The cross-language experiment extends the benchmark beyond Python to validate that scorecard-anchored prompting generalizes. Six repos, three policies:
- JS/TS:
express(Mocha + node:assert + supertest),jsonwebtoken(Mocha + Chai + sinon),zod(Vitest, monorepo). - Go:
chi(httptest router),gjson(pure parse + fuzz),golang-jwt(crypto signing). - Policies:
oneshot,iter2,iter20(same semantics as the quality-driven stage).
The prompt parts (common_header, quality_contract,
quality_scorecard, and the three policy bodies) are language-
parameterized; scripts/setup_cross_language.py substitutes per-repo commands
({RUN_CMD}, {SCORE_CMD}, {DELETE_CMD}, …) and materializes a
.rex_prompt.md into each wt-r3-<policy> worktree.
The defining difference from the quality-driven prompts: self-scoring is a single
score.py --lang <L> --baseline … call (the multi-language scorer),
not the Python-specific grep/wc recipes.