Summary
Current paraphrase test (Scenario 4) runs on English only. We need to know how much of the 10/10 grade is English-specific vs architectural before making general claims.
What to do
- Port Scenario 4's 5-query rubric to at least one other language (Czech is a natural choice — we already have Czech fixtures from the cross-language concept binding test).
- Use the SAME coverage logic (no per-language stopword lists added).
- Record the grade. If the grade drops significantly, that quantifies the English-preprocessor dependency flagged in M1's coverage issue.
Acceptance
tests/honest_agent_paraphrase_cs.rs (or similar) exists.- Results logged in
docs/ with the English-vs-non-English grade delta.
Links
tests/honest_agent_paraphrase.rs- M1 issue: "Coverage signal has a hidden English prior"
Summary
Current paraphrase test (Scenario 4) runs on English only. We need to know how much of the 10/10 grade is English-specific vs architectural before making general claims.
What to do
Acceptance
tests/honest_agent_paraphrase_cs.rs(or similar) exists.docs/with the English-vs-non-English grade delta.Links
tests/honest_agent_paraphrase.rs