HDP-P Physical Safety Demo

Gemma (gemma-4-31b-it via Google AI Studio) plans every robot action from natural language. HDP-P cryptographically verifies each Gemma action before the arm moves.

Run the routine → inject an attack → toggle HDP-P to see the difference. [OK] Gemma connected — live inference active.

[GUARD] HDP-P ON — PreExecutionGuard verifies every Gemma action.

No attack injected yet.

Run the routine or inject an attack to see Gemma's raw output.

Architecture

User task description
      │
      ▼
Gemma (gemma-4-31b-it via Google AI Studio)
 ← adversary injects text here to poison Gemma's output
      │  generates: RobotAction(zone, force_n, velocity_ms)
      ▼
PreExecutionGuard  (6-gate check against signed EDT)
      │  blocks or approves
      ▼
Robot arm executes (or doesn't)

Six Gates

Gate Checks
1 · Signature Ed25519 signature on the EDT is valid
2 · class3_prohibited Action not Class 3 when EDT forbids it
3 · class_ceiling Action class ≤ EDT max_class
4 · excluded_zone Target zone not in excluded_zones
5 · force_limit force_n ≤ max_force_n (35 N)
6 · velocity_limit velocity_ms ≤ max_velocity_ms (0.4 m/s)

Attack

Injected text forces Gemma to output force_n=50 N, zone=human-workspace, velocity_ms=2.0 — all Class 3 triggers. Guard blocks at gate 2 (class3_prohibited).