Code and data for evaluating passive fingerprints for tabular foundation models under context changes and distillation

Frozen numerical-reproduction data and analysis code for a bounded evaluation of passive model fingerprints under support-context changes and student distillation. The data archive contains the experimental support contexts and row allocations, probe inputs, recorded model responses, student training targets and raw regression predictions, frozen calibration thresholds, training warnings, checkpoint source identifiers, and the 22 numerical tables used in the manuscript. The code archive recomputes these tables and all eight figures on CPU and provides the ModelDiff search implementation and an optional single-student refit entry point. It does not require foundation-model weights for numerical reproduction. The scope comprises nine development splits, six confirmation splits, three controlled context-replacement tasks, and 120 students divided into separate calibration and test cohorts. Calibration thresholds were frozen before test execution. Results do not establish ownership, licensing status, independent pretraining lineages, or a population false-positive-rate guarantee. Research data are licensed under CC BY 4.0. Source code is licensed under MIT, as specified in the code archive. UCI dataset attributions and upstream model sources are recorded in data/sources.json. Third-party pretrained weights are not redistributed. Search oracle caches, redundant transforms, internal documentation and submission materials are omitted. Code version 1.0.0 is frozen at Git commit 0b84886f03792d813a021741f2ea194edd945025 (tag v1.0.0). The code ZIP was generated from that Git tag and every file was verified against the public Git blobs. The data archive SHA-256 is 943f4d090ae1efce7a1cbeb59b86d25c33c7183d8877be3d5a711834c115ff63.