THE FIRST CASE RUN. A Preregistered Protocol for the Discrimination Lemma, with Pre-Coding Calibration on Synthetic Data
Year: 2026
Type: preprint
Study design: Study Protocol
Abstract: Preregistered Protocol This document contains no results. It is deposited before any case is coded. HIGHLIGHTS ▸ Cases are selected by the availability of exogenous funding variation, never by suspicion of inversion. This is the protocol’s central defence and it operates at the sampling stage, before any test runs. A diagnostic for institutional capture that selects its cases from institutions the analyst already distrusts cannot produce evidence, only confirmation. The inclusion rule here is indifferent to what the tests will find. ▸ The register contains no expected-verdict column, and this is deliberate. Forty cases are listed with their funding instrument and their external outcome series. None carries a prior. An analyst who cannot state in advance which cases will return inversion is an analyst who cannot unconsciously arrange for them to. ▸ The analysis script refuses to run on unverified cases. Each case must clear four gates before its verified flag may be set: exogeneity checked against a primary source, an admissible external outcome series located, the named function sourced to a constitutive document, and confirmation that the institution was not created by the putative occupying power in the first place. ▸ The falsifier thresholds were calibrated on synthetic data before coding began — and one of them was found to be broken. The redundancy check asked whether any single test predicts the aggregate verdict above 90%. Simulation under the pure null returned 0.88, meaning the check nearly failed on noise. The metric was tautological: the aggregate is largely the heavily weighted tests, so of course they predict it. Recomputed leave-one-out, the null falls to 0.40 and the check becomes informative. ▸ Calibration also fixes the sample size. With per-test accuracy at 0.70, thirty cases separate cleanly from the null. At 0.55 the separation is thin and forty are needed. Roughly a quarter of cases will return INDETERMINATE, so forty must be coded to yield thirty decided. ▸ One falsifier is preregistered as likely to fire, and what its firing means is stated in advance. If the six tests are individually accurate above about 0.80, the redundancy check fails — correctly. Six measures that agree that often are one measure with six names, and the lemma would have to be restated as a single test with five corroborations. That outcome is not a defeat; it is a finding, and saying so now prevents it from being explained away later. ▸ Obstruction is counted as evidence, not as missing data. A case where the required series cannot be obtained for reasons traceable to the institution is coded as a positive result on the measurement test, not as a null. Without this rule the procedure would systematically under-detect exactly the condition it exists to detect. ABSTRACT The Discrimination Lemma supplies a six-test decision procedure for distinguishing institutional inversion from decay, corruption and cost disease. It has never been applied to a case. This document specifies the first application in advance of it: the population, the inclusion rule, the coding protocol, the aggregation rule, the falsifiers, and the analysis script, all fixed before any case is coded. Forty candidate cases are registered across six service domains and twenty-six countries, selected by a single criterion — the existence of a plausibly exogenous funding change of at least twenty-five per cent in real terms, arising from a resource windfall, an allocation formula, or an eligibility threshold — together with an admissible external outcome series spanning the change. Selection is deliberately indifferent to whether inversion is suspected. No expected verdict is recorded for any case. The protocol was calibrated on synthetic data before deposit. That calibration identified a defect in one preregistered falsifier: the redundancy check, which asks whether any single test predicts the aggregate verdict above ninety per cent, returned 0.88 under the pure null, because the aggregate is dominated by the same heavily weighted tests being examined. Recomputed leave-one-out, the null falls to 0.40 and the check becomes informative. Calibration further establishes that thirty decided cases suffice at per-test accuracy of 0.70, that forty are required at 0.55, and that an indeterminate rate near one quarter should be expected. It also establishes that the redundancy check will fire if per-test accuracy exceeds approximately 0.80, and the meaning of that outcome is preregistered rather than left to later interpretation. Keywords: preregistration · institutional capture · inversion · decay · exogenous variation · regression discontinuity · case selection · diagnostic validity
Author: Mattos, José Caetano de
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7204098967
Cites:
Cited by:
Related to: