Reviewer quickstart

Everything below runs offline on commodity hardware. No accounts, no data upload, no external datasets.

1. Install and test (~2 min)

npm ci
npm test          # unit + integration suite (deterministic, pure cores)

2. Reproduce the evaluation (~30 s)

npm run repro

This runs the real analysis cores over deterministic synthetic fixtures with analytic ground truth and writes:

• benchmarks/out/metrics.md — the evaluation table (also in the paper)
• benchmarks/out/metrics.json — the raw numbers
• benchmarks/out/registration_bias.{png,pdf} — vertical-change preservation
• benchmarks/out/calibration.{png,pdf} — uncertainty-band coverage

The figure step needs Python + matplotlib (pip install matplotlib); the metrics table is written even without it.

What the metrics show:

• M1 — a full-3D rigid registration absorbs a true uniform vertical change into its z-shift (detected-change error grows with the change), while the horizontal-only constraint preserves it (≈ 0 error). This is the change-detection design choice, measured.
• M2 — planar alignment recovers a known horizontal misregistration.
• M3 — the reported stockpile ±1σ band is calibrated: empirical coverage sits near the nominal 0.68 over hundreds of noise realisations.
• M4 — the integrity-report digest is deterministic and tamper-evident.

3. Run the application (~1 min)

npm run build && npm run preview
# open the printed URL, drag in a LAS/LAZ/PLY/E57 scan (or pick a sample),
# place a measurement, export the "Integrity report (JSON)", then run
# the command palette action "Verify integrity report…" on that file.

Everything happens on your machine; no data leaves the browser.