Cheap Signals, Costly Proof¶
The Reach and Limits of Award-Layer Screening in Cartel Enforcement
Insper Institute of Education and Research, Sao Paulo, Brazil
Download paper (PDF) Online appendix
JLEO version v25 — major revision: the over-crediting bias promoted to the lead contribution and into the main body; enforcer stopping rule stated modestly; three-document package (paper + online appendix + Online Supplement), two platforms (BEC + ComprasNet), June 2026.
CV(T) Validating a volume-loaded award screen against a contact-defined cartel label over-credits it — and that inflation is an estimable size-bias, governed by the dispersion of participation volume and summarized by a one-number diagnostic — the coefficient of variation of participation counts — computable before any bid file opens. (The gap equals the Gini coefficient of participation volume, with an exact rare-target closed form; the base rate admits no monotone comparative static; magnitude is a synthetic surface anchored at one empirical point per platform, not a fitted curve.) Companion organizational result: the award-to-bid decision is a budget-dependent burden–recall envelope, the image of an ordinary cost–benefit tangency — so "there is no fixed cutoff to defend" is a consequence, not a grand result.
Abstract¶
Cartel enforcement must allocate costly proof-producing effort before legal proof exists. This paper asks how agencies should govern cheap administrative screens before opening bid-level records. Using São Paulo's BEC procurement platform and CADE adjudications, we audit a minimal award-layer construct—persistent zero-win participation—against a reproducible adjudication-anchored cobidder label that never uses the screen itself. The audit identifies an over-crediting problem: when both screen and label load on participation volume, raw metrics overstate evidentiary value—in BEC, an award-layer AUC of 0.761 falls to chance once opportunity is held fixed. Raw reach is exposure, not conduct, and strict prospective ranking fails outside incumbent pools. We then organize follow-up as a burden–recall envelope: its featured point opens 6% of firms but 53% of bid rows, so firm-count savings overstate forensic burden. A second-platform stress test shows that the audit logic travels; the score does not. Liability remains in the richer record.
JEL Classification: D44 D73 H57 K21 K42 L41
Keywords: cartel screening award-layer data enforcement triage incomplete observability bid rigging
Key Findings¶
The paper leads with the over-crediting bias — an estimable analytical object the screening literature has not written down — followed by the organizational cost–recall result and the reproducible non-circular audit protocol that produces both.
Lead contribution — the over-crediting bias as an estimable object (main body, §4)
Validating an administrative screen against adjudicated cases without adjusting for procurement opportunity does not merely risk inflation — it inflates by an amount we characterize, now stated as a Proposition in the main body. Because a contact-defined label and a participation-based score both load on the same volume, the positive class is the size-biased volume distribution, so the raw area is a size-bias gap that is strictly positive for any non-degenerate participation distribution (it equals the Gini coefficient of participation volume) and grows with its dispersion — with an exact closed form in the rare-target Gamma benchmark. The base rate admits no monotone comparative static (exact counterexamples; the calibrated surface is flat over the empirically relevant range). The practitioner deliverable is a one-number diagnostic — the coefficient of variation of participation counts — computable from award data before any bid file is opened. The magnitude is a synthetic surface anchored at one empirical point per platform, not a fitted curve: only the signs and the leading-order statistic are claimed.
Companion result — the award-to-bid decision as a cost–recall frontier (stated modestly)
The award-to-bid recovery decision admits an ordinary cost–benefit (marginal-benefit = marginal-cost) tangency: an agency descends a cheap award ranking until the marginal adjudicable case recovered per forensic dollar meets its cost–value ratio. The cost–recall frontier we report is the empirical image of that textbook optimum. So "there is no fixed cutoff to defend" is a consequence, not a grand result — the optimum is budget-dependent, so the object an enforcement designer reports is the locus of budget-dependent optima, the frontier itself. Each non-dominated operating point (e.g. the envelope point \(K_1 = 1{,}000\), where the firm-count footprint falls ≈ 94% but the bid-row footprint only ≈ 47%) is the tangency for an agency at one implied cost–value ratio; dominated policies — including \(K_1 \ge 2{,}000\) at the \(k = 500\) queue — are optima for no cost–value ratio and are marked as such.
Companion result — a reproducible, non-circular audit protocol
The validation label is built end-to-end from current scripts: 651 unique always-loser cobidders — zero-win firms sharing at least one tender-item with a BEC-active adjudicated CADE defendant (defendants excluded). The frequent-loser flag is never used to construct the label: 341 positives are frequent losers, 310 are not. The protocol sequences label construction, opportunity-cell adjustment, rolling-origin timing, leave-one-case-out tests, a bid-layer benchmark, and cost–recall accounting — components individually familiar, but sequenced for the award-to-bid decision in a way that is, to our knowledge, new.
What the framework correctly catches — the deflation, stated precisely
The raw award-layer score concentrates cobidders (ROC-AUC 0.761, lift 5.6× at the top 500). The framework then catches that this concentration is mostly opportunity, not conduct: genuine label-blind opportunity structure already ranks the label at 0.553, and within comparable opportunity sets the residual ordering is ≈ chance (0.47–0.51), with matched-stratum permutation and enrichment tests unable to reject the opportunity-only null. This deflation is exactly what the over-crediting characterization predicts — not a screen that "fails", but a framework that correctly prices the inflation. The audit is armored: a positive-control score recovers AUC 0.953 in the very same strata (the within-stratum test is not dead by construction), the permutation test rejects with probability 0.97 at true within-AUC 0.55 (non-rejections genuinely bound any residual), and a fully label-frozen timing design ranks new 2017–2019 defendant contact at 0.713 among 13,051 incumbents — generic contact forecasting, not a cartel-specific signal.
Reach and limits — the estimated ranking is not a portable cartel score
Strict prospective ranking fails outside incumbent pools (full-universe ROC 0.47; zero true positives in the top 500 in every rolling-origin year; 23.5% of positives are unrankable entrants). One adjudicated case (rail/metro) supplies 32% of positives and 45% of top-500 true positives — dropping it cuts PR-AUC by 37%. The estimated ranking is therefore not a portable cartel score — that is the object the evidence denies. What carries to other settings is the cost–recall stopping rule, the over-crediting characterization, and the protocol — not a deployable score. Liability remains in the richer evidentiary record.
The two platforms land where the over-crediting characterization predicts
The identical audit battery is re-run on a second, institutionally distinct procurement platform — federal ComprasNet (2013–2019, pure Pregão) — against the same family of CADE cartel anchors, integrated into the manuscript as §5 (The Audit on a Second Platform) and Appendix G. The two platforms are not the same null twice — they are two anchor points that fall where the over-crediting characterization predicts (the synthetic size-bias surface is anchored at one empirical point per platform, not a fitted curve): the federal system's rarer target is exactly why its opportunity-only model edges past the raw score there (exposure-only ROC 0.754 ≥ raw 0.744), as the over-crediting characterization predicts. The over-crediting characterization replicates at full power; the within-stratum residual collapses to chance (0.462, holding subject to a stated federal power bound), the matched permutation does not reject, and the negative controls reproduce the order. The construct ports and deflates; it does not break. This is a provisional robustness leg, not a promotion to "Confirmed": the federal CADE anchors partially overlap the São Paulo portfolio (same cartels, establishment-anchored), and the federal positive base is more case-concentrated (top-two cases ≈ 64% of positives). See AN-043.