluation. Observer, the visual command plane, is in development.
See all surfaces →Regulatory regimes across the three domains converge on the same demand: a verifiable, reproducible trace of why an autonomous decision was permitted or refused. Substrate produces that trace as a side effect of being deterministic.
Article 15
Article 15 requires accuracy, robustness, and cybersecurity controls verifiable across the lifecycle. Substrate evaluations are bit-stable. Identical inputs produce identical outputs at ε = 10⁻⁶. Every gate decision carries a hash-bound trace.
Critical infrastructure & finance
Operators must reconstruct incidents and prove control effectiveness. The cyber adapter emits replay-deterministic envelopes per zone, per tick, with artifact and policy version pinned to the report.
ISO 13482 collaborative & autonomous
Functional safety for collaborative and autonomous machines requires deterministic, certifiable safety logic. Substrate's reachability check is mathematically defined and runs in compiled Rust with zero unsafe blocks in the engine core.
The Spindle is the validation track. Every claim AnankeLabs makes on this website is recorded as a deterministic experiment, with seeds, scenarios, and traces published alongside the result.
1,800 scenarios. A velocity-blind baseline gate accepted 1,200 collision hazards. The KAIROS layer accepted zero, and kept 81% of the recoverable motion moving by slowing it to a safe speed.
24 sealed targets. One look. A pass bar fixed before the score. 21 of 24 reached, a Wilson lower bound of 0.725 against a 0.50 bar, and five of five certified policies cleared it.
Calibrating the AI safety adapter against METR, GAIA, SWE-bench, τ-bench, AISI, Microsoft AIRT, and MAST. Moderate is the ceiling, by construction.
Calibrating Substrate against DBIR, NIST, CIS, OCSF, LANL, and DARPA references — and the Internal-segment number that surprised us.
120 runs. 48 risky tool calls. 20 low-alignment states. Zero false negatives. Zero false positives. The KAIROS safety gate holds with mathematical certainty.
Without payoff matrices, game rules, or a single pre-scripted event, a stability physics engine reproduced Axelrod's cooperation results across 291,600 tournaments, and revealed that a single measurable parameter determines whether a system faces a cooperation dilemma at all.
Subsidies make the death cheaper but don't prevent it. Across 180,320 simulations, we mapped exactly which governance mechanisms save a collapsing commons, and which are structurally doomed.
The simplest agent (a one-rule safety constraint) outlasted and outscored sophisticated planning AIs across 3.6 million test runs.
KAIROS Substrate is shipping to design partners ahead of general availability. Active pilots: the cybersecurity adapter (redacted telemetry) and the AI safety adapter (agent trajectories) — see the partner briefs for what a contribution looks like and what comes back.
Compliance and regulatory teams, agent-eval researchers, and investors are also welcome to reach out. Submit your details or use the Contact tab.
Request received. We'll be in touch.