Skip to content

Changelog

v25.4 — Length pass — June 7, 2026 (Current)

Headline: a length/compression pass with no change to identification, data, or numerical results. Two sections whose payoff was repetition were tightened: the second-platform (federal ComprasNet) section, which re-ran the full audit battery, had its stage-by-stage re-narration cut (detail kept in the appendix); and the cobidder-profile section folded its descriptive-texture recitation into a single clause pointing to the appendix. Every claim, number, macro, and citation is preserved; the comparative table and the "audit travels, the score does not" result stay in the body. Main text 45 → 44 pp (double-spaced 61 → 59 pp).

v25.3 — Editorial craft pass — June 7, 2026 (superseded by v25.4)

Headline: an editorial craft pass with no change to identification, data, or numerical results. The abstract now states the deflation magnitude explicitly, the introduction is concretized and tightened, the cost-recall table is de-bloated, and the conclusion is tightened.

  • Abstract now states the deflation magnitude — an award-layer AUC of 0.76 falls to chance once opportunity is held fixed — replacing the more abstract "raw screening metrics overstate evidentiary value" phrasing. Quoted verbatim on the landing and paper pages.
  • Introduction concretized and tightened.
  • Cost-recall table de-bloated to a single \(k = 500\) panel; the \(k = 1000\) panel and the full operating-point grid moved to the Online Supplement.
  • Conclusion tightened.
  • Manuscript 45 pp (was 46); online appendix 28 pp; JLEO double-spaced 61 pp.
  • econ-write craft score ≈ 91/100.

v25.2 — Author-final abstract; Box 1 removed — June 6, 2026 (superseded by v25.3)

Headline: the author's final abstract (148 words) replaces the v25 text, and Box 1 is removed from the body. No numerical change.

  • New abstract opens with the allocation problem ("Cartel enforcement must allocate costly proof-producing effort before legal proof exists") and closes on the transfer result ("the audit logic travels; the score does not"). Every claim verified against the manuscript (12% of firms / 67% of bid rows at the §6 operating point; transfer phrasing is verbatim from the comparative section).
  • Box 1 ("From Cheap Suspicion to Bid-Layer Follow-Up") removed from the body; its audit-component content remains covered by §4 and the conclusion. Paper ≈ 46 pp.

v25.1 — Final editorial polish — June 6, 2026 (superseded by v25.2)

Headline: a light final editorial pass on top of v25 with no substantive or numerical change — the over-crediting lead contribution, the cost–recall organizational object, the abstract, and every claim are unchanged from v25.

  • Submitted appendix compressed to 28 pp (from 30), with the residual duplication moved into the Online Supplement; paper ≈ 48 pp and Online Supplement ≈ 23 pp are unchanged.
  • Box 1 retitled "From Cheap Suspicion to Bid-Layer Follow-Up."
  • Three hostile-panel rounds converged — ending 2 accept / 3 minor; pre-submission status READY_AFTER_MINOR.

v25 — Major revision: over-crediting promoted to the lead contribution — June 6, 2026 (superseded by v25.1)

Headline: a major revision for storytelling, compression, and humanization. The over-crediting bias — that validating a volume-loaded screen against a contact-defined cartel label inflates it by an estimable size-bias — is promoted to the lead contribution and moved into the main body (§4) as a Proposition. The enforcer's cost–recall / stopping rule is restated modestly, as an ordinary textbook cost–benefit tangency, not a grand result. The package is compressed into three documents with a new Online Supplement. A hostile pre-submission panel ran three rounds and ended 2 accept / 3 minor, zero rejects — both prior rejects were converted.

Concrete changes in this version:

