Supplementary Artifact for "Default Zero-Shot LLM APIs Over-Flag Indian Financial SMS: A Policy-Aligned Benchmark and Template-Level Pre-Deployment Audit"

This record contains the GroundLM camera-ready supplementary analyses for “Default Zero-Shot LLM APIs Over-Flag Indian Financial SMS: A Policy-Aligned Benchmark and Template-Level Pre-Deployment Audit,” accepted at the GroundLM Workshop through the ARR commitment track.

Included file:
scamshield-groundlm-2026-supplementary.zip

The supplement contains prompt-ladder predictions and analysis, consensus analysis, regenerated label-policy sensitivity, and aggregate negative-label, human-annotation, structural-cluster, and calibration results. It includes exact prompt definitions and runnable analysis scripts. It contains no raw SMS text or verbatim provider responses.

Companion artifact:
The original checkpoints, evaluation data, and historical source files are available separately at:
https://doi.org/10.5281/zenodo.20480151

The original archive is not included in this record. Both records should be consulted together.

Reproducibility limitations:
The companion artifact lacks required training modules, the official IIITD SMSAssassin distribution, and an automated reconstruction script. It does not reproduce training from scratch or full checkpoint re-inference as packaged. Its 169 IIITD-sourced evaluation messages are represented by hashes; the original messages must be obtained separately and rejoined locally.

This GroundLM supplement supports recomputation of selected results from frozen parsed predictions. Analyses released only as aggregates cannot be independently reconstructed from this supplement alone. Hosted API behavior may change.

Rights:
Apache-2.0 applies to original project code and documentation. Any third-party material retains its upstream licenses or access terms.

Clarification: The earlier artifact description overstated its contents and reproducibility. The original archive does not include the official IIITD dataset ZIP or a complete runnable training workflow. This GroundLM release provides additional analyses, with the original models and evaluation materials available through the companion link.