H:award-bid-complementarity — Award and bid layers are complementary (division of labor, not dominance)

Sequencing matters under costly observability. The hypothesis is that the cheap award layer and the expensive bid-distribution layer carry complementary information on the same adjudication-anchored target — a division of labor, not a contest. The award layer ranks where to look; the bid layer evaluates what is found. Combining the two beats either alone, but the point is the division of labor, not that one dominates.

Intuition (plain-language)

There are two layers of procurement data: cheap administrative award records (who participated, who won) and expensive bid-level microdata (every bid amount in every tender). The hypothesis: they carry complementary, non-redundant information. The data support this only conditionally. On the pooled diagnostic the two layers are comparable (award ≈ bid: 0.760 vs 0.717). The combined model beats award on precision–recall under a random cross-validation split (PR 0.188 vs 0.143), but falls below award-only once folds are grouped by case (PR 0.103 vs 0.143). So the complementarity is conditional on this implemented benchmark and case-fragile — a division of labor, not a dominance or a robust gain. The combined number is leakage-sensitive and is not an operational claim.

Evidence strength: Mixed (conditional, case-fragile). The complementarity claim on the 651-cobidder target: (i) Award ≈ bid pooled (AN-010): the transparent bid-moment random forest reaches ROC 0.717 / PR 0.116; the award continuous score 0.760 / PR 0.143; FL14 0.688. Comparable, not a clean ceiling above both (same-sample gatekeeping pool: bid 0.665, award 0.665, combined 0.727). (ii) Combined gain is fold-dependent (AN-033): the combined model 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). The increment is conditional on the implemented benchmark and case-fragile. (iii) Division of labor, not dominance (AN-033): the award layer cheaply ranks where to look; the bid layer evaluates what is found. The gain from combining is a complementarity diagnostic, not an outperformance claim (Spearman award–bid 0.544). (iv) Same-sample horse race (AN-011, AN-015): continuous loss intensity carries the award-layer signal slightly better than the FL14 cut. (v) Leakage / case sensitivity (AN-014, AN-034): the combined number attenuates out of sample and under case-grouped folds; it is a division-of-labor diagnostic, not an operational deployment guarantee. The verdict is Mixed: complementarity holds conditionally (random CV) but is case-fragile (case-grouped folds). Non-BEC replication is the path to any generalizable claim.

Theory

Bid-distribution screens \citep{imhof2018screening,imhof2019detecting, wallimann2023machine} evaluate suspicious bidding once bid-level data are recovered. Award-layer signals are visible earlier. If the two carry different information margins, the joint signal should beat either alone. This is the cost-of-evidence framing in \citet{chassang2022robust,harrington2008detecting}.

Prediction

On the 651-cobidder target:

• award ≈ bid on the pooled diagnostic (0.760 vs 0.717);
• combined beats award on PR under random CV (0.188 vs 0.143);
• but the combined increment is not robust to case-grouped folds (combined PR 0.103 falls below award-only 0.143) — the gain is conditional.

Competing prediction

Award is redundant / combination is fragile. If the bid layer already encodes the loser-side information, or if the combined gain disappears once folds respect case structure, the complementarity carries no robust value. The data land in between: a random-CV gain that does not survive case-grouped folds — hence the Mixed verdict.

Case evidence

Imhof-style bid-distribution screens have been shown to discriminate cartelized auctions in Swiss procurement \citep{imhof2018screening,imhof2019detecting}. The Brazilian setting has both layers available — the test asks whether they substitute or complement.

Empirical test

• Sample: BEC firms with both award and bid features available.
• Outcome: cobidder indicator.
• Specifications:
• award-only: FL14 and log(1+tenders_count);
• bid-only: Imhof-style features (within-tender bid moments, CV, relative rank);
• joint: union of features in a single classifier.
• Identification: same-sample horse race; DeLong test for AUC difference.

Data requirements and limitations

Requires the LANCES bid-level export. Limitation: the bid layer is more expensive to recover, so the joint score is a full-observability upper bound rather than an operational benchmark. The complementarity result therefore supports — but does not require — the gatekeeping deployment of H:gatekeeping-cost-of-evidence.

Evidence

Analysis	Bearing	Status	Key takeaway
AN-010 (bid-moment benchmark)	Mixed	done	Bid RF 0.717 / award continuous 0.760 / FL14 0.688 / combined 0.756 (random CV) — comparable; combined PR falls below award-only under case-grouped folds
AN-011 (horse race)	Supports	done	Continuous loss intensity carries the award signal slightly better than the FL14 cut
AN-015 (D1 harmonized)	Supports	done	D1 passes; price coefficients align in single-score specs
AN-033 (incremental decomposition)	Mixed	done	Combined beats award on PR under random CV (+0.045) but falls below award-only under case-grouped folds; complementarity conditional and case-fragile (Spearman 0.544)
AN-034 (sequential envelope)	Direct	done	Award ranks where to look, bid evaluates what is found; combined number is leakage- and case-sensitive, not an operational guarantee

Open tests

• Decomposition of complementarity into modality (Convite vs Pregão).
• Marginal AUC contribution of each Imhof feature in the joint model.
• Sequential envelope under temporal holdout (some data already in AN-013).
• Cross-jurisdiction replication of the same incremental DeLong tests on ComprasNet federal or a non-BR procurement panel.

Why not confirmed?

Under the non-circular label H6's within-data evidence is mixed, not clean:

• On the pooled diagnostic the two layers are comparable (award 0.760 ≈ bid 0.717); there is no clear ceiling above both.
• Combining the layers beats award on PR under random CV (0.188 vs 0.143) but falls below award-only once folds are grouped by case (0.103 vs 0.143) — the combined gain does not survive case-grouped folds.
• The combined number is an in-sample full-observability diagnostic and is leakage- and case-sensitive, so it is read as conditional complementarity, not as an operational guarantee or a dominance claim.

Two artifact families further remain untested by the within-data evidence:

1. BEC-specific bid-data structure. The Imhof features (cv_mean, cv_sd, skew_mean, kurt_mean, spread_mean, minmax_mean, second_low_mean) are computed from BEC's specific bid-recording convention. If BEC bid microdata has a structured noise pattern that interacts with award-layer participation in a particular way, the complementarity could be partially a feature-construction artifact. Imhof's original work uses Swiss data with different bid conventions; the complementarity might look different there.

2. Same-sample CADE label structure. Both the bid-moment benchmark and the award layer are evaluated against the same 651-cobidder positive class. If CADE adjudications systematically select cartels where loser-side participation differs from bid distribution (e.g., cases driven by tip-offs about bid patterns rather than participation patterns), the conditional complementarity could be specific to CADE's adjudication selection.

Both can only be assessed by replicating the incremental tests on a non-BEC panel with an independent cartel anchor. Under the current evidence H6 is Mixed (conditional, case-fragile), consistent with the project-wide rule documented in findings/index.md.