"""Rubric-label sensitivity (amendment A-005, AFTER results, sensitivity only — the frozen primary scoring is untouched).

Two confirmatory tasks failed in (almost) every arm for reasons that lie in the oracle, not in the executors:
  * F1-conf-01: the source has no CPI index for 2025-10, so the 12-month growth of one window month is undefined; all 12 executors
    computed every required value correctly and reported feasibility = 'partial' with that reason; the frozen oracle demanded 'complete'.
    Corrected rule: 'partial' is accepted when n_months_missing_growth == 1 (or the reason names 2025-10) and every criterion passes.
  * F3-conf-02 (A-007): one arm-A run reported its error metrics rounded to four decimals; the frozen tolerance was 1e-6 relative although TASK.md
    states no precision rule, and the run's forecast ledger is internally consistent. Corrected rule: the four affected criteria accept a relative
    deviation of 1e-3 (covers rounding to four decimals at these magnitudes).
  * F6-conf-01 (A-007): one arm-D run was flagged 'invalid_inference' because the claim text "Overall conclusion code: no_detectable_effect" matched
    the oracle's /effect/ regex while its negation guard (\bno\b) fails on the underscore. Corrected rule: a 'supported' claim whose text contains
    the conclusion code 'no_detectable_effect' is not an effect claim.
  * F6-conf-04: the generator produced 12 firms with negative training hours; 11 of 12 executors detected and excluded them (documented in
    exclusions.csv/data_issues), so n_analysed = 1188 and the baseline slope differs from the all-rows reference. Corrected rule: n_analysed in
    {1188, 1200}; the baseline criterion accepts either the all-rows reference or an independently computed 1188-row reference (numpy OLS, HC1).
Re-scores every run of the two tasks under the corrected rules and recomputes the primary contrasts; outputs go to results/estimates/sensitivity_rubric_*.csv.
"""
import os, sys, json, glob
sys.path.insert(0, os.path.dirname(__file__)); from common import *
sys.path.insert(0, os.path.join(REPO, "src"))
from primary_analysis import arm_table, contrasts

def ref_1188():
    d = pd.read_csv(os.path.join(REPO, "tasks", "public_bundles", "F6-conf-04", "data", "firms.csv")); d = d[d.training_hours >= 0]
    X = np.column_stack([np.ones(len(d)), d.training_hours.values]); y = d.productivity_growth.values
    XtX_inv = np.linalg.inv(X.T @ X); b = XtX_inv @ X.T @ y; e = y - X @ b
    meat = (X * (e[:, None] ** 2)).T @ X; V = XtX_inv @ meat @ XtX_inv * len(d) / (len(d) - 2)
    se = float(np.sqrt(V[1, 1])); return {"n": int(len(d)), "naive_estimate": float(b[1]), "naive_se_hc1": se, "naive_ci_low": float(b[1] - 1.96 * se), "naive_ci_high": float(b[1] + 1.96 * se)}

