AN-015: Gate D1 — harmonized same-sample horse race¶
Intuition (plain-language)
D1 is the first of four 2026 "gate" diagnostics that decided the paper's framing. On a single harmonized firm set it pits the continuous score against the binary FL14 head to head. The continuous version dominates (0.939 vs 0.924; DeLong p ≈ 10⁻⁵) and the price coefficients agree in sign. The economic payoff: it confirms that loss intensity — not a particular cutoff — is the primitive, locking the rule that the binary is a deployable simplification of a continuous signal.
Question¶
D1 gate diagnostic: does the continuous score dominate FL14 on a harmonized same-sample horse race, and do the price coefficients align? This is the diagnostic D1 from the 2026-04-30 gate battery that locked the JLEO submission strategy.
Design¶
- Sample: harmonized same-sample set with both FL14 and continuous variants available; N = 1,653,658 item-firm observations.
- Specifications:
- AUC test: DeLong paired AUC comparison.
- Price test: coefficient on each score in (a) FL14 alone, (b) continuous alone, (c) joint.
Results¶
AUC comparison:
| Score | AUC | 95% CI |
|---|---|---|
| FL14 binary | 0.924 | [0.921, 0.926] |
| Continuous log_tc | 0.939 | [0.932, 0.946] |
D1 passes: continuous dominates the binary flag (DeLong Z = −4.38, p = 1.2 × 10⁻⁵); the gap is 0.015 under the corrected FL14 (≥ 14) definition. The D1 re-run (2026-05-25) confirmed the direction; the earlier Z = −4.30 / p = 2 × 10⁻⁵ were computed under the superseded > 14 (FL15) cut.
Price coefficients (item × year × PBU FE):
| Spec | FL14 coef | SE | log_tc coef | SE |
|---|---|---|---|---|
| FL14 alone | +0.0653*** | 0.0216 | — | — |
| log_tc alone | — | — | +0.0188*** | 0.0055 |
| Joint | −0.0746* | 0.0383 | +0.0349*** | 0.0108 |
Single-score specifications show same-sign price coefficients (both positive); the joint specification reveals the sign reversal of FL14 once the continuous score absorbs the persistent-loss signal.
Macros: \valHorseAUCBin, \valHorseAUCCont, \valDeLongZ,
\valDeLongP, plus the \valHorse* price-coefficient series.
Interpretation¶
D1 passes unambiguously. The continuous score dominates the binary in both discrimination and in carrying the persistent-loss signal in the price equation. FL14 is demoted from "the construct" to "the deployable implementation of the construct"; the continuous log(1+tenders_count) is the empirical primitive.
The sign reversal of the FL14 coefficient in the joint specification flips the price reading: under the joint model the binary picks up truncation noise that disconnects from the price signal. This is part of the scope-vs-damages story in AN-019 and H:price-scope-sign-reversal.
D1 is one of four gate diagnostics (D1–D4) from 2026-04-30. D2 (AN-016) disqualified γ++ reframing. D3 (AN-017) preserved the loser-side thesis without FL14. D4 (AN-018) confirmed CADE-winner-heavy structure.