REFLEXEDGE
VERIFIED ARM64 RUN
ARM CREATE 2026 ↗

PHYSICAL AI · SENSOR → INFERENCE → ACTION

A brake reflex
you can audit.

ReflexEdge turns a raw 64-beam distance and radial-velocity frame into a learned collision-risk score and a deterministic actuator command. Every speed and safety claim replays from raw evidence on real Arm hardware.

01 / RAW SENSOR→02 / FUSED INT8 NEON→03 / BRAKE
Apple M4arm64NEON ONDOTPROD ON
DETERMINISTIC REPLAYClear path
ACTIONGO0.0% risk
VERIFIED DELTA6.06×

median p95 · raw sensor → action

P95 · FINAL RUN
10738.54 ns4703.13 ns
THROUGHPUT · FINAL RUN
194.95K/s338.82K/s
MODEL BYTES
584 B160 B
ACCURACY
98.24%98.20%
ADDED GROUND-TRUTH FALSE NEGATIVES0

15 full action changes · 3 BRAKE-boundary · 0 missed scalar BRAKE · 3 additional int8 BRAKE

THE EVIDENCE CONTRACT

Optimization is only real
when safety survives it.

01

Freeze the baseline

2,500 unseen test frames. 5 independent alternating-order paired trials plus a 500-pass final run. Same threshold and actuator policy.

02

Change one mechanism

Reference feature encoding and scalar FP32 become LUT + one-pass summaries, vectorized quantization, and an Arm NEON dot-product kernel. A validation-only safety bias favors an extra brake over a missed brake.

03

Retain raw proof

Dataset 40c9395be7cc… · MIT synthetic corpus · hardware identifiers removed.

JUDGE-READY OFFLINE PATH

One command.
No cloud. No secrets.

REPRODUCE THE FULL CLAIMMIT
./scripts/reproduce.sh
  • Regenerates the rights-clean sensor corpus
  • Trains and freezes the FP32 model
  • Builds scalar and Arm NEON engines
  • Runs safety, performance, rights, and negative-control gates

CLAIM BOUNDARY

Measured locally, not generalized globally.CPU time per inference is an energy proxy, not joules. Synthetic frames prove deterministic regression behavior, not field safety certification.