We use cookies to keep the site working, learn how it's used, and improve what we ship next. Accept everything, reject the non-essential, or pick category by category. See our privacy policy.
Technical reference
Models propose. Policy decides. Devices execute.
PIRA keeps the trust boundary explicit: AI model → typed action proposal → deny-by-default policy engine → scoped authorization → trusted Android executor → physical capability → auditable result.
01
AI model
The model provides untrusted reasoning input. It cannot turn its output into shell, ADB, arbitrary URL, or background-surveillance execution.
02
Typed action proposal
Every request names the exact capability and parameters PIRA must evaluate.
03
Deny-by-default policy engine
Policy checks run before execution; unsupported actions are rejected and logged.
04
Scoped authorization
Authorizations are short-lived and tied to the exact action parameters the user approved.
05
Trusted Android executor
The executor rechecks approved operations, uses Keystore-backed encrypted provider credentials, and remains independent from model output.
06
Physical capability
Android performs the permitted camera, speech, or supported device operation—never unrestricted access.
07
Auditable result
Allowed and denied action events, outcomes, and errors form an audit record. An independent emergency stop can block execution.
