Validation
What was checked, what held, and what the available ground truth can and cannot establish — reported straight.
Hand verification of shipped candidates
Every headline candidate pattern was hand-verified against the cited, effective-dated schedule PDF before publication. Examples: a paid rate of $16.50 screened against the Rockland Electrician rate of $51.50; $42.55 against the Suffolk Carpenter (Building) rate of $53.56; the largest-exposure contractor’s floors (Laborer Heavy&Highway Group A $42.93 and Operating Engineer Heavy&Highway Class C $53.82) confirmed in the 2025–2026 schedule’s July 2025 column. Two Tier-2 apprentice downgrades were also hand-verified in the raw PDFs — one an exact dollar-term match, one a published 60% term applied to the classification’s own journeyworker rate.
False-positive controls
The engine was checked against lines it must not flag: a Laborer (Demolition) line paid exactly the scheduled rate of $32.33 is not flagged, and a line paid $22.87 — the scheduled first-term apprentice rate — is not flagged. Effective-dating discipline was proven on a rate that changed mid-window ($61.59 → $62.59): work before the change date is screened against the earlier rate.
The debarment back-test, reported honestly
The natural ground truth is the NYSDOL Bureau of Public Work debarment list — 76 companies barred for willful prevailing-wage violations. Matching them against the 3,503 contractor accounts active in the 2026 payroll data yields essentially no overlap, and that is structural, not a screen failure: debarment bars a firm from public work, so debarred firms exit the certified-payroll population. The debarment list cannot validate an active-payroll screen.
An earlier, looser name-matcher produced a flattering “56% recall” figure. Inspection showed it was built on generic-word false matches (for example, matching two unrelated firms because both names contain “Contracting Group”). That number was discarded and the matcher rebuilt with distinctive-token requirements. We report the corrected result only.
Exactly one debarred firm genuinely remains in the 2026 payroll data: EGL Drainage Solutions & Repairs, debarred effective July 17, 2025, with a Nassau County fiscal officer. Its 2026 wage lines were initially invisible to the screen because the project ZIP was a PO-box-type code absent from the Census file; after the county-attribution fallback was built, the engine independently flagged EGL — 33 findings, 29 high-confidence, roughly $13.6K indicative exposure, in Nassau County, with both key floors hand-verified in the schedule PDF. All 33 findings also survived the Tier-2 apprentice and alternate-classification screens.
That is a recall of 1 of 1 on the screenable adjudicated population — a single genuine corroboration that the screen points at the kind of contractor that ends up adjudicated, not a precision/recall study. Two important cautions:
- The sample size is one, and the name match is “probable” (“SOLUTION” vs. “Solutions” — clearly the same firm, but noted for completeness).
- EGL’s 2026 wage lines postdate its debarment start. Debarment bars new public-work awards; work on pre-existing contracts may lawfully continue. The 2026 lines are therefore not themselves debarment violations and are not characterized as such — the debarment is cited only as an adjudicated fact about the firm’s history, corroborating that the screen’s flags landed on a firm with an adjudicated violation record.
What a real precision/recall study would take
A credible precision number requires case-level settlement or determination data — firm, underpaid classification, and work period — that overlaps the 2026 window. The public dashboards summarize such cases but do not expose them in machine-readable, case-level form. Until that data is in hand, the honest claim is the one made throughout this site: the screen produces evidence-cited candidates for investigation, validated by hand-verification, false-positive controls, and one genuine adjudicated corroboration.