Reproducible drift and trust-boundary evidence

Modelstamp packages its controlled dependency-drift and trust-boundary evaluations as one rerunnable Python 3.11 protocol. A clean checkout and one command reproduce 14 pinned drift cases and eight trust-boundary cases:

python3.11 experiments/reproducible_evidence/run_evidence.py

The protocol produces an aggregate JSON result, a Markdown summary, per-scenario observations, persisted model artifacts, and manifests. The scenario catalog is hashed into the result to connect observations to the predeclared package pins and expected outcomes.

Evidence covered

The drift suite includes sklearn patch and minor changes, LightGBM and its embedded sklearn dependency, XGBoost, CatBoost, NumPy, SciPy, joblib, identical-environment controls, and unrelated-package noise controls. Each case fits a real estimator in environment A and checks it without deserialization in environment B.

The trust-boundary suite covers artifact tampering, an unsigned pair swap, manifest hash and identity edits, unsigned replacement of a signed pair, untrusted-key replacement, shared-key-holder replacement, and replay of an older valid signed pair. Accepted cases are documented limitations, not security successes.

See the experiment README for prerequisites, selective-run commands, acceptance criteria, and the interpretation boundary.