Aletheia
The Unforgetting.
A river of fading records, recovered by an independent replay.
Time flows left to right, or downward in portrait. Recorded native transitions weave the river; its path fades into grey ink as it ages. A separate engine reconstructs the same run. Only matching per-tick SHA-256 digests restore crisp brass behind the needle. Tiny seals mark advisory changes, warning onsets, native losses and Gamma/Lambda regimes. Warning motion pinches the current and deepens its haze. Select a retained tick to re-enter the record; Return to live resumes the original run.
Independent replay
Independent reconstruction from origin, within the same fixed WASM build in this browser. The digests compare the versioned semantic observation at the displayed tick: state, agents, reachability, warnings, decisions and native event records.
Records cover the latest 256 ticks. Re-entry reconstructs earlier engine steps before comparing the retained stretch. Digest agreement describes this observation schema; it is not a signature, physical-safety result or cross-platform guarantee.
Live KAIROS
Six artistic impressions of one evolving simulation. The KAIROS engine is written in Rust and compiled to WebAssembly, running locally in your browser in a dedicated Web Worker. Its calculations of reachability, branching futures and basin loss shape the artwork as it unfolds.
Switch artworks to see the same run through a different lens. Choose a scenario or adjust the seed, Lambda and Gamma, then apply your settings to restart. In Aletheia, Replay independently reconstructs the retained record; in the other field views it restarts the active run.
Two live KAIROS engines
Cooperative Blocks runs its own simulation: two KAIROS engines in a separate Web Worker, each given the board as an explicit topology. The page knows the rules of the game; the engines decide every move. The six field artworks pause while it plays and resume where they left off.
- Where the piece can still go, as the position engine sees it: brighter where more landing value remains.
- Around the piece: the orientations the rotation engine can still reach.
- The path toward the shared target; the dashed outline is the target landing.
- Landings that would leave a hole.
- Each agent's move. Amber when the board refuses it.
Explore the field
Offsets are relative to the current position. Future: 0–48 ticks. Branch: −48 to +48. Use integers.
Space plays or pauses. ← / → switches artworks. On the canvas, Shift + arrows explores a point; Enter queries it.
Tap the artwork to explore a point.
01 / REPLAY CONTROLS
PENDING · Waiting for original records.
Original — · replay — · verified —
Waiting for original history.
Shift + arrows on the canvas select a retained tick; Enter re-enters it.
Digest pair · pending
Structural Safety and Evidence for Autonomous Systems
A 90 kg humanoid has no backspace key.
Autonomous robots and AI agents now act at machine speed in the physical and digital world. A single bad command crushes hardware, drains an account, or takes down a live network in milliseconds. Hardware kill-switches trip only after impact; AI guardrails gamble on probabilities. AnankeLabs stops the catastrophe before it executes. KAIROS measures how close a system is to losing control, arrests the command that would cross the line, and seals every verdict into deterministic evidence.
Scope an EvaluationRobotics
Physical AI policies carry real mass and kinetic energy. The robotics adapter sits between the learned policy and low-level execution, evaluates each proposed motion against stopping distance and workspace geometry, and gates the command before the controller commits.
Explore RoboticsSafety at every stage of a robot’s life.
The same deterministic margin calculation governs the robot from its first simulation sweep to its fielded conformity assessment.
- 01
Calibrate
Sim-to-real calibration
Lock safety thresholds from deterministic simulation sweeps before hardware moves.
- 02
Train
Safe-by-design training
Feed structural-margin feedback into the training loop, so policies learn from the same boundary they face at runtime.
- 03
Deploy
Physical firewall
Gate every proposed motion at runtime, before the actuator command is sent.
- 04
Validate
Automated safety validation
Freeze the system, run adversarial scenario populations, and produce audit-grade reports that replay from committed bytes.
AI Safety
Deterministic governance for frontier agents. Language classifiers guess at intent; autonomous agents optimize around them. The AI safety adapter instruments the decision loop itself, measures structural pressure against the alignment floor, and arrests destructive tool calls before they execute outside the model.
Explore AI SafetyEvery action passes three gates.
Every proposed action passes through a layered gate chain. Any gate can reject an action that would cross the safety boundary.
- 01
State Gate
Checks the system's structural health before any action is considered. Below the deployment floor, every action is rejected, however the prompt is phrased.
- 02
Action Gate
Previews each proposed tool call against the reachability field. Calls that stabilise pass; calls that push toward the boundary are rejected.
- 03
Hazard Gate
Detects basin collapse and multi-agent paradoxes. Triggering a hazard gate locks the session immediately and zeroes the retry budget.
Cyber
Signature scanners only catch attacks they have seen before. The cyber adapter treats each defended network zone as a structure under load, reads normalized telemetry in real time, and scores the zone's stability before an intrusion becomes a breach.
Explore CybersecurityThe zones an intrusion has to cross.
Every defended surface reduces to a zone with a finite capacity to absorb pressure. Four archetypes cover the surfaces most intrusions cross.
- 01
Identity Plane
Privilege escalation
Authentication, authorization, and credential stores. Tracks escalation load against the credential boundary before privilege control gives way.
- 02
Edge Device
Initial access
The perimeter where external traffic meets infrastructure. Scores exploitation load on exposed services before the perimeter is breached.
- 03
Internal Segment
Lateral movement
East-west traffic between interior systems. Detects propagation drift across trust boundaries before lateral movement locks in.
- 04
Data Plane
Bulk egress
High-value records, models, and secret stores. Flags abnormal staging and egress pressure before exfiltration commits.
KAIROS Substrate
The engine under every adapter.
Robotics, AI Safety, and Cyber all run on the same compiled Rust kernel. It evaluates every proposed action against the safety boundary in under a millisecond, and seals each verdict into an auditable log.
Determinism
The artwork at the top of this page runs on KAIROS.
Same seed, same step, same state digest on every run of this build.
- Engine
- kairos-wasm @ 62bcfdd · sha256 a3315013…
- Run
- —
- Tick
- —
- Digest
- —
- Status
- Loading engine
Specification
- Replay
- Replaying the committed log reproduces every verdict exactly.
- Timing
- The verdict is computed in under a millisecond, before the action executes.
- Integration
- Policy- and model-agnostic: sits beside any planner, model or controller.
- Build
- Compiled Rust. Zero runtime dependencies, no unsafe code in the engine core.
- Deployment
- Air-gapped. Runs fully offline.
Where it runs
- Native LibraryLinks directly into C, C++, or Rust control loops for sub-millisecond in-process evaluation with zero network overhead.
- CLI & Observe TUIStandalone binary for live terminal telemetry, transcript replay, and automated pass/fail gates in CI pipelines.
- Python SDK & REST APINative PyO3 bindings for simulation and training sweeps, plus an OpenAPI 3.1 HTTP endpoint for agent harnesses and cloud pipelines.
- WebAssemblyDeterministic browser target compiled from the exact same Rust core, powering the live simulation at the top of this page.
Scope an Evaluation
Scoped evaluations test the KAIROS robotics adapter directly against a team's own embodiment and workspace geometry. An engagement scopes a single control-loop decision point, ingests the robot's kinematic limits alongside a recorded trajectory cohort or MuJoCo/Isaac simulation scene, and returns a digest-anchored containment and operability report.
Design partnerships are also open for teams deploying the AI safety adapter (agent trajectories) and the cybersecurity adapter (redacted zone telemetry), as well as conformity-assessment and regulatory specialists. Select a domain on the right or use the Contact tab to start a technical conversation.
Request received. We'll be in touch.