  • Over-crediting bias is now the lead analytical object, in the main body (§4). The Proposition Signs of the over-crediting bias moves from the appendix into the 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 increasing in the dispersion (coefficient of variation) of participation counts and decreasing in the adjudicated base rate (signs analytic). The practitioner deliverable is a leading-order sufficient statistic — the CV of participation counts (CV(T)) — computable from award data before any bid file opens — that bounds how much of a raw screening number is mechanical.
  • Synthetic-surface honesty. The over-crediting magnitude is a synthetic size-bias surface anchored at one empirical point per platform — NOT a fitted or estimated curve. Only the signs and the leading-order statistic (CV of T) are claimed; the phrase "inflation curve" is retired wherever it could read as a fitted object. The two platforms (BEC + ComprasNet) are two anchor points that fall where the characterization predicts, not points on an estimated curve.
  • Stopping rule restated modestly. The award-to-bid decision is presented as a standard cost–benefit (marginal-benefit = marginal-cost) tangency, with the cost–recall frontier as its empirical image. "There is no fixed cutoff to defend" is stated as a consequence (the optimum is budget-dependent), not a headline result. No overselling.
  • Abstract reordered (~150 words) to lead with the over-crediting object, then the organizational cost–recall frontier, then reach-and-limits; quoted verbatim on the landing and paper pages.
  • §4 compressed; the full permutation/rolling-origin grids, threshold sweeps, negative-control grids, bid-feature dictionary, full price-regression grid, complete ComprasNet construction detail, and the long proofs of both propositions are lifted into the new Online Supplement.
  • New three-document package: paper ≈ 48 pp + online appendix ≈ 30 pp + new Online Supplement ≈ 23 pp.
  • Sánchez Graells (2019, screening/transparency objection to publishing a procurement red-flag) cited in the price-scope section.
  • Site: landing key-result and Key-Findings reordered to lead with the over-crediting object (stopping rule second and modest, protocol third); paper.md structure, formal-results, and contribution list updated to place the over-crediting Proposition in §4 and the modest stopping rule in online Appendix B, with the Online Supplement named.

Hostile-panel trajectory: three rounds (review → Action Plan + armor → verification), ending 2 accept / 3 minor, zero rejects; both prior rejects converted.

Honest caveats preserved (front-paged, not hidden):

  • No closed-form inflation magnitude. Only signs (in CV of T and base rate) plus the leading-order CV-of-T sufficient statistic are claimed; the magnitude surface is synthetic, anchored at one empirical point per platform.
  • Partially overlapping legal anchors wherever the federal evidence appears — the federal CADE cases are the same cartels as the BEC portfolio, establishment-anchored; a robustness leg under partially correlated ground truth, not independent out-of-sample proof.
  • Power-bounded federal null — the over-crediting characterization replicates at full power, but the within-stratum deflation holds only subject to a stated federal power bound.
  • Not a cartel detector. The estimated ranking is not a portable cartel score (that is the object the evidence denies); cobidders are adjudication-anchored exposure, not cartel members (651 broad always-loser cobidders on BEC); AUC ≈ 0.49 against direct CADE defendants is the loser-side scope boundary by design; no platform-wide prospective deployment; no damages/overcharge claim; liability remains in the richer evidentiary record.

Manuscript home: the over-crediting Proposition is in §4 (full grids + proof in the Online Supplement); the cost–recall / stopping-rule result is in online Appendix B (long proof in the Online Supplement). Paper ≈ 48 pp + online appendix ≈ 30 pp + Online Supplement ≈ 23 pp; builds 0 errors / 0 undefined.

v24 — Positive-contribution reframe (enforcer stopping rule + over-crediting bias) — June 6, 2026 (superseded by v25)

Headline: the paper is no longer cast as an audit that finds a screen mostly fails. It now leads with two positive analytical objects the screening literature has not written down, plus the reproducible non-circular audit protocol that produces both. This reframe responds to a hostile pre-submission panel that read the prior framing as tilted toward its own deflationary conclusion; the answer was to surface the two objects that were already implicit in the computation and state them as formal propositions, not caveats.

Object 1 — the enforcer's evidence-acquisition problem, solved (online Appendix B). The award-to-bid recovery decision is formalized as an optimal-stopping problem: an agency with per-record forensic cost \(c\) and per-case recovery payoff \(V\) descends a cheap award ranking to the unique tangency where the marginal adjudicable case recovered per forensic dollar meets the cost–value ratio, \(\rho(K^*)/\phi(K^*) = c/V\) (new Proposition: optimal stopping depth). The cost–recall frontier reported in the body is the empirical image of this solved stopping rule, not a descriptive scatter. Consequently "there is no optimal cutoff" is now a result, not a confession — the optimum is budget-dependent, so the design object is the locus of budget-dependent optima (the frontier), and each operating point (including \(K_1 = 2{,}000\)) is the tangency for an agency at one implied \(c/V\).

