Changelog¶

v17 — May 2026 (current)¶

Route 2 substantive revision push and storytelling polish for JLEO.

Title: Frequent Losers: A Cover-Bidder Screen Without Bid Microdata. Cunha "Frequent Losers" as a noun-phrase technical term and promises the most original methodological move (operating without bid microdata at the operational layer).
New section §forensic carries the architectural test as the contribution-central section. Three findings: same-target AUC comparison (FL alone 0.881 vs Imhof full 0.846), non-redundant signal in combination (+0.096 AUC, DeLong p = 1.2 × 10−26), and the sequential gatekeeper rule that delivers an 83% bid-microdata footprint reduction at an 8% recall cost relative to joint scoring.
Temporal-holdout audit of the gatekeeper added (script 64): the sequential rule matches or exceeds joint scoring under temporal holdout in absolute true-positive count, so the data-envelope reduction does not trade off against out-of-time recall.
Bid-level extension of the within-stratum theory bridge (script 62): per-bid Valor Unitário Proposta and Flag Vencedor from the LANCES export, restricted to losing bids in convite or pregão. Cobidders bid plausibly close to winners (median gap d = −0.28) with elevated within-firm dispersion (d = +0.15). Multivariate logit holding log(1+tenders_count) constant confirms both signs at p < 10−3. The signs adopt the credible-cover-bidding (R2) reading of Marshall–Marx and Asker over the textbook (R1) reading of Porter–Bajari–Ye.
Sign-reversal segment-level decomposition (script 61): the broad-sample positive concentrates in tender-value Q4 (+0.046, robust to overlap restriction and ATT reweighting). Trim sensitivity strengthens the negative ATT estimate (from −0.097 to −0.118 at top-decile trim), foreclosing the few-cell-artefact reading.
Observability framing reframed: distinction between operational layer (audit-court-queryable in minutes; carries award envelope only) and forensic-recoverable layer (per-bidder bid amounts, accessible through administrative request with weeks of latency). The screen operates on the operational layer; the bid-microdata pipeline appears in §forensic as the benchmark and as the Stage-2 input the screen triages, never as a screening input itself.
Online Appendix A formal R1/R2 distinction added at the end of the proof appendix; framework predictions about firm-level participation intensity, deployment proximity, and portfolio specialisation are unaffected by R1/R2; only the bid-level reading of \(b_C\) is.
Manuscript: 47 pp body + 20 pp online appendix. All thirteen main sections and seven OA sections at storytelling 9.5. Disclosure footnote on LANCES export access embedded in §emp_estimand.
v16 byte-frozen as recoverable hedge under work/v16-editor/.

v15 — May 2026¶

Single-thesis rewrite for JLEO.

Title realigned with body: Frequent Losers as a Cartel Screen: Detecting Cover Bidding Without Bid Microdata. Construct scope explicit (cover-bidder population, not cartel ringleaders).
Abstract reframed in eight sentences: problem, puzzle, construct, headline range, oversight gradient, framework, identification stance, three-stage administrative pathway.
Identification scope made explicit and descriptive throughout. RDD at the convite/pregão statutory caps and DiD using Decreto 9.412/2018 both return null at the available bandwidths and are reported openly.
Headline range reported as +3.6% to +7.7% across five estimators (full-sample OLS, PBU fixed effects, IPW, CEM, cross-fit), with the cross-fit decomposition isolating a structural component (+3.6%) from within-sample classification noise (~+2.8 percentage points).
Procuring-unit-size oversight gradient reported as 12.6× between extreme quartiles (+21.4% in Q1 to +1.7% in Q4), the strongest source of heterogeneity in the data.
Minimum-bidder rule (convite Article 22 §3) treated as institutional context rather than as the channel the construct exploits. The data flatly reverse the institutional reading: the construct's signal is sharper in pregão (no minimum-bidder rule) than in convite. We read this as scope information about where the construct fires most cleanly, not as positive evidence for any institutional theory.
CADE validation reframed as bounded discrimination: lower bound AUC 0.748 (4 cases pre-2020, 210 cobidders, 108 frequent losers; contemporaneous reading) and upper bound AUC 0.924 in-sample / 0.864 temporal holdout (full 12 cases, 193 cobidders; prospective reading). Three named firms (Sol Tecnologia, Nova Esperança Locadora, Jofran Comércio) flagged before their cartel adjudications.
Operational metrics reported under temporal holdout (train 2009–2016, test 2017–2019) as the deployable projection. In-sample precision overstates the holdout column by ~50% at top-500 (0.132 vs 0.070, retention 53%); the gap is disclosed explicitly and the holdout column carries all operational claims.
Horse race against bid-distribution detectors reframed as informational complementarity, not dominance. The construct reaches AUC 0.903 against the full Imhof–Wallimann pipeline's AUC 0.888 at substantially lower data cost; combined classifiers reach AUC 0.955–0.962, with the gain identifying non-redundant signal in participation-only and bid-distribution dimensions.
Mechanism heterogeneity reorganized into three subsections by interpretive stake: continuous loss intensity dominates the binary cutoff (DeLong p = 2×10⁻⁵), modal asymmetry between pregão and convite as scope information, and what the data do not support (cell-heterogeneity grid not a positive test of cartel rotation; first-time-FL prediction in the predicted direction but compresses substantially under propensity-score matching).
Limitations section reorganized into three subsections by stake: identification scope, construct scope (with three competing readings of the participation footprint), external validity and portability (with three structural prerequisites: commodity-heavy item composition, electronic platforms with publicly auditable participant lists, institutional asymmetry between winner and loser sides).
Cover-bidding framework compressed to three formal results: a lemma establishing wins=0 as the equilibrium choice of cover bidders, a proposition identifying log(1+tenders_count) as a sufficient ranking statistic for the cover-bidder type, and a proposition delivering ∂m*/∂θ_k < 0 as the comparative static the procuring-unit-size oversight gradient instantiates.

Manuscript:

60 pages, single-thesis discipline, lean appendix.
Bid-level structural mixture, Bajari–Ye coordination tests, network-split heterogeneity, dyadic-linkage permutation, Cox survival, market-selection / Regime 1 / Regime 2 cover-bid-distribution propositions, welfare illustrative calculations, and cross-sector replication removed: not used by the v15 main body.
Manuscript pipeline preserved; analysis blocks mapped to v15 sections in Replication.

v5.1 — April 2026¶

Final referee-report edits before the JLEO repositioning. Combined-AUC fix, sample-size clarifications, measurement-error wording tightened, Bajari–Ye tender FE reframed, MkDocs site updated.

v5 — April 2026¶

JLE repositioning and institutional framing. Headline narrowed to +3.6 to +7.7%. AUC = 0.94 against CADE cobidders. Strategic complementarity, regime selection, three-stage enforcement pathway introduced. Subsequently superseded by the v15 single-thesis rewrite.

v4 — March 2026¶

Manuscript polished with H1/H2 contributions, expanded IV section, Bajari–Ye restructured, DiD as complementary exercise, implementation blueprint added.

v3 — February 2026¶

CADE validation with CNPJ enrichment. Network-split IV regressions. Bajari–Ye tender fixed effects. Welfare with cross-fit bounds. Staggered DiD with Rambachan–Roth sensitivity.

v2 — January 2026¶

Bajari–Ye coordination tests. Network-based FL classification. Regime test (complementary vs coordinated). Cinelli–Hazlett sensitivity. CEM and IPW matching.

v1 — December 2025¶

Initial manuscript: OLS, IV, FL definition (always-losers + IQR threshold), leave-one-out instrument, threshold sensitivity, cross-fitting robustness.