Skip to content

The FL binary flag is null on direct CADE defendants — by design

Intuition (plain-language)

The same logic that makes the loser-side flag concentrate risk also fixes its limit. A binary flag 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 FL flag scores essentially at random (AUC ≈ 0.49) against the convicted ring-leaders. Participation volume, on the other hand, ranks defendants moderately above chance (0.66–0.70), because heavy participants are easier to flag generically. The boundary is on the loser-side binary flag, not on participation arithmetic: the flag tells you where to look among losers, never who is guilty.

🟢 The frequent-loser binary flag, evaluated against the 47 direct CADE defendants (the legal anchors) in 2009–2019 procurement-cartel adjudications, returns AUC ≈ 0.49 — indistinguishable from random. This is a predicted scope boundary, by design, not a failure: a binary flag built from persistent zero-win participation cannot rank designated winners, who are not zero-win firms by construction (AN-007). The continuous participation score, by contrast, ranks defendants moderately above chance (0.658 full / 0.695 strict) — heavy participants are easier to flag generically — so the asymmetry is specifically about the loser-side binary flag, not participation volume. The paper front-pages the binary null as the honest limit of an award-layer flag.

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 — whereas cobidders, being always-losers, have win rate ≡ 0 by construction. 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 "cartel detector" reading and pins the interpretation to a loser-side flag that concentrates forensic priority among losers, never proves conduct. Without it, the flag would overclaim; with it, the reach-and-limits map 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; the binary FL flag is silent on defendants (0.49) while the continuous participation score ranks them moderately above chance (0.66–0.70) — the asymmetry is on the binary flag, not participation volume).
  • Cross-refs: H:direct-defendants-null; docs/results.md.
  • Macros: \valAUCdirectCADE (0.491), \valAUCdirectCADECI, \valDirectCADE (47), \valDirectShareAL (14.9%), \valDirectMedWR (0.261). (The old \valOthersMedWR = 0.086 is dead — never cite: cobidders are always-losers, so their win rate is ≡ 0 by construction.)
  • Validation: backing scripts scripts/33_auc_direct_cade.R, scripts/39_gate_d4_cade_winner_heavy.R, scripts/40_leakage_audit_d3.R.