Object 2 — the over-crediting bias as an estimable object (online Appendix C). Raw award-screen AUC inflates because a contact-defined label and a participation-based score both load on participation volume (size-bias). New Proposition: signs of the over-crediting bias proves the raw-minus-opportunity-adjusted gap \(\Delta\) is increasing in the dispersion (coefficient of variation) of participation counts and decreasing in the adjudicated base rate (signs analytic; magnitude by simulation, no closed form). The practitioner deliverable is a sufficient statistic — the CV of participation counts — computable before any bid file is opened, that bounds how much of a raw screening number is mechanical. The two platforms (BEC + ComprasNet) are reframed as two points on this one inflation curve, not "the same null twice": the federal system's rarer target is exactly why its opportunity-only model edges past the raw score there, as the characterization predicts.

The deflation, reframed. Raw concentration (ROC 0.761; mostly opportunity, residual ≈ chance) is now presented as what the framework correctly catches, not as "the screen fails." The within-opportunity null is the framework correctly pricing the over-crediting that Object 2 predicts.

Other changes in this version:

  • Abstract reframed to the canonical v24 abstract (~125 words): opens on the enforcer's optimal-stopping problem and the budget-dependent cost–recall frontier, then the over-crediting size-bias and its CV-of-\(T\) sufficient statistic.
  • Citations added: Seibel & Škoda, Collusion by Exclusion in Public Procurement (AEJ:Micro, forthcoming) and Kawai–Nakabayashi–Ortner–Chassang (KNOC) 2023 bid-rotation/RD detection.
  • Page counts updated to ≈ 53 pp main + ≈ 46 pp online appendix.
  • Site: landing key-result and Key-Findings rewritten to lead with the two objects + protocol; paper.md manuscript-structure and formal-results sections updated to name Appendix B (stopping rule) and Appendix C (over-crediting bias) as the analytical core.

Honest caveats preserved (front-paged, not hidden):

  • Partially overlapping legal anchors wherever the federal evidence appears — the 7 federal CADE cases are the same cartels as the BEC portfolio, establishment-anchored; the federal leg is a robustness leg under partially correlated ground truth, not independent out-of-sample proof.
  • Power-bounded federal null — the matched-strata permutation is correctly sized but the federal sample bounds the residual only to a coarser tolerance than BEC.
  • Not a cartel detector. The estimated ranking is not a portable cartel score (that is the object the evidence denies); cobidders are adjudication-anchored exposure, not cartel members; AUC ≈ 0.49 against direct CADE defendants is the loser-side scope boundary by design; no platform-wide prospective deployment; liability remains in the richer evidentiary record.

Manuscript home: the two propositions are in online Appendix B (Theory, Exit, Survival — optimal stopping depth) and Appendix C (Opportunity & Timing — signs of the over-crediting bias); the federal column remains bound via the \valFedAud* macro namespace. Paper ≈ 53 pp + appendix ≈ 46 pp; builds 0 errors / 0 undefined.

v23 — The Audit on a Second Platform (ComprasNet) — June 6, 2026

Headline: the award-layer evidence-triage protocol is now demonstrated on two institutionally distinct procurement platforms. Beyond São Paulo's BEC (2009–2019), we re-run the full audit battery on the federal ComprasNet platform (2013–2019, pure Pregão) against the same family of CADE cartel anchors, integrated into the manuscript as a new Section 5, The Audit on a Second Platform, with the full federal battery in Appendix G. The federal leg replicates the deflation anatomy: under opportunity adjustment the within-stratum residual ordering (score net of defendant-contact exposure) falls to chance (within-stratum AUC 0.462), the matched-stratum permutation does not reject (nested increment +0.005, p = 0.191), and the negative controls return the surviving ordering to the realm of generic opportunity/volume geometry. The construct ports and deflates; it does not break. The strict-temporal carrier replicates too: the prospective ranking collapses outside the incumbent pool on both platforms (strict full-universe ROC 0.489 federal / 0.474 BEC), while the incumbent-pool continuous score deflates in the same direction and magnitude class (0.666 federal vs 0.684 BEC).

