Function is_error

scripts/rescore_clean_lambda0.py:23–25

(r)

        "deepseek-v4-flash_strict_rejudge.json"}

def is_error(r):
    # empty response with no genuine refusal text -> treat as query error, exclude
    return (not (r.get("model_response") or "").strip()) and r.get("verdict") == "REFUSAL"

def score(records):
    ts = {t: {"total":0,"correct":0,"wrong":0,"refusal":0} for t in TIERS}