Replication package: Who Gets the House? Financing-Associated Price Gaps in New York City

Replication package and manuscript for Who Gets the House? Financing-Associated Price Gaps in New York City, by Pablo Loschi, independent researcher, Berlin. Version v2.3 (documentation correction). The manuscript and the replication folder are byte-identical to v2.2 (https://doi.org/10.5281/zenodo.22921254; local release identifier WGTH-2026-09-23-V2.2, which the manuscript and code still carry). Earlier version v2.1: https://doi.org/10.5281/zenodo.22915008. Concept DOI (always resolves to the latest version): https://doi.org/10.5281/zenodo.22421850. The companion paper on the mortgage recording tax and transfer-tax thresholds, The Shape of the Tax, has its own deposit at https://doi.org/10.5281/zenodo.22880953 and is not the parent of this version. The paper estimates financing-associated repeat-sales price contrasts in New York City administrative records, fitting financing-switch indicators jointly with permit restrictions, institutional-name exclusions, buyer and seller controls and neighbourhood-by-time effects. For houses the least-squares contrast is 9.25 log points (999-refit parcel-bootstrap interval 7.17–11.35). Robust estimators using conventional tuning constants on the same draws give 5.15–7.66 log points, every interval excluding zero (Huber 6.73 [5.32, 8.76]). Condominiums under common controls show a contrast that is much smaller and sensitive to the estimator (least squares 0.77 with 95% CI [−0.01, 1.56], which includes zero; Huber −0.49 [−1.00, −0.03]). A seller indifference condition shows that execution risk easily accommodates the condominium contrast, but under model assumptions at a 10% failure probability, an 87% loss on failed sales (63% at Huber) would be required to reconcile the house contrast alone. Rather than an empirical bound or mechanism share, this calibration illustrates the extreme demands placed on execution risk by the house point estimates. None of these estimates is a causal effect of financing, execution certainty or mortgage taxation. Version 2.3 corrects one number in the package documentation and changes nothing else. The v2.2 record description and five documentation files (START_HERE, CHANGELOG_V2_2, CONTRIBUTION_AND_REFEREE_BRIEF, VALIDATION_REPORT, VALIDATION.json) gave the condominium least-squares interval as [−0.12, 1.66], a value no output supports. The manuscript was always correct: [−0.01, 1.56] (the robust-estimator bootstrap on the same 999 draws gives [−0.02, 1.55]). Version 2.2 had conditioned the execution-risk calculation as an illustrative calibration check, described the condominium contrast as much smaller and sensitive to the estimator, adopted the term 'robust estimators using conventional tuning constants', and removed speculative international references; those changes are unchanged here. All empirical results reproduce identically. The package includes archived inputs, preserved public extracts, code, results, manuscript, submission documents and a blinded instrument-review queue. From the ZIP's replication folder, install Python 3.12 and Tectonic, run bash setup.sh, then execute env -u PYTHONPATH -u PYTHONHOME PYTHONNOUSERSITE=1 .venv/bin/python code/reproduce.py (about 10 minutes on eight cores; the robust-estimator bootstrap dominates). The analysis uses supplied inputs without live API calls. --from-results only rebuilds exhibits and documents. The reproduction scope is archived analysis inputs to paper. The original full raw DOF/ACRIS panel has not been reconstructed. Instrument classification remains incompletely validated: the four-case AI-assisted pilot is not a population audit, and the blinded 53-pair queue has no completed independent human reviews. No population classification-error rate or fully validated financing measure is claimed. Author code is MIT-licensed; author-created documentation and compilation contributions are CC-BY-4.0. City, Freddie Mac and StreetEasy source materials retain their underlying terms, as described in the README and third-party notices. This is a working-paper release, not journal acceptance or a completed journal submission.