What this adds — and what it does NOT. This is a second-platform demonstration of the audit protocol, not a promotion to Confirmed. The federal CADE anchors partially overlap the BEC portfolio (same cartels, different establishments and tender footprints), so the two legs are correlated, not fully independent. Federal panel: 51.0M participation rows, 92,600 firms, 35,943 always-losers, FL cut re-estimated at 32 (≥; not transported from BEC's 14), 6,491 federal frequent losers. Main target: 3,850 broad-rule cobidders → 195 broad-AL cobidders (mirrors BEC's 651), of which 94 are FL and 101 non-FL, anchored on 7 numbered CADE cases and 25 platform-active direct-defendant establishments.

The deflation battery (federal, read from the named CSVs):

  • Opportunity adjustment. Raw award-layer ROC 0.744; exposure-only model 0.754 (≥ raw); within-stratum residual ordering 0.462 (≈ chance); nested increment +0.005 (DeLong p = 0.191); matched-stratum permutation does not reject. Genuine label-blind opportunity structure (cell rate from non-cobidder rows only) ranks the label at 0.611 — the raw concentration is mostly mechanical co-participation exposure.
  • Negative controls. Placebo anchors reproduce the order (AUC 0.728, p = 0.258 ns); non-CADE high-volume-winner anchors exceed it (AUC 0.746, p = 0.582 ns). The surviving ordering is generic opportunity/volume geometry, not cartel-specific.
  • Power. The matched-strata permutation is correctly sized — rejection probability 0.35 at true within-AUC 0.55 and 0.90 at 0.60 — so the non-rejection genuinely bounds any residual at the federal sample size.
  • Label-frozen timing. With pool, score, and label ingredients frozen on 2013–2016, the frozen score ranks new 2017–2019 defendant contact at 0.595 among 24,888 incumbents (N = 98) — generic contact forecasting, not a cartel-specific signal.

Honest caveats (front-paged, not hidden):

  • Partially overlapping legal anchors. The 7 federal cases are the same cartels as the BEC portfolio, establishment-anchored at the federal level. This is a robustness leg under partially correlated ground truth, not a clean out-of-sample replication on an unrelated cartel set.
  • Shorter window (2013–2019). Four training years federally vs eight on BEC. The strict full-universe prospective collapse is robust (ROC 0.489), and the frozen-pool prospective row (0.595, N = 98) is reported as such.
  • Pure Pregão. Convite is extinct federally, so the federal leg cannot reproduce the BEC modality stratification. No federal bid microdata exists, so no bid-layer/Imhof forensic benchmark is constructible from the public federal data.
  • Case concentration is worse than on BEC. The federal positive base is more cartel-concentrated: top case 35.4% of positives, top-two 64.4%, and the realized top-500 detections are 87.5% one case. The clustered-randomization ordering test is robust (leaving out the largest case keeps the ordering distinguishable from random, RI p < 0.001), but the operational precision/recall-at-k ranking is one-to-two-cartel-dominated and the case-coverage-breadth test is non-significant (p = 0.487). The federal leg is therefore a single-system, concentrated-anchor stress test of the protocol, not a multi-case operational generalization. Both the breadth disclosure and the per-case leave-one-case-out table appear in Appendix G.
  • Direct defendants null reproduces by design. Direct-defendant FL-binary AUC is ≈ chance (0.465–0.503), reproducing the loser-side scope boundary.

New site material (this version):

  • AN-043 — Federal opportunity-adjusted validation (ComprasNet). The two-platform audit anatomy with the federal numbers.
  • Scorecard: new cross-platform portability row, verdict "deflation replicates — audit protocol ports (two platforms)" — explicitly NOT a promotion to Confirmed (partially overlapping anchors).
  • Landing page: two-platform extension announced, with the partially-overlapping-anchors caveat stated in the text.

