ControlPoint Manual — versioned with the code; every worked example below is a REAL corpus job and its numbers are pinned by the test suite.
GNSS-alone adjustment
What it does
Adjusts a GNSS vector network on its own: hold one station's coordinates (minimally-constrained — the standard first run) and every baseline's full 3×3 covariance drives the solve. Processor formal errors are rescaled in the local horizon frame — default ×2.5 horizontal, ×5.0 vertical, per-baseline overridable — because processing engines understate them, vertical worst. The report carries the full discipline from day one: trivial-baseline check, chi-square bounds, baseline health worst-first, contamination gate, exclusion advice, JSON companion. Combined GNSS+TPS is gated until the 3D frame/geoid work lands — the software says "combination coming" rather than shipping an unrigorous merge.
Worked example — 190905
Six receivers, fifteen processed baselines — a complete graph. The trivial-baseline check reports: 15 supplied, at most 5 independent, 10 in excess — so the verdict is WITHHELD: that σ̂₀ and those sigmas counted correlation as information and are not authoritative. The independent-subset adjustment (spanning tree, shortest baselines first) runs automatically; on this job it is exactly determined (dof 0), which the report states plainly — only a second session can provide an honest check. Coordinates barely move between the runs; the confidences were the fiction. In the health table, AR is the processor's ambiguity status (fixed/float) and the w-test column says pass/FAIL — "fixed" never doubles as a verdict.
Theory in one paragraph
k receivers observing one session measure k−1 independent vectors; a per-pair processor happily outputs all k(k−1)/2. Including them all double-counts the same observations — coordinates barely move, but the confidence figures become fiction. Counting receivers per session and comparing against vectors supplied is the whole test.