TrustShift

A mechanism-aware audit of machine-learning deployment shift.

TrustShift is a single audit protocol that measures four axes of model trustworthiness — discrimination, calibration, operating-point performance, and subgroup reliability — before and after a distribution shift, attempts a staged, largely label-free mechanistic diagnosis of the shift (label / covariate / concept / mixed), and reports inconclusive honestly when the evidence does not support a firm label. Applied identically to four deliberately dissimilar domains, it shows that scalar shift magnitude alone does not consistently explain which trustworthiness axis fails; whether shift mechanism specifically carries more information than magnitude is tested directly and not established by the current evidence.

The four deployment shifts

Domain	Source → Target	Diagnosis	What fails
Mental-health NLP	Kaggle → Reddit / Twitter	mixed	everything (AUC ↓0.28–0.39, calibration, subgroup gap)
Network intrusion	CIC-DDoS2019 → CICIDS2017	inconclusive	discrimination + calibration on unseen attack families (AUC 0.998 → 0.684)
Clinical risk	NHANES → BRFSS	inconclusive	calibration + operating point far more than discrimination (AUC ≈flat, ECE 0.17 → 0.28)
Mortgage lending (temporal)	HMDA 2020–21 → 2022–24	covariate	none of the four measured axes; ranking improves, operating point flat
Mortgage lending (geographic)	HMDA, 60% states → held-out states	inconclusive	operating point degrades (macro-F1 −0.03) while ranking still improves

Headline: a lending model under large measured covariate shift (domain-classifier AUC up to 0.80) improves in ranking while its macro-averaged operating-point performance stays essentially flat — even though positive-class F1 alone falls 0.06–0.08 — a case where which single metric you report changes the story. A text model at comparable domain-classifier magnitudes collapses on every axis. Scalar shift magnitude did not predict which of these would happen; three of the five targets above are diagnosed inconclusive by the current label-free/labeled probes, reported honestly rather than forced into a category.

Reproduce every number

Every figure and table in the paper regenerates from committed prediction files — no raw third-party data required. Regenerating from the raw source repos (clinical, NLP) additionally needs the sibling repos on PYTHONPATH; see Configuration below.

python -m venv venv
venv\Scripts\pip install -e .# installs the package + deps
venv\Scripts\python -m audit.engine # writes results/audit_{domain}.json
venv\Scripts\python -m synthesis.tables
venv\Scripts\python -m synthesis.figures # includes the Figure 1 audit-flow diagram

Results land in results/ (audit/diagnosis/remediation/meta JSONs, tables/*.csv, figures/*.png+.pdf). The manuscript sources are in paper/.

Configuration

All paths and constants live in config.py. The clinical/NLP source-repo paths and the security notebook path are only needed to regenerate predictions from raw source repos (most users won't need them — the released, committed predictions in results/ and on the Hugging Face Hub are sufficient to reproduce every paper number). Each defaults to this project's own development machine layout and can be overridden with an environment variable without editing code:

export TRUSTSHIFT_CLINICAL_REPO=/path/to/diabetes_prediction_project/federated
export TRUSTSHIFT_NLP_REPO=/path/to/mental-health-fairness-nlp
export TRUSTSHIFT_HMDA_FEATURES=/path/to/features_panel.parquet
export TRUSTSHIFT_HMDA_FEATURE_SETS=/path/to/feature_sets.json
export TRUSTSHIFT_DDOS_NOTEBOOK=/path/to/CrossDataset_DDoS_Colab.ipynb
export TRUSTSHIFT_FAIRSCOPE=/path/to/fairscope

Layout

config.py single source of paths + constants
schema.py standardized prediction schema + validator
domains/ one adapter per domain (clinical, nlp, lending, security)
audit/ engine.py (trustworthiness axes), diagnosis.py (shift diagnosis), remediation.py
synthesis/ tables.py, figures.py (incl. Figure 1), meta.py (cross-domain regressions)
results/ committed JSONs, tables, and 300-dpi figures (source of every paper number)
paper/ main.tex, references.bib, numbers.py (numbers ↔ results check)
tests/ schema + engine unit tests

The source datasets are public (NHANES, BRFSS, HMDA, GoEmotions, CIC-DDoS2019, CICIDS2017). Consistent with their licenses, this repository redistributes standardized model predictions and metadata, not raw third-party records.

What this benchmark does not claim

Per the manuscript's own frozen scientific position: TrustShift does not claim shift mechanism outpredicts shift magnitude (tested directly, not established by the current evidence — see the paper's Discussion); it is not a new dataset-shift taxonomy, and not a formal performance-decomposition estimator. Its contribution is the joint multi-axis audit, the staged diagnostic protocol that reports inconclusive as a first-class outcome, and the remediation analysis, across four real, non-synthetic cross-modality deployment shifts.

Status

Research code accompanying the TrustShift manuscript (in preparation). The prediction parquets that constitute the benchmark are hosted separately on the Hugging Face Hub.

Author

Rajveer Singh Pall.

License

MIT (code). Source datasets retain their original licenses.