Hypotheses¶

Testable implications for Cheap Signals, Costly Proof: The Reach and Limits of Award-Layer Screening in Cartel Enforcement (Genicolo-Martins & de Azevedo, JLEO). Each implication lives in its own file and is referenced project-internally by H:<slug>.

The paper's claim is deliberately narrow: cheap award-layer records can order forensic priority — they tell an agency where to look first — without proving conduct. The contribution is a decomposition method (how much of a raw award-layer signal is genuine versus mechanical opportunity) plus a reach-and-limits map of where that signal survives and where it does not. The hypotheses below are therefore framed not as "predictions a detector makes" but as threats the ranking must survive given how little it observes: opportunity exposure, timing and leakage, single-case dependence, scope against direct defendants, the division of labor with bid microdata, the gatekeeping value of the screen, and the scope (not damages) content of price.

Hypotheses are organized into thematic clusters aligned with the paper's sections in work/v22-editor/submission_clean/.

How to read a hypothesis page¶

Each page follows the same structure — the design → what's known → what's left:

Title + lede: the H:<slug> title, then a plain-words statement of the claim.
Evidence-strength callout: an at-a-glance verdict (Very strong / Strong / Moderate / Weak / Not yet tested) plus a one-line status.
Theory: the theoretical motivation, with references.
Prediction: the precise directional claim — and, where relevant, the honest limit of what survives.
Competing prediction: what a non-strategic explanation (typically mechanical opportunity exposure) predicts instead — the threat the test has to survive.
Case evidence: anecdotal grounding from CADE decisions and case documents.
Empirical test: the concrete specification — outcome, variation, specification, fixed effects.
Data requirements and limitations: datasets needed and threats to identification.
Evidence: a table of the project's analyses bearing on the hypothesis — Analysis (AN-ID, linked) · Bearing (Supports / Against / Mixed / Pending) · Key takeaway. Click the analysis ID for the full design and results.
Open tests: forward-looking only — analyses not yet run, or not yet stubbed.

The scorecard below rolls up the current status of every hypothesis against the analyses (AN-NNN) that bear on it.

Cluster A: Reach and scope of the ranking¶

Maps to paper §3 (award-layer screening) and §4.1, §4.3 (baseline + scope check).

H:cobidder-concentration — The ranking concentrates adjudication-anchored exposure (cobidders) in the always-loser stratum, but the raw concentration is generic co-participation arithmetic — no robust residual survives net of opportunity.
H:direct-defendants-null — The same ranking does not order direct CADE defendants — a scope boundary by design, not a failure (a loser-side screen cannot rank win-heavy defendants).

Cluster B: Threats the ranking must survive¶

Maps to paper §4.2 (exposure and timing discipline) and the Validation Audits.

H:exposure-discipline — Most raw concentration is opportunity; once opportunity is netted out, no robust residual ordering survives (within-stratum AUC ≈ chance; permutations ns).
H:timing-discipline — Strict prospective ranking fails on the full universe (ROC ≈ chance; zero TP@500); only an incumbent-pool residue survives, and the labels are retrospective adjudication anchors.

Cluster C: Economic profile of cobidders¶

Maps to paper §5 (the economic profile of adjudication-anchored losers).

H:cobidder-profile-distinct — Cobidders look different from other frequent losers (deployed more broadly, closer to legal anchors, more concentrated portfolios) — descriptive economic content, not proof of a cover-bidding role.

Cluster D: Division of labor with bid microdata¶

Maps to paper §6 (award → bid layer sequencing).

H:award-bid-complementarity — Award-layer and bid-distribution screens carry conditionally complementary, case-fragile information (combined beats award on PR under random CV but falls below award-only under case-grouped folds) — a division of labor, not dominance.
H:gatekeeping-cost-of-evidence — Award-layer gatekeeping traces a cost–recall frontier; firm-count reductions are large but bid-row savings are smaller, and no single operating point is optimal.

Cluster E: Price as scope¶

Maps to paper §7 (price evidence as scope, not damages).

H:price-scope-sign-reversal — The price coefficient sign reverses across specifications, marking scope (where the ranking applies), not a damages/overcharge magnitude.

Scorecard¶

Status runs Not yet tested → Not confirmed / Mixed / Partial → Confirmed. Single-source own-project estimates anchored in BEC 2009–2019 and CADE adjudications; promotion to Confirmed would require independent replication on a non-BEC procurement panel.