Manuscript home of these numbers: §5 The Audit on a Second Platform and Appendix G (the federal audit battery); the federal column is bound via the \valFedAud* macro namespace. Paper 45 pp + appendix 40 pp; builds 0 errors / 0 undefined.

v22.2 — June 4, 2026 (hostile-review armor + narrative repair)

Five-referee hostile pre-submission review executed, and its Action Plan applied. Integrity verdicts were clean (≈190 numbers traced, 0 fatal; 0 fabricated references); the substantive critique — "the audit is tilted toward its own deflationary conclusion" — was answered with computation, not prose:

  • Exposure-benchmark leakage tiers. Observed contact ranks the label at 0.905 and plug-in expected contact at 0.985 (mechanical label encoding); the firm-leave-one-out benchmark gives 0.855; recomputing every cell's defendant density from non-cobidder rows only leaves genuine label-blind opportunity at 0.553. The earlier "exposure-only is the better model" framing was itself partially mechanical and is retired; the deflation now rests on the anchor-agnostic battery (negative controls, powered permutation, label-frozen timing).
  • Falsifiability. A positive-control score ranks at 0.953 within the same strata that put the screen at chance; the within-stratum null is stable across granularities (0.49–0.60).
  • Power. The matched-strata permutation is correctly sized and rejects with probability 0.97 at true within-AUC 0.55 — non-rejections now bound any residual.
  • Label-frozen timing. Pool, score, and label ingredients all frozen on 2009–2016: the frozen score ranks new 2017–2019 defendant contact at 0.713 among 13,051 incumbents — generic contact forecasting, not cartel-specific signal.
  • Narrative: contributions reordered (organizational result first: the award-to-bid frontier is the design object); "transferable framework" → disciplined audit protocol + portable principle; stakes documented via the institutional red-flag lineage (Fazekas–Kocsis 2020 BJPolS; OECD 2022); §5 leads with what adjustment removes; §7 gaming reframed as instability of a published cutoff; §2.1 folded into §1; appendix deduplicated.
  • Abstract 133 words; paper 40 pp + appendix 34 pp (single-spaced); builds 0 errors / 0 undefined. Post-repair calibration: P(desk) ≈ 0.25.

v22.1 — June 4, 2026 (superseded by v22.2 — canonical-target re-estimation)

Credibility repair: reproducible, non-circular validation target. The previous main target (193 cobidders, a static file whose builder was archived and whose membership was conditioned on the frequent-loser flag) is retired. The new main label is built from current scripts: 651 unique always-loser cobidders — zero-win firms sharing ≥ 1 BEC tender-item with a BEC-active adjudicated CADE defendant; defendants excluded; the frequent-loser flag is never used to construct the label (341 of 651 positives are frequent losers, 310 are not). All downstream tables, figures, and macros re-estimated.

Headline numbers under the honest label (supersede every earlier figure):

  • Raw award-layer score: ROC 0.761 (PR 0.143; lift 5.6× at top 500). The earlier 0.946/0.924 figures were inflated by the label's circularity.
  • Opportunity-only model: 0.905 unconditional — outperforms the score.
  • Within-opportunity AUC ≈ 0.47–0.51 (chance). Nested increment +0.010 (p = 0.013); matched-stratum permutation p = 0.127 (ns); FL-enrichment p = 0.067 (ns). No robust residual signal net of opportunity.
  • Intensity-restricted sensitivity (contact ≥ 2, 368 positives): deflationary result strengthens — exposure-only 0.956; within-stratum 0.506; increment +0.003 (p = 0.47).
  • Case concentration: largest case 32% of positives (drop-largest PR −37%). Strict prospective ranking: zero true positives in the top 500, every year.
  • Bid benchmark (16,731 firms / 651 positives): award 0.760 ≈ bid RF 0.717; combined falls below award-only under case-grouped folds.
  • Cost-recall at K1 = 2000/k = 500: TP 116; firm reduction 88% vs bid-row reduction 33%.

Note on analysis-note (AN) pages: AN pages dated before June 4, 2026 document historical runs under the superseded label; where they conflict with the paper, the paper wins. Main pages (results, robustness, findings, hypotheses) are reconciled to the canonical numbers.

Main 39 pp + online appendix 33 pp; builds 0 errors / 0 undefined references; claims scan critical = 0.

v22 — June 3, 2026 (superseded by v22.1 — JLEO submission)

Title: Cheap Signals, Costly Proof: The Reach and Limits of Award-Layer Screening in Cartel Enforcement. Canonical source: work/v22-editor/submission_clean/. JLEO submission-ready (GO_WITH_MINOR_REPAIRS); main 38 pp + online appendix 31 pp.

Substantive reframe — "reach and limits." The paper is no longer cast as a deployable cartel screen. The contribution is now (1) reframing procurement-cartel screening as evidence allocation under costly observability, (2) a transferable decomposition method that separates a genuine ranking signal from mechanical opportunity exposure and single-case concentration, and (3) a map of reach and limits.

