"""E-M2b · resultado de los jueces (DEFINICIONES_PREVIAS §6), stdlib, en el host: consenso prudente, aceptación, kappa, lo que EM2b pierde y la precisión de E0b re-estimada. Uso: python3 sql/60_jueces.py""" import csv import json from collections import Counter from pathlib import Path W = Path("/em2b") J = W / "jueces" clave = {r["id"]: r for r in csv.DictReader(open(J / "clave_jueces.csv", encoding="utf-8"), delimiter=";")} P = {r["par"]: r for r in csv.DictReader(open(W / "clave/parejas.csv", encoding="utf-8"), delimiter=";")} OK = {"SI", "DUDOSA", "NO"} RANGO = {"NO": 0, "DUDOSA": 1, "SI": 2} out = [] def p(*a): s = " ".join(str(x) for x in a) out.append(s) print(s) def leer(n): v, malas = {}, [] for parte in (1, 2): for r in csv.DictReader(open(J / f"juez_{n}_parte{parte}.csv", encoding="utf-8"), delimiter=";"): ver = (r.get("veredicto") or "").strip().upper().replace("Í", "I") if r.get("id") in clave and ver in OK and r["id"] not in v: v[r["id"]] = (ver, (r.get("motivo") or "").strip()) else: malas.append(r) return v, malas j1, m1 = leer(1) j2, m2 = leer(2) p(f"Juez 1 (Opus): {len(j1)} veredictos válidos, {len(m1)} filas malas; juez 2 (Fable): {len(j2)} válidos, {len(m2)} malas; " f"faltan J1 {len(set(clave) - set(j1))}, J2 {len(set(clave) - set(j2))}") ids = sorted(clave) cons = {} for i in ids: vs = [x[0] for x in (j1.get(i), j2.get(i)) if x] cons[i] = min(vs, key=lambda v: RANGO[v]) if len(vs) == 2 else ("NO" if not vs else min(vs + ["DUDOSA"], key=lambda v: RANGO[v])) def kappa(a, b): n = len(a); po = sum(x == y for x, y in zip(a, b)) / n ca, cb = Counter(a), Counter(b) pe = sum(ca[c] * cb[c] for c in OK) / n / n return po, (po - pe) / (1 - pe) ambos = [i for i in ids if i in j1 and i in j2] po, k = kappa([j1[i][0] for i in ambos], [j2[i][0] for i in ambos]) p(f"\n## Acuerdo entre jueces: {100 * po:.1f} %, kappa {k:.3f} ({len(ambos)} parejas)") for nom, jj in (("Juez 1 (Opus)", j1), ("Juez 2 (Fable)", j2)): c = Counter(jj[i][0] for i in ids if i in jj) p(f"- {nom}: SÍ {c['SI']}, DUDOSA {c['DUDOSA']}, NO {c['NO']}") def tabla(titulo, sel): c = Counter(cons[i] for i in sel) n = len(sel) p(f"| {titulo} | {n} | {c['SI']} ({100 * c['SI'] / max(1, n):.1f} %) | {c['DUDOSA']} ({100 * c['DUDOSA'] / max(1, n):.1f} %) | " f"{c['NO']} ({100 * c['NO'] / max(1, n):.1f} %) |") return c, n A = {} B = {} for f, d in (("A", A), ("B", B)): for l in sorted((W / f"pasada{f}").glob("lote_*.jsonl")): for x in l.read_text(encoding="utf-8").splitlines(): if x.strip(): o = json.loads(x); d[o["par"]] = o alta = lambda i: A[clave[i]["par"]]["confianza"] == "alta" and B[clave[i]["par"]]["confianza"] == "alta" # noqa: E731 p("\n## Consenso prudente (NO < DUDOSA < SÍ)") p("| Grupo | n | SÍ | DUDOSA | NO |") p("|---|---:|---:|---:|---:|") em = [i for i in ids if clave[i]["grupo"] == "em2b"] de = [i for i in ids if clave[i]["grupo"] == "descartada"] c_em, n_em = tabla("**Parejas EM2b**", em) tabla("… que además son E0b", [i for i in em if clave[i]["e0b"] == "1"]) tabla("… que no son E0b (solo E0)", [i for i in em if clave[i]["e0b"] == "0"]) tabla("… con confianza alta en las dos pasadas", [i for i in em if alta(i)]) tabla("… sin confianza alta en alguna pasada", [i for i in em if not alta(i)]) c_de, n_de = tabla("E0b que EM2b descarta", de) no_ok = c_em["NO"] <= 0.10 * n_em si_ok = c_em["SI"] >= 0.60 * n_em p(f"\n**Aceptación (precisión)**: NO {c_em['NO']}/{n_em} = {100 * c_em['NO'] / n_em:.1f} % (≤ 10 %: {'sí' if no_ok else 'NO'}) · " f"SÍ {c_em['SI']}/{n_em} = {100 * c_em['SI'] / n_em:.1f} % (≥ 60 %: {'sí' if si_ok else 'NO'}) → {'CUMPLE' if no_ok and si_ok else 'NO CUMPLE'}") # Con la opinión de un solo juez (sensibilidad, informativo) for nom, jj in (("solo juez 1", j1), ("solo juez 2", j2)): c = Counter(jj[i][0] for i in em if i in jj) p(f"- {nom}: SÍ {c['SI']}, DUDOSA {c['DUDOSA']}, NO {c['NO']} de {len(em)}") # Precisión de E0b re-estimada con estos jueces (informativo): pesos por parejas E0b elegibles viejas = set() EM2 = Path("/em2/cotejo") for f in ("jueces_clave.csv", "origen_e0.csv", "origen_e0b.csv", "a2_e0.csv"): for r in csv.DictReader(open(EM2 / f, encoding="utf-8"), delimiter=";"): if "serie_m1r" in r and "licitacion" in r: viejas.add((r["serie_m1r"], r["licitacion"])) em2b_par = {k for k in P if k in A and k in B and A[k]["etiqueta"] == B[k]["etiqueta"] == "misma_necesidad"} elig = [k for k in P if P[k]["tipo_unidad"] == "serie" and P[k]["de_2023"] == "1" and P[k]["en_lista"] == "1" and not P[k]["ancla_id"] and (P[k]["unidad"], P[k]["licitacion"]) not in viejas and P[k]["e0b"] == "1"] w_pos = sum(k in em2b_par for k in elig) / len(elig) s_pos = [i for i in em if clave[i]["e0b"] == "1"] r_pos = Counter(cons[i] for i in s_pos) r_neg = c_de est = {v: w_pos * r_pos[v] / max(1, len(s_pos)) + (1 - w_pos) * r_neg[v] / max(1, n_de) for v in ("SI", "DUDOSA", "NO")} p(f"\n## Precisión de E0b re-estimada con estos jueces (informativo)") p(f"- parejas E0b elegibles: {len(elig)}; de ellas EM2b: {sum(k in em2b_par for k in elig)} (peso {w_pos:.3f})") p(f"- P(SÍ | E0b) ≈ {100 * est['SI']:.1f} %, DUDOSA {100 * est['DUDOSA']:.1f} %, NO {100 * est['NO']:.1f} % " f"(E-M2, con los jueces de entonces y una pareja por serie de 2019-2024: 48,3 / 21,7 / 30,0)") # Detalle de las EM2b que no son SÍ p("\n## Parejas EM2b con consenso DUDOSA o NO") for i in em: if cons[i] != "SI": p(f"