In short
Rebuilt a published radar-and-communications framework that shipped without code, cut its dominant compute cost 64-fold and found three errors in its results.
The problem
Integrated sensing and communication is how operators get radar out of spectrum they already own: one waveform carries data and detects objects at the same time. Autonomous driving and drone tracking need exactly that. The framework we worked from had been published without code, so nobody outside the authors could verify or extend it.
What we built
We wrote our own implementation. Cutting the delay-Doppler grid from 256 × 16 to 16 × 4 reduced the dominant correlation-search cost 64-fold, and every qualitative trend still held across a hundred Monte Carlo trials.
64 times less search, drawn to scale
Original delay-Doppler grid · 256 × 16 = 4,096 cells
Reduced grid · 16 × 4 = 64 cells · 64× fewer
The correlation search scales with the number of cells. Every qualitative trend held across 100 Monte Carlo trials.
Source: COEP TU Cornerstone-II reproduction study, 2026
Three problems surfaced along the way. The bistatic case subtracted a velocity from a frequency, a unit mismatch. Two results were hardcoded as fixed multiples instead of being simulated. We replaced all three with direct simulation.
The implementation has not been released.
You are reading a shorter version of this technical edge.