Executive summary
A real radar DSP hot path was ported from Python/NumPy to embedded C with Claude Code. Claude's initial plan was sound: reuse the existing C FFT, validate against NumPy, then optimize. LOCI added an execution signal from the compiled executable and revealed an arithmetic difference hidden by the source: fused multiply-add behavior. Claude used that signal to change its plan before relying on target-hardware discovery.
The co-reasoning flow
- Developer intent. Port the radar Doppler hot path to embedded C.
- Claude plans. Reuse the C FFT and validate against NumPy.
- LOCI signal. Compiled execution shows fused multiply-add behavior.
- Claude reasons. The 1e-9 mismatch is execution behavior, not just tolerance.
- Plan adjusts. Port the NumPy FFT behavior operation by operation, and control fusion.
- Code. Implement the execution-aware plan.
Measured result
| FFT before | FFT after | Accuracy |
|---|---|---|
| 71.1 µs | 43.4 µs | Bit-exact with the Python reference |
Why not just give Claude the workload or binary?
| Binary or workload as context | LOCI's execution signal |
|---|---|
| Claude can inspect artifacts and infer intent from what is visible. | LOCI provides a purpose-built signal about compiled execution behavior and its implications. |
| The agent still has to infer which low-level details materially change behavior. | The signal surfaces execution differences that can change the plan — here, arithmetic fusion hidden behind equivalent-looking source. |
| Typical loop: Plan → Code → Run → Discover → Fix | Execution-aware loop: Plan → LOCI → Reason → Adjust → Code |
Claude reasons about what to build. LOCI reasons about how it will behave.
LOCI is pre-run execution prediction for AI coding agents. No target-hardware run was needed to discover the execution difference in this use case.