r-bridge-execution

Running R from Python (the certified bridge)

Psychometric estimation uses R's field-standard packages (lavaan, mirt, MASS/difR) through a deterministic bridge — because Python's psychometrics libraries did not meet the certification bar, and because LLM-generated R is not certifiable.

The one rule

Generated code NEVER writes R code and NEVER calls Rscript, subprocess, or src.r_bridge directly. It calls the certified Python wrappers in analysis_helpers:

# Signatures (use EXACTLY these parameter names):
res_cfa = analysis_helpers.psy_cfa(items_df, model="F =~ it1 + it2 + it3")
res_inv = analysis_helpers.psy_invariance(items_df, group=group_series,
                                          model="F =~ it1 + it2 + it3")
res_grm = analysis_helpers.psy_grm(items_df)              # Likert GRM
res_dif = analysis_helpers.psy_dif(items_df, group=group_series)
res_ctt = analysis_helpers.psy_ctt(items_df)               # pure Python
res_om  = analysis_helpers.psy_omega(res_cfa)              # from CFA loadings

Each wrapper serializes the item matrix to JSON, runs the matching certified script in r_helpers/ via Rscript --vanilla, and returns a plain dict. Hand-rolling any of these models (in R OR Python) is a contract violation — the recipes are certification-locked (scripts/psychometric_gates.py: CFA loading error .016, GRM a-error .066, DIF hit rate 1.0 / false positives 0.0, invariance null honesty).

Data conventions at the bridge

• Items go in as NUMERIC columns; NCES sentinels must already be NaN.
• Items are never imputed. lavaan uses FIML (missing="fiml"); mirt handles NA natively; the ordinal-DIF models use listwise deletion and report the analyzed n. lavaan drops all-missing ROWS — expect the returned n to be ≤ the input rows.
• Likert items must be positive-integer categories (1..k) for GRM/DIF; reverse-coded items must be recoded BEFORE the call (per the registry's item bank flags).

Environment facts (Critic-checkable)

• R ≥ 4.4 resolved via config r_bridge.rscript_path → EDM_ARS_RSCRIPT env → common install dirs → PATH. Missing R fails loudly with remediation — never silently skip an analysis.
• R is NOT in the Docker sandbox image: psychometrics runs use the subprocess executor (sandbox.enabled: false).
• The bridge rejects any script path outside r_helpers/.