Substrate
Compute the boundary before the action commits.
Autonomous systems now act faster than any human can review. Yet the safety layers wrapped around them are built from the same material as the models they guard: neural classifiers, confidence scores, and statistical thresholds.
A probabilistic guard inherits the failure modes of the system it watches. It drifts with its training distribution, bends under optimization pressure, and reduces every post-incident audit to a guess when two identical runs return two different answers.
A safety boundary must be calculated from state geometry. KAIROS Substrate evaluates the system against a cryptographically signed policy, previews the consequence of the proposed action without advancing the clock, and commits the verdict to a deterministic log.
KAIROS Substrate is a compiled Rust kernel with zero runtime dependencies and no unsafe code in its core. On every call, it validates telemetry freshness and signatures, clones the live state to preview the proposed action across the reachability field, and issues a verdict in under a millisecond: before an actuator moves, a tool call fires, or a network zone collapses.
Kernel invariants
Every evaluation runs through the same fixed-order pipeline: cryptographic policy verification, reachability solve, preflight action preview, and digest-anchored trace emission.
01 / 04 · Same inputs, same verdict
Fixed-order execution from seed to digest.
02 / 04 · Fly-by-wire action preview
Clone the state. Preview the move. Gate before impact.
03 / 04 · Air-gapped and monotonic
Certified floors tighten; they never relax.
04 / 04 · Cryptographic provenance
Every verdict and human escalation is sealed.
Developer surfaces
A single Rust codebase compiles into five integration targets. Every surface calls the same reachability solver and enforces the same signed deployment policy.
$ cat request.json | kairos evaluate --config artifact.json --policy policy.json
{
"requestId": "eval-001",
"decision": "PASS",
"reasonCode": "NONE",
"mode": "state_gate",
"evaluation": {
"currentGamma": 0.68,
"gammaFloor": 0.20,
"currentLambda": 0.89,
"stability": 2.40
}
}$ curl https://api.anankelabs.io/v1/simulations -H "X-Api-Key: krs_…"
{
"data": [
{ "id": "a1b2c3d4-…", "scenarioId": "capability-divergence", … }
],
"pagination": { "limit": 50, "offset": 0, "total": 142 }
}kairos · kairos observeSingle binary for preflight evaluation, RFC 6901 trace diffing, and the live observe TUI with 70/30 drift-vs-optimal radar projections, Sentinel assessments, and HITL operator controls.C ABI · kairos_engine.hShared (.so / .dylib / .dll) or static (.a) library with auto-generated C headers. Panic-isolated, mutex-synchronized handles link directly into real-time C, C++, or Rust control loops.kairos_engine · PyO3Compiled wheel for high-throughput seed sweeps, closed-loop RL reward shaping, and deterministic trace comparison across MuJoCo, Isaac Lab, and gym environments.api.anankelabs.io/v1 · kairos hitlHTTPS endpoints with OpenAPI 3.1 schemas for hosted simulation and trace evaluation, plus the self-hosted HITL Coordinator for RSA-PSS override issuance and distributed CAS retry ledgers.kairos-wasm · wasm32Size-optimized browser and Node build exposing WasmEngine and advisory WasmSession bindings. Runs the live simulation on our homepage.Deployment topologies
In-process control loop
Linked directly beside the motion planner, agent harness, or controller. Evaluates candidate actions in-process and returns a verdict in under a millisecond before execution commits.
Automated regression gate
Runs committed scenario suites on every pull request, diffs trace bundles and SHA-256 digests against frozen baselines, and fails the build on any behavioral or margin divergence.
Air-gapped & cluster HA
Operates completely offline inside classified or industrial networks with machine-fingerprinted licenses, backed by SQLite or multi-node HITL Coordinator clusters for durable escalation state.
Calibration & RL sweeps
Executes thousands of seeded episodes through native Python bindings to calibrate domain artifacts, shape safe-by-design RL rewards, and generate portable conformity evidence bundles.
Specification
Rosetta domain adapters
Pre-Actuation Safety for Physical AI
Explore RoboticsDeterministic Governance for AI Agents
Explore AI SafetyStructural Early Warning for Attack Pressure
Explore CyberKAIROS Substrate is available for technical evaluation and design partnerships across all three domains. Evaluations scope a concrete control loop or decision boundary, run against recorded trajectories, simulation scenarios, or redacted zone telemetry, and return a digest-anchored containment report that replays from committed bytes.
Reference specifications are available for robotics, the AI safety adapter(agent trajectories), and the cybersecurity adapter (redacted zone telemetry). Select a domain on the right to request evaluation access or start an integration review.
Request received. We'll be in touch.