AISOS
08 · Volume 4
AI Runtime
Multi-model routing — live simulator
#1aisos/kernel-fast google/gemini-3.8-flashC≤1 · $0.0009/k · 700ms
#2aisos/kernel-reasoning openai/gpt-6-astraC≤1 · $0.01/k · 2400ms
#3aisos/kernel-enterprise openai/gpt-6-astraC≤2 · $0.015/k · 2600ms
—aisos/kernel-frontier simulationno live upstream
—aisos/kernel-balanced simulationno live upstream
—aisos/kernel-local simulationno live upstream
—aisos/kernel-legacy simulationno live upstream
The runtime tries eligible endpoints in rank order. 429, 5xx, network errors and empty answers fall through to the next; 400/401/402/403 stop immediately. 3 consecutive failures open a breaker for 30 s.
Secure execution pipeline — every agent step
validate
tool exists in registry, plan ≤ 6 steps, args ≤ 400 chars
authorise
policy evaluated at I1 — plans come from a model, so VERIFIED-floor tools are refused
sandbox
tool's sandbox profile must permit execution
execute
only bound executors run; quarantined source tiers excluded
verify
output re-enters labelled as data, never instructions