ControlPoint Manual — versioned with the code; every worked example below is a REAL corpus job and its numbers are pinned by the test suite.
The levelling QC report
What it does
The levelling report carries the full QC standard: input hashes, health by fieldbook line, benchmark consistency (a minimally-constrained run holding ONE benchmark, misfit of the rest), loop misclosures from the observation graph, fore/back reciprocal sections, a collimation diagnostic with three honest verdicts, paired-plate candidates, staff-spike detection, and the full residual appendix with audit keys and both a-priori and a-posteriori w.
Worked example — 250115-2HPX
Error factor σ̂₀ = 0.943 on 97 observations. Five benchmarks held — but the consistency check shows only three are ever observed: B01RTZ01 and C01RTZ01 constrain nothing and are flagged UNTESTABLE. The two testable ones misfit +1.49 and −0.10 mm: noise, not datum strain. Collimation: every line TESTED CLEAN, e.g. M_17042023 slope +0.24 ± 2.70 mm/km with a minimum detectable of 5.4 mm/km over its 50 m imbalance range.
Theory in one paragraph
"OK" is only meaningful with power: a perfectly balanced line CANNOT reveal collimation (balancing is how you cancel it in the field), so the verdict set distinguishes tested-clean from untestable-by-design. The same honesty drives the benchmark check — a held point that no observation touches is not confirmed, it is merely repeated.