SAS R2.1-b, Function-Level Extension — Reference-Free Validation of Python API Calls via Sandboxed Introspection

Gonzalo Emir Durante (Independent Researcher; Symbiotic Autoprotection System)
Registry: TAD EX-2026-18792778 (Argentina)

This deposit extends R2.1-b's import-level phase (DOI 10.5281/zenodo.21420392) to function/method-level reference-free validation: given that a Python module is real, does a specific call against it actually exist? A knowledge base of real callable API surface for five widely used libraries (numpy, pandas, matplotlib.pyplot, requests, json) was built via sandboxed live introspection (isolated subprocess, no network, timeout-bounded), and an AST-based validator checks calls against it without ever executing the candidate code being evaluated.

Validated against the public replication package of arXiv:2601.19106 (accepted at FORGE '26), the method reaches precision 100.00%, recall 98.04%, F1 0.9901 on a scope-aligned 140-row subset, exceeding that paper's own reported recall (87.6%) and F1 (0.934) on comparable ground.

Reaching that result required correcting a data-quality defect found during validation: the external dataset's own reference code used pandas.groupby(...) as a module-level call in five rows — independently confirmed to not exist (groupby is exclusively a DataFrame/Series instance method). This is the third documented instance in this research line of a third-party benchmark's ground truth containing a verifiable defect, following similar findings in R2.1 (MBPP) and R2.1-b's import-level phase (Slopsquatting).

Disclosed limitations: no instance-method resolution (calls like df.to_csv(...) require type inference, out of scope here); "missing import" and "contextual misnaming" categories are explicitly unaddressed, as no existence-based method can resolve them in principle.

Contents: full code (knowledge-base builder, AST validator, evaluation pipeline), generated knowledge base, per-row-ID results manifest (no third-party code content redistributed), PDF paper, README with full scope disclaimer.

Registry: TAD EX-2026-18792778 (Argentina)
DOI (R2.1-b, import-level phase): 10.5281/zenodo.21420392
DOI (R2.1): 10.5281/zenodo.21365707
DOI (SAS Standard): 10.5281/zenodo.19702379

Timestamp proof: RFC 3161 token (FreeTSA), included as .tsr