  • Opportunity decomposition is the headline. Pooled/exposure-only discrimination AUC 0.946 falls to 0.7715 once procurement opportunity is held fixed; the genuine increment over mechanical exposure is +0.042 (DeLong p ≈ 2×10⁻⁶). The earlier firm-level 0.924/0.864 figures are now reported as exposure-inflated pooled numbers, not the discrimination headline.
  • Limits front-paged. The ranking is retrospective among incumbents (full-universe ROC ≈ 0.55; sequential strict-timing infeasible) and leans on one case (leave-largest-case-out PR-AUC 0.126 → 0.036, −71%; one case ≈ 55% of positives). AUC ≈ 0.49 against direct CADE defendants is a scope boundary by design.
  • "83% reduction" retired. The cost-recall frontier reports firm-pool reduction ≈ 88% but bid-row footprint reduction ≈ 33% at one operating point (K₁ = 2000); the frontier, not a universal cutoff, is the design object. Combined award+bid AUC 0.962 is a full-observability upper bound.
  • Price = scope, not damages. Broad +0.064 → overlap-cell ATT −0.097 → Q4 +0.041 (only positive cell); cover-bidding "theater" mechanism not identified.
  • Language disciplined throughout: "adjudication-anchored exposure" (not cartel-adjacency / membership); "orders forensic priority" (not detects / proves); cobidders are a validation target, not cartel members.
  • Site: all sections rewritten to this reframe; the Mind Map was replaced by a paper DAG of the argument; one fabricated reference removed after verification.

v20 submission-clean — May 23, 2026 (superseded — JLEO)

Title: Cheap Signals, Costly Proof: Award-Layer Evidence Triage in Cartel Enforcement. Canonical source: work/v20-editor/submission_clean/ (full integral copy of the v18-editor build, last recompiled 2026-05-25). Headline: an 83% bid-microdata reduction recovering 131 of 193 adjudicated cobidders; firm-level discrimination AUC 0.864 under temporal holdout (0.924 in-sample).

2026-05-25 refresh.

  • D1 horse-race re-run completed (scripts/36_gate_d1_harmonized.R, FL14 = tenders_count ≥ 14). Continuous AUC 0.939 [0.932, 0.946] dominates binary FL14 0.924 [0.921, 0.926]; DeLong Z = −4.38, p = 1.2 × 10⁻⁵ — direction preserved, marginally tighter than the earlier (superseded) Z = −4.30 / p ≈ 2 × 10⁻⁵ that had been computed under the > 14 (FL15) cut. AN-011, AN-015, robustness, and the award-bid-complementarity hypothesis page updated; the "pending the D1 re-run" caveats are retired.
  • Site PDFs (paper.pdf, online_appendix.pdf) re-synced to the 2026-05-25 v20-editor build (intro / validation / forensics / price-scope section edits + macro corrections).

New result:

  • AN-041 — volume-matched cobidder audit. The "strongest remaining within-data audit": matching cobidders to non-cobidder FLs on tenders_count (standardized difference 0.49 → 0.00, PS nearest-neighbour
  • CEM), the within-FL profile distinctness is not a volume artifact — product HHI (d +0.47), winner-pair spread (−0.56), and median gap-to-winner (−0.25) hold or strengthen. Honest casualty: the AN-031 bid-dispersion elevation collapses (+0.05, n.s.) and is dropped. Script scripts/74_volume_matched_cobidder_audit.R.
  • AN-042 — volume-matched timing audit (null). Candidate second bid-conduct channel: pregão bid timing (revision intensity, inter-bid interval, last-bid position, engagement span) under the same matching. No dimension survives (all Wilcoxon p ≥ 0.23). Bid-conduct distinctness stays single-channel (gap-to-winner). Script scripts/75_volume_matched_timing_audit.R.
  • H5 promoted Mixed → Partial (strongly supported) (structural scope). The structural distinctness survives volume matching (AN-041); AN-021 / AN-032 re-labelled as scope limits on a causal-mechanism reading the hypothesis does not assert. Incorporated into the manuscript: §5 body + Appendix D.3 ("Volume-Matched Within-FL Profile") + \valVM* macros. Confirmed still blocked by the non-BEC replication constraint.

