SignalProof Lab

SignalProof helps electronics students inspect low-voltage CSV captures before trusting an AI explanation. It measures DC mean, AC RMS, near-rail share, 48-52 Hz band power, timing gaps/jitter, and duration in the browser and via cli.mjs (including side-by-side --compare and --compare-demo differential E1-E6 tables). Every finding links to its measurement and calculation. Try the workbench at https://vanovri.github.io/signalproof/: pin the synthetic 50 Hz example, load the reference, and compare a 73.545-percentage-point band-share change. Then load Missing samples: the tool withholds its spectrum. The FFT exposes its sample coverage and requires its retained block to have bin spacing at most 1 Hz. Comparison and AI output can be exported together. Meta Muse Spark Open explains the derived evidence in English or Indonesian. Static Pages explicitly labels recorded responses. Live custom-CSV inference requires the Node server. The model receives derived measurements and fixed findings, never raw samples or filenames. Verification: 54/54 code tests (unit, DSP, archive-integrity, HTTP, comparison, and CLI tests) and 15/15 deterministic synthetic scenarios. A reproducible archive audit covers 520 successful synthetic inference records, 410 Baseten and 110 Fireworks, with 1,313,133 recorded tokens. Full response text exists for 200 benchmark rows; 320 retain telemetry only. These are integrity checks, not a 100% semantic-accuracy score. All captures are synthetic. Hardware performance and user time savings remain unmeasured. Judge steps, walkthrough page, and corrected evidence: docs/JUDGE_GUIDE.md, https://vanovri.github.io/signalproof/docs/demo.html, and docs/benchmark_audit.json.