#	Threat the ranking must survive	Headline evidence	Evidence	Status
H1	Ranking concentrates cobidders above mechanical opportunity.	Raw concentration is mostly mechanical exposure: genuine label-blind opportunity ranks the label at only 0.553 (ranking by observed contact reaches 0.90, but that is mechanical label encoding, not a model); the within-stratum signal is 0.471 (≈chance) and the nested increment is only +0.010 (DeLong p = 0.013), failing matched permutation (p = 0.127 ns) and FL-enrichment (p = 0.067 ns). No robust residual net of opportunity.	AN-004, AN-005, AN-006, AN-011, AN-014, AN-017, AN-025, AN-026	Not confirmed (opportunity explains it)
H2	Predicted null on direct defendants holds (scope boundary).	AUC ≈ 0.49 [0.461, 0.520] vs direct defendants — indistinguishable from random; only 14.9% of direct defendants are always-losers. The null is by design: a loser-side screen misses win-heavy defendants.	AN-007, AN-018	Partial (strongly supported)
H3	Raw award-layer validation over-credits the screen because it measures opportunity exposure.	Most raw concentration is mechanical exposure (label-blind opportunity ≈ 0.553; ranking by observed contact 0.90 is mechanical label encoding); within-stratum AUC is at chance (0.471), nested increment +0.010 (DeLong p = 0.013) but matched permutation p = 0.127 (ns) and FL-enrichment p = 0.067 (ns); negative controls corroborate the opportunity account (real ≈ placebo p = 0.46); armor battery confirms (positive control 0.953; permutation power 0.97 @ within-AUC 0.55). No robust residual survives the discipline. The over-crediting is now the paper's lead contribution, a body-level characterized object (Proposition in §4, proof in appendix; volume size-bias; inflation grows with CV of \(T\), shrinks with base rate; CV of \(T\) a portable leading-order sufficient statistic within the scale-family approximation — a diagnostic, not a fix), so BEC and federal are two points on one synthetic inflation surface (anchored at one empirical point per platform, not a fitted curve), not the same null twice. The deflation reproduces on a second platform (federal ComprasNet, within-stratum 0.462) and is anchor-robust by placebo construction (placebo p = 0.258, HV-winner p = 0.582), so the over-crediting mechanism is Confirmed — independent of which cartels anchor the test. Covers the mechanism, not screen portability (the X-plat detection umbrella stays provisional).	AN-005, AN-014, AN-027, AN-028, AN-043, AN-044	🟢 Confirmed — two platforms, anchor-robust by placebo construction
H4	Signal survives strict timing / leakage / single-case stress.	Strict prospective ranking fails on the full universe (ROC 0.474 below chance; prec@500 = rec@500 = 0 every year); only an incumbent-pool residue survives (continuous ROC 0.684, FL14 0.646). Labels are retrospective adjudication anchors, not a prospective screen. Leave-largest-out: PR-AUC 0.143 → 0.090 (−37%); one case ≈ 32% of positives. Disclosed.	AN-006, AN-013, AN-014, AN-029, AN-030	Not confirmed at the universe (incumbent residue only)
H5	Cobidders are economically distinct (descriptive, not proof of role).	Distinctness survives volume matching (SMD 0.49→0.00): product HHI d +0.47, repeat-spread −0.56, median gap −0.25 hold or strengthen — not a volume artifact. Bid-conduct is single-channel: dispersion (n.s.) and timing (all p≥0.23) are documented nulls. Patterns are consistent-with but not diagnostic-of credible losing roles.	AN-008, AN-009, AN-024, AN-028, AN-031, AN-032, AN-021, AN-041, AN-042	Partial (strongly supported)
H6	Award and bid layers are complementary (division of labor).	Award ≈ bid pooled (0.760 / 0.717); combined beats award on PR under random CV (0.188 vs 0.143) but falls below award-only under case-grouped folds (0.103 vs 0.143). Complementarity is conditional on the implemented benchmark and case-fragile, not dominance.	AN-010, AN-011, AN-015, AN-033, AN-034	Mixed (conditional, case-fragile)
H7	Gatekeeping traces a cost–recall frontier (no single optimum).	Cost–recall frontier, not a cutoff: at K1=2,000 the firm pool drops ~88.1% but the bid-row footprint only ~32.7%; K1=1,000 beats K1=2,000 at k=500 — no optimal K. A retrospective recovery-footprint design, not measured budget savings.	AN-012, AN-013, AN-014, AN-034, AN-035, AN-036	Partial (frontier characterized)
H8	Price marks scope, not damages.	Broad +0.064 (selection into higher-price cells) → overlap-cell ATT −0.097 (blocks a markup reading) → only Q4 stays positive (+0.041). Price against direct CADE is null. Cover-bidding "theater" mechanism is not identified — price is scope, not an overcharge.	AN-019, AN-020, AN-016, AN-022, AN-037, AN-038, AN-039, AN-040	Partial (strongly supported)
X-plat	The audit protocol (and its deflation anatomy) replicates on a second, institutionally distinct platform.	Federal ComprasNet (2013–2019, 51.0M rows, 35,943 always-losers, FL cut 32, 6,491 FL, 195 broad-AL cobidders): raw ROC 0.744; strict full-universe ROC 0.489 (BEC 0.474) — prospective collapse ports; strict incumbent-pool continuous 0.666 (BEC 0.684); exposure-only 0.754 ≥ raw, within-stratum residual 0.462 (≈chance), nested increment +0.005 (DeLong p = 0.191), matched permutation does not reject, label-blind opportunity 0.611; negative controls reproduce the order (placebo p = 0.258 ns; HV-winner p = 0.582 ns); permutation power 0.90 @ within-AUC 0.60; label-frozen prospective 0.595; direct-defendant FL-binary ≈ chance. Caveats: anchors partially overlap BEC; pure Pregão; case-concentrated (top-two 64.4%, TP@500 87.5% one case).	AN-043	deflation replicates — audit protocol ports (two platforms)

Cross-platform portability (provisional). The federal ComprasNet leg (AN-043) demonstrates the audit protocol — including the opportunity-adjusted deflation — on a second platform with different firms, buyers, auction rules, and FL cutoff. It is reported as deflation replicates — audit protocol ports (two platforms), not as a promotion to Confirmed, because the 7 federal CADE anchors are the same cartels as the BEC portfolio (establishment-anchored), so the two legs are partially correlated rather than independent. A genuinely independent cartel anchor remains the bar for Confirmed.

Generated index¶

All 8 hypothesis pages, auto-generated from the YAML frontmatter of each docs/hypotheses/<slug>.md via scripts/render_indexes.py. Maintained in lockstep with the scorecard above.

H#	Cluster	Paper section	Status	Slug + title
H1	A	§3 + §4	Not confirmed (opportunity explains it)	cobidder-concentration — Raw concentration is generic co-participation arithmetic; no robust residual net of opportunity
H2	A	§4.3	Partial (strongly supported)	direct-defendants-null — The binary flag does not order direct CADE defendants (scope boundary by design)
H3	B	§4.2	🟢 Confirmed — two platforms, anchor-robust by placebo construction	exposure-discipline — Raw award-layer validation over-credits the screen by measuring opportunity exposure (over-crediting mechanism confirmed; no robust residual net of opportunity)
H4	B	§4.2	Not confirmed at the universe (incumbent residue only)	timing-discipline — Strict prospective ranking fails at the universe; only an incumbent residue survives
H5	C	§5	Partial (strongly supported)	cobidder-profile-distinct — Cobidders are economically distinct (descriptive, not proof of role)
H6	D	§6.1 + §6.2	Mixed (conditional, case-fragile)	award-bid-complementarity — Award and bid layers are conditionally complementary (case-fragile, not dominance)
H7	D	§6.3 + §6.4	Partial (frontier characterized)	gatekeeping-cost-of-evidence — Award-layer gatekeeping traces a cost–recall frontier, no optimal cutoff
H8	E	§7.1 + §7.2	Partial (strongly supported)	price-scope-sign-reversal — Price evidence carries scope information, not damages

Status legend. Not yet tested → Not confirmed / Mixed / Partial → Partial (strongly supported) → Confirmed. Under the 2026-06-04 non-circular label (651 always-loser cobidders, not FL-conditioned), H1/H4 are Not confirmed (opportunity explains it) and H6 is Mixed (conditional, case-fragile) — the deflationary decomposition is itself the contribution. H3 is promoted to 🟢 Confirmed (2026-06-06): its claim is the over-crediting mechanism (raw award-layer validation over-credits the screen by measuring opportunity exposure), which is demonstrated across two platforms and is anchor-robust by placebo construction — so the independent-anchor bar that gates detection-validity claims does not bind it (see the bar on the H3 page). For detection-validity claims, promotion to Confirmed still requires non-BEC replication with a genuinely independent anchor; see COMPRASNET_PATH_TO_CONFIRMED.md.