Skip to content

The FL ranking is null on direct CADE defendants — by design

Intuition (plain-language)

The same logic that makes the screen work also fixes its limit. A score built on "this firm never wins" is, by construction, blind to the firms that did win the rigged contracts — designated winners are not zero-win firms. So the screen scores essentially at random (AUC ≈ 0.49) against the convicted ring-leaders. That is not a failure to hide; it is the honest boundary of a loser-side tool: it tells you where to look, never who is guilty.

🟢 The frequent-loser ranking, evaluated against the 47 direct CADE defendants in 2009–2019 procurement-cartel adjudications, returns AUC = 0.491 [0.461, 0.520] — indistinguishable from random. The null is the predicted finding under the loser-side scope: a score built from persistent zero-win participation cannot rank designated winners, who are not zero-win firms by construction (AN-007).

Diagnostic D4 (AN-018) confirms the null mechanistically: only 7 of 47 direct defendants (14.9%) are always-losers, and their median win rate is 0.261 (vs 0.086 for cobidders, Mann–Whitney p < 0.05). Direct defendants are structurally winner-heavy.

The null survives every evaluation regime: 0.506 in the leakage audit (AN-014) raw item-level, 0.511 under temporal holdout — all within sampling distance of 0.5.

This finding 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. Without it, the screen would overclaim; with it, the scope is honest.

Caveat. The null is the predicted finding under the loser-side scope interpretation; reading it as a failure of the screen would misunderstand the construction. The reading is 🟢 because the result is load-bearing for the framing of the paper and survives every audit in §4 of the manuscript.

Sources.

  • Own analysis: AN-007 (direct-defendant AUC), AN-018 (D4 gate — CADE winner-heavy), AN-014 (null in audit regimes), AN-027 (scope matrix; row 4: participation count vs direct CADE = 0.383, below random — score actively repels winner-heavy defendants).
  • Cross-refs: H:direct-defendants-null; docs/results.md.
  • Macros: \valAUCdirectCADE (0.491), \valAUCdirectCADECI, \valDirectCADE (47), \valDirectShareAL (14.9%), \valDirectMedWR (0.261), \valOthersMedWR (0.086).
  • Validation: backing scripts scripts/33_auc_direct_cade.R, scripts/39_gate_d4_cade_winner_heavy.R, scripts/40_leakage_audit_d3.R.