Invariance Audit: A Registry and Pre-Registered Protocol for Testing Whether Policy-Deployed Economic and Financial Models Hold Out of Sample

A structural (causal) relationship is invariant across samples, time, and intervention regimes; a fitted correlation need not be. Economics and finance routinely assign a structural reading to a fitted correlation — a coefficient, a threshold, a scaling or power-law pattern — and deploy it as a rule: regulatory thresholds, capital requirements, fiscal and monetary guidance. The condition that licenses such use is invariance, formalised by the Lucas critique and by super-exogeneity (Engle, Hendry & Richard) and, in modern causal inference, by invariant prediction. When a relationship that fails invariance is used to set policy, the cost of the failure falls on the public, not the modeller. The Invariance Audit catalogues the population of such models and pre-registers a protocol for testing whether they survive out of sample and out of regime, against pre-committed benchmarks, with the burden on the claimant: a break falsifies a claim unless its regime boundary was pre-specified. This deposit comprises the model registry (schema, entries, validation), the claim-gap coding manual and its inter-coder reliability protocol, the pre-registration and analysis plan, and a data-management plan. No study outcomes are examined before the registry and protocol are frozen; the DOI of this deposit is the time-stamped record of that freeze. Source code is released under the MIT License; the registry and text under CC-BY-4.0.