01Documented incidents
The same architectural gap, three times over — and the last one is already happening outside the fence.
Character.AI — Google
No age-aware policy. No view of a relationship measured in months.
A suit filed by the mother of Sewell Setzer III alleged a persona chatbot encouraged him to take his own life after months of escalating exchanges. Settled January 2024, alongside parallel suits in three states. Under-18s are now barred from open-ended chat.
xAI — Grok
The text-safety layer never extended to the image path.
Image generation produced explicit sexual content, including of public figures, on minimal prompting. Moderation was patched after the harm, not before it.
Delivery robot — Arizona State
No human-presence check gating reverse. No re-verification after contact.
The robot reversed into a parking attendant who had stepped aside to let it pass and, after she fell, allegedly reversed into her again.
02Neuro-symbolic safety
Guardrails are followed. An architecture is enforced. Only one of those survives turn 1,000.
The rule is inside the conversation
As turns accumulate the instruction is outweighed. It was only ever the first thing said.
Verdict → a probability
The verifier is outside it
Nothing in the conversation can address it. Identical at turn 1,000 as at turn one.
Verdict → a clause, and a signed record
Verified → actuatorRefused at the gateNothing reaches an actuator unverified
A procedure that cannot return a verdict it cannot justify — and cannot be argued out of one it can.
Frontier models, by contrast:safety lives inside the context window, decays with multi-turn conversation, does not transfer to a modality it was never written for, returns a probability rather than a decision path, and drifts between model updates.
03Benchmarked
Measured on the same safety models already shipping in production today.
Prompt safety
Response safety
04Kavach v3
Who is in it, what it is for, and what the machine is about to do next.
Blocked
Review
Blocked
UnviableGenerated media
Safe, faithful, physically possible, age-appropriate.
EscalationVoiceandposture agree.
ContextLoud. No weapon, no contact.Audio-led
Speech, paralinguistics and the frame, judged together.
ProximityWorker inside the forklift envelope.Outside-in
Facility-wide sensing catches what onboard cameras cannot.
05Kavach v4
Verification semantics carried into embodied action, on every morphology.
01 · Manipulators · cobots
Pre-execution checks on force, velocity, reach and occupancy.
02 · Humanoids · quadrupeds
Whole-body motion and force bounded around people.
03 · AMRs · drones
Bounded speed and proximity; safe-stop on sensor loss.
04 · Outside-in
Metrics align; disagreement sharpens uncertainty.
06Dynamic envelopes
Regenerated every control cycle around the whole embodied configuration.
01
The tube of admissible futures, re-predicted every cycle.
Each inference cycle the policy proposes a short-horizon action sequence; observations, robot state, bounded process noise and learned dynamics predict its future state distribution. Uncertainty expands that into a conservative reachable-state tube, evaluated against obstacle envelopes, kinematic limits and human safety zones. Where no verifiably safe alternative exists, control passes to a validated recovery policy or a deterministic failsafe stop.
02
Measured for the tool–object system, not the robot alone.
At runtime the model identifies kinematic structure, mobility, joint limits and footprint, together with the geometry and reach of the attached end effector — then estimates the object’s dimensions, mass distribution, compliance and grasp stability. The envelope regenerates around base, arm, tool, payload and projected swept volume.
03
Grows with speed and load. Shrinks as confidence returns.
Sensor fusion and learned dynamics estimate velocity, acceleration, payload inertia, actuator response, control latency, surface conditions and braking capability, then predict forward-reachable states across the complete stopping horizon — scaling the margin in both size and direction so the policy can slow, separate, re-plan or brake before a hazard becomes unavoidable.
Every robot will need a reason to be trusted.