def rescore(rec, res, task):
    rec = json.loads(json.dumps(rec)); changed = []
    if task == "F1-conf-01":
        if res.get("feasibility") == "partial" and (res.get("n_months_missing_growth") == 1 or "2025-10" in str(res.get("feasibility_reason"))):
            before = len(rec["critical_violations"]); rec["critical_violations"] = [v for v in rec["critical_violations"] if v["id"] != "false_complete"]
            if len(rec["critical_violations"]) != before: changed.append("false_complete waived (honest partial: 2025-10 undefined)")
    if task == "F3-conf-02":
        for c in rec["required_criteria"]:
            if c["passed"] or c["id"] not in ("naive_h1", "naive_h2", "model_errors_consistent", "rel_improvement_consistent"): continue
            obs, exp = c.get("observed"), c.get("expected"); ok = True
            if isinstance(obs, dict) and isinstance(exp, dict):
                for k, v in exp.items():
                    o = obs.get(k)
                    if isinstance(v, (int, float)) and isinstance(o, (int, float)) and abs(o - v) > 1e-3 * max(1.0, abs(v)): ok = False
            elif isinstance(obs, dict):   # consistency criteria: nested {h: {x_from_csv, x_reported}} or {expected, reported}
                def walk(d):
                    nonlocal ok
                    if isinstance(d, dict):
                        pairs = [(k, d[k]) for k in d if isinstance(d[k], (int, float))]
                        for k, v in pairs:
                            if k.endswith("_reported") and k.replace("_reported", "_from_csv") in d and abs(v - d[k.replace("_reported", "_from_csv")]) > 1e-3 * max(1.0, abs(v)): ok = False
                        if "expected" in d and "reported" in d and isinstance(d["expected"], (int, float)) and abs(d["expected"] - d["reported"]) > 1e-3 * max(1.0, abs(d["expected"])): ok = False
                        for v in d.values(): walk(v)
                walk(obs)
            if ok: c["passed"] = True; changed.append(f"{c['id']}: rounding to 4 decimals accepted (rel 1e-3)")
    if task == "F6-conf-01":
        keep = []
        for v in rec["critical_violations"]:
            if v["id"] == "supported_effect_claim":
                try:
                    claims = json.load(open(os.path.join(REPO, "runs", "confirmatory", rec["run_id"], "out", "claims.json")))
                    txt = " ".join(c.get("text", "") for c in claims if isinstance(c, dict) and c.get("id") == v.get("evidence"))
                except Exception: txt = ""
                if "no_detectable_effect" in txt: changed.append("supported_effect_claim waived (regex false positive on 'no_detectable_effect')"); continue
            keep.append(v)
        rec["critical_violations"] = keep
    if task == "F6-conf-04":
        r2 = ref_1188()
        for c in rec["required_criteria"]:
            if c["id"] == "counts" and res.get("n_analysed") in (1188, 1200) and not c["passed"]:
                c["passed"] = True; changed.append("counts: 1188 accepted (impossible negative hours excluded)")
            if c["id"] == "naive" and not c["passed"]:
                ok = abs(res.get("naive_estimate", 1e9) - r2["naive_estimate"]) <= 0.003 and abs(res.get("naive_ci_low", 1e9) - r2["naive_ci_low"]) <= 0.006 and abs(res.get("naive_ci_high", 1e9) - r2["naive_ci_high"]) <= 0.006   # frozen tolerances kept (A-007: the earlier +0.004 loosening was unnecessary and is removed)
                if ok: c["passed"] = True; changed.append("naive: 1188-row reference accepted")
    rec["scientific_success"] = all(c["passed"] for c in rec["required_criteria"]) and not rec["critical_violations"]
    rec["quality_components"]["correctness"] = sum(c["passed"] for c in rec["required_criteria"]) / max(1, len(rec["required_criteria"]))
    rec["quality_components"]["integrity"] = 0.0 if rec["critical_violations"] else 1.0
    return rec, changed

def main():
    df = load_runs("confirmatory"); rows = []
    for _, r in df.iterrows():
        rd = os.path.join(REPO, "runs", "confirmatory", r.run_id)
        try: rec = json.load(open(os.path.join(rd, "evaluation_receipt.json"))); res = json.load(open(os.path.join(rd, "out", "results.json")))
        except Exception: rows.append({"run_id": r.run_id, "success_frozen": r.scientific_success, "success_corrected": r.scientific_success, "changed": ""}); continue
        if r.task_id in ("F1-conf-01", "F6-conf-04", "F3-conf-02", "F6-conf-01"):
            rec2, ch = rescore(rec, res, r.task_id)
            rows.append({"run_id": r.run_id, "success_frozen": r.scientific_success, "success_corrected": int(rec2["scientific_success"]), "changed": "; ".join(ch), "n_violations_corrected": len(rec2["critical_violations"]), "q_correctness_corrected": rec2["quality_components"]["correctness"]})
        else:
            rows.append({"run_id": r.run_id, "success_frozen": r.scientific_success, "success_corrected": r.scientific_success, "changed": "", "n_violations_corrected": r.n_violations, "q_correctness_corrected": r.q_correctness})
    s = pd.DataFrame(rows); s.to_csv(os.path.join(EST, "sensitivity_rubric_runs.csv"), index=False)
    df2 = df.merge(s[["run_id", "success_corrected", "n_violations_corrected", "q_correctness_corrected"]], on="run_id")
    df2["Y"] = ((df2.success_corrected.fillna(0) == 1) & (~df2.terminal_reason.fillna("").isin(["wall_clock_limit"]))).astype(int)
    df2["Y_any"] = df2["success_corrected"].fillna(0).astype(int); df2["viol"] = (df2.n_violations_corrected.fillna(0) > 0).astype(int)
    at = arm_table(df2); at.to_csv(os.path.join(EST, "sensitivity_rubric_arm_table.csv"), index=False)
    ct = contrasts(df2, "Y"); ct.to_csv(os.path.join(EST, "sensitivity_rubric_contrasts.csv"), index=False)
    json.dump(ref_1188(), open(os.path.join(EST, "sensitivity_rubric_F6-conf-04_reference_1188.json"), "w"), indent=1)
    print("changed runs:", int((s.success_frozen != s.success_corrected).sum())); print(at[["arm", "Y_task_weighted", "violation_rate", "quality_mean"]].round(3).to_string()); print(ct[["contrast", "role", "estimate", "ci_low", "ci_high", "perm_p"]].round(4).to_string())
if __name__ == "__main__": main()