RealityCheck · criterion-validity ledger

Demo panel — 60 orgs, 2 regimes

Do the scores predict independent reality? Preregistered hypotheses, circularity-guarded outcomes, reversed-lag placebos.

🔒 PRESPECIFIED — ledger hash matches lock of 2026-07-09T07:52:16
hypotheses 5 guard exclusions 1 generated 2026-07-09
3
confirmed
0
directional
1
null
0
reversed
1
guarded

The ledger

Spearman's ρ with two-sided permutation p-values; verdicts against the prespecified direction. A confirmed association whose reversed-lag placebo is also significant is downgraded — predicting the past is confounding, not foresight.

idhypothesisnρpplacebo (reversed lag)verdict
H1decoupling_score → Independent incident-tracker rate
expects positive · lag 1 · scope: all · evidence: third_party_administrative · exposure: low
decoupled governance leaves real exposure unmanaged
1200.6110.0010ρ=-0.158, p=0.2108 ✓clean✓ CONFIRMED
H2decoupling_score → Regulatory enforcement actions
expects positive · lag 1 · scope: regime=market · evidence: third_party_administrative · exposure: medium
market regimes rely on ex-post enforcement, so the gap surfaces there
600.7500.0010ρ=-0.280, p=0.1319 ✓clean✓ CONFIRMED
H3commitment_depth → Days from commitment to enacted control
expects negative · lag 0 · scope: all · evidence: behavioral_record · exposure: low
deeper verifiable commitment should shorten enactment
120-0.1880.0380✓ CONFIRMED
H4rhetoric_intensity → Independent incident-tracker rate
expects positive · lag 1 · scope: all · evidence: third_party_administrative · exposure: low
falsification probe: tone alone should predict nothing
1200.0330.7123ρ=0.153, p=0.2478 ✓clean○ NULL
H5decoupling_score → Own press-release volume
expects positive · lag 0 · scope: all · evidence: self_published · exposure: high
guard demonstration: self-published outcome must be excluded
evidence class 'self_published' — excluded by the circularity guard; an index must not predict documents with documents
🛡 GUARDED

Limitations — standing, non-suppressible

Endogeneity is not designed away. The circularity guard excludes self-published outcome evidence, but third-party outcomes — rankings, accreditation actions, enforcement, press coverage — are themselves shaped by the same institutional discourse the scores measure. A reversed-lag placebo does not touch this. Where declared, outcomes are stratified by exposure to the measured discourse; prediction concentrated in high-exposure outcomes should be read as discourse reverberation, not enactment.

Confirmed associations are correlational. The design supports forecasting claims, not causal ones, and a validity ledger is evidence about the index, not about any single organization.

Stratified by discourse exposure: low: confirmed 2, null 1 · medium: confirmed 1

Circularity guard. Outcomes with evidence class self_published are mechanically excluded (🛡 GUARDED): an index must not predict institutions' documents with institutions' documents. Eligible evidence: third-party/administrative records and behavioral records only.

Reading the ledger. The deliverable is the whole ledger, not one coefficient: how many prespecified associations landed, where, and under which governance regime. NULL rows are findings too — a stable, consequence-free say–do gap is a first-order policy result.

Generated by RealityCheck — phase 2 of the measurement family: gate the rubric, gate the judge, route the judgments, report it all, then check it against reality.