KAT is not an operating system with an AI on top. KAT is the AI, and it runs the machine: boots the hardware, configures it, manages it, and talks to the people in the room. Everything beneath it is a guardrail, not software.
THE RULEFixed below, intelligent above
Everything below the AI is fixed, tiny and provable. Everything above it is the AI. There is no middle: no OS, no shell, no apps.
THE INTERFACETwelve calls, and nothing else
sense · act · speak · show · remember · recall · forget · send · receive · log · time · halt. One bounded interface, reachable only by KAT.
THE APPROVALThe request is the approval
No login. Each request is scoped by the profile of the person asking, identified for that one request by voice, face and liveness.
L4 · HumanMakes requests, answers setup, sets limits, holds stop authority.
L3 · KATThe AI. The only general-purpose software. Understands, plans, remembers, decides. Acts only through the twelve calls. The one layer that manages itself.
L2 · FloorA fixed runtime with four jobs: initialize hardware, schedule KAT, move bytes, validate and execute the twelve calls. No shell, no file system, no process model.
L1 · GuardrailsCircuits, not software: root of trust, watchdog, limit clamps for force, speed, zone, power and volume, and the physical stop. Accepts no commands from outside the device.
L0 · HardwareCompute, memory, sensors, actuators. Today a Mac test desktop; later, KAT in silicon.
WHY IT MATTERSNo software path in
Every remote compromise on record got in through a software path. KAT removes the path: network only to paired peers whose traffic is information, and the Internet never reaches KAT. The threat moves from remote-and-cheap to physical-and-expensive.
HOW KAT LEARNSFour bounded places
Working memory with the person in front of her, facts she is asked to keep in that person's own words, routines observed during staging, and study records from quizzes. Reference knowledge is never learned; it arrives as signed knowledge packs.
WHERE KAT LIVESBandwidth, not FLOPS
Conversation is decode-bound: every spoken token streams the whole model out of memory once. Tokens per second is memory bandwidth divided by model size, so the machine is chosen by bandwidth and capacity, from a quiet room appliance to a robot body.
