AN-007: AUC against direct CADE defendants¶
Intuition (plain-language)
The most important number in the paper is a near-coin-flip. The screen ranks loser-side roles, and a cartel's ringleaders are its winners — they capture the rents by taking the rotated contracts. So tested against the 47 direct CADE defendants the screen scores AUC 0.491, indistinguishable from random. This is not a failure to hide; it is the design made visible. A tool claiming to flag both cover bidders and ringleaders would be overclaiming, and this null is what disciplines every other claim in the paper.
Question¶
Does the FL score discriminate direct CADE defendants? Under the loser- side scope interpretation, it should not — the score is built over zero- win firms and cannot rank designated winners. A null AUC is the predicted finding.
Design¶
- Sample: all BEC firms (not restricted to always-losers).
- Positive label: 47 direct CADE defendants (AN-003).
- Specifications: AUC of FL14 (binary) and
log(1+tenders_count). - Identification: scope-check placebo; the null is the predicted finding.
Results¶
| Score against direct defendants | AUC | 95% CI |
|---|---|---|
| FL14 + continuous (combined) | 0.491 | [0.461, 0.520] |
| Item-level (raw) | 0.506 | (close to random) |
| Item-level under temporal holdout | 0.511 | [0.510, 0.513] |
Macros: \valAUCdirectCADE, \valAUCdirectCADECI,
\valAUCitemDirect, \valAUCItemDirectTemp.
D4 gate (AN-018) corroborates the null mechanistically: 7 of 47 direct defendants (14.9%) are always-losers; their median win rate is 0.261 vs 0.086 for the cobidder set.
Figure: AUC against 47 direct CADE defendants across three regimes — full panel (0.491, [0.461, 0.520]), item-level raw (0.506), item-level under temporal holdout (0.511). All within sampling distance of 0.5 (red line). The null is regime-invariant.
Interpretation¶
The result is exactly what the loser-side scope predicts. Direct defendants are winner-heavy by construction — the rank cannot cover them, and it shouldn't. This is the anti-claim that disciplines the rest of the paper: it rules out the generic-cartel-detector reading and pins the interpretation to loser-side adjacency. The 95% CI includes 0.5; the null survives every audit variant (AN-014: 0.506 raw, 0.511 temporal).
The result is load-bearing for the framing in §4.3 of the manuscript and confirmed by the auxiliary D4 diagnostic (AN-018).
Follow-ups¶
- Decomposition of direct defendants by role (declared winner, coordinator, other) and AUC by role.
- Sensitivity to alternative direct-defendant labels (e.g., extended CADE case set).
- Triangulation with the unified mechanism profile (AN-024).