JLEO editorial optimization pass (main paper 42 → 40 pp):

  • Abstract microcorrection: "exposure audits" → "validation audits"; "procurement overlap" → "opportunity-set exposure".
  • Section 4 sharpened: the Three-Classifier Timing Battery and the Universe-Anchored Scope Matrix moved from the main text to Appendix D (D.6 and D.4); the full sham-FL distribution and the price-coefficient-under-sham moved to D.2. Main text keeps the compact Validation Architecture threat table.
  • Section 7 re-downgraded to scope/limits: the monetary-scope range (R$ figures / % of spending) and the full selection/within-cell decomposition table moved to Appendix E (E.3 and E.2); a bounded one-sentence pointer remains in the main text.
  • Mechanism-identification language softened throughout to descriptive scope evidence that does not identify a mechanism, a causal price effect, overcharges, or damages.

Site reconciliation (this version):

Note: the v18 entry below describes the Three-Classifier Timing Battery and Universe-Anchored Scope Matrix as main-text Tables 3/4; this pass moved both into Appendix D.

v18 — May 2026 (JLEO submission-clean)

Full hypothesis audit + manuscript integration:

  • All 8 hypotheses graduated to Partial (strongly supported) except H2 (Confirmed by structural design). Promotion to Confirmed for H1/H3/H4/H6/H7/H8 requires non-BEC replication; see the COMPRASNET_PATH_TO_CONFIRMED.md planning memo in the repo root.
  • 14 new AN pages added (AN-025 through AN-038) documenting the formal sham permutation, cutoff sweep, subsample robustness, universe-anchored stratum scope matrix, standardized differences battery, three-classifier timing battery, firm/market persistence, bid-level behavioral profile, matched-heterogeneity audit, formal Imhof incremental DeLong, sequential gatekeeping envelope, full architecture cost-of-evidence matrix, K-fold CV precision stability, sign-reversal decomposition, and item-group cell audit.
  • 38 total AN pages now with figures (19 from existing R outputs, 4 from PDF→PNG conversion, 19 generated via new scripts/gen_an_figures.py).
  • Manuscript integration: surgical inserts in §4.2, §4.3, §5.4, §7.2 bring the new audit-chain numbers into prose. Two new tables (Table 3 Three-Classifier Timing Battery and Table 4 Universe-Anchored Stratum Scope Matrix) added to §4. tab:price_scope_submission extended with broad-sample and overlap-cell-unweighted rows showing the full three-spec progression. Paper now 39 pages.
  • values.tex extended with ~95 new macros for the new audits; 6 macro names corrected (LaTeX names cannot contain digits).
  • Site sections updated: advanced.md rewritten from stub to substantive methodological reference; results.md extended with formal sham permutation + three-classifier + scope matrix below- random subsection; robustness.md extended with placebo / scope matrix / CV stability / cutoff sweep / subsample / matched-het audit subsections; extensions.md extended with architecture cost-of-evidence matrix and three-classifier battery.

v5.1 --- April 2026

Referee-report edits (mr-frequent review):

  • Fixed combined AUC (0.502 → 0.608): included omitted z_skew component in combined score
  • Added "Why combination degrades" paragraph explaining AUC = 0.61 as empirical signature of Regime 2
  • "nearly orthogonal" → "largely non-overlapping" with theoretical ceiling note (corr/max ≈ 1/5)
  • "lower bound" → "conservative estimate" for cross-fit
  • Sample size clarified: body reports 1,654,447 (full sample) with footnote explaining 1,654,401 (prices) and 1,653,156 (non-FL firms)
  • AUC interpretation: added sentences clarifying AUC reflects high-participation proximity, not conditional discrimination
  • Measurement error: "Classical binary ME" → explicit assumption that FP > FN justifies attenuation direction
  • Bajari--Ye tender FE: acknowledged ambiguity ("suggestive but not definitive")
  • Fixed pre-existing LaTeX bug in tab_fl_crossfit.tex (unescaped underscores)
  • Updated MkDocs site to reflect all changes

v5 --- April 2026

JLE repositioning and institutional framing:

  • Repositioned for Journal of Law and Economics (institutional framing throughout)
  • Abstract leads with minimum-bidder rules giving cartels a reason to deploy cover bidders
  • Price range narrowed to 3.6--7.7% (cross-fit, OLS, matching), replacing 4--21% from v4
  • IV (19.4%) reframed as measurement-error diagnostic, not a primary estimate
  • Finding now Regime 2 (coordinated), not Regime 1 (complementary): \(\sigma_c / \sigma_g = 0.72\)
  • \(\gamma > 0\): strategic complementarity (more cover bidders in MORE competitive tenders)
  • Convite 3.8% vs pregao 9.3% --- larger where voluntary, signal dilution in convite
  • AUC = 0.94 against CADE convictions (primary validation)
  • Horse-race: correlation 0.06 with Imhof-style CV proxy; FL rises to 0.084 when CV added

New content:

  • Three new figures: corner solution, dispersion paradox, enforcement flowchart
  • Five JLE-relevant references added (Posner 1970, Caoui 2022, Ghosal & Sokol 2014, Harrington 2015, Baranek & Titl 2024)
  • Lei 14.133/2021 discussion with two testable predictions
  • Three-stage enforcement pathway (screen, triage, investigate)
  • Welfare reframed as cost-benefit of screening deployment
  • 12.5x gradient in oversight heterogeneity across PBU size quartiles
  • Minimum-bidder constraint variation: 7.6% voluntary premium vs. \(-0.160\) forced interaction
  • Five supporting diagnostics (M1--M5) with joint assessment
  • Non-claims paragraph and explicit limitations section

Tables and figures:

  • 35+ tables, 26 figures
  • ~56 pages, natbib/bibtex compilation

Pipeline:

  • 24 R scripts (22 numbered + figures_new.R + 00_setup.R), full pipeline in ~8 minutes
  • New scripts: 14_quality_checks.R through 22_threshold_heatmap.R plus figures_new.R

v4 --- March 2026

Manuscript:

  • Rewrote contributions paragraph with explicit H1/H2 structure and scope limitation
  • Softened over-claims throughout (9 instances: "establish" to "show", "confirming" to "indicating")
  • Expanded IV section: preferred spec declaration, lambda interpretation, LATE discussion
  • Bajari--Ye restructured around 3 sequential claims with magnitude hedge
  • DiD shrunk to 2 paragraphs (complementary, underpowered)
  • Mechanisms: added connective sentences and bridge paragraph
  • Heterogeneity: relabeled "high/low-suspicion" to "concentrated/competitive-market FL"
  • Added robustness roadmap paragraph (6 groups)
  • Added implementation blueprint (Flag, Triage, Investigate, Monitor)
  • Added enforcement integration and legal boundary sentences
  • Abstract trimmed to 136 words with policy contribution sentence
  • Broader implications paragraph added to conclusion
  • Float numbering fixed: appendix counters reset (A.1, B.1, etc.)
  • Added missing cross-references for 3 main-body floats

Tables and figures:

  • 35 tables (12 main body, 23 appendix)
  • 14 figures (6 main body, 8 appendix)
  • 59 pages, 0 errors, 0 undefined references

Pipeline:

  • 13 R scripts, full pipeline in ~8 minutes
  • New tables: IV balance, IV Panel C, Bajari--Ye tender FE, network interactions, homogeneous CV, welfare bounds (updated)

v3 --- February 2026

  • CADE validation section with CNPJ enrichment (3 FL firms matched)
  • Network-split IV regressions
  • Bajari--Ye tender fixed effects specification
  • Homogeneous sub-samples via price-CV approach
  • Welfare analysis with cross-fit bounds
  • Staggered DiD with MDE and Rambachan--Roth sensitivity

v2 --- January 2026

  • Bajari--Ye exchangeability and conditional independence tests
  • Network-based FL classification (winner HHI, repeat partners)
  • Regime test: complementary vs. coordinated cover bidding
  • Sensitivity analysis (Cinelli & Hazlett, 2020)
  • Matching estimators (CEM and IPW)
  • Year-by-year coefficient stability

v1 --- December 2025

  • Initial manuscript with OLS and IV regressions
  • FL definition (always-losers + IQR threshold)
  • Leave-one-out instrumental variable
  • 18 tables, 6 figures
  • Threshold sensitivity and cross-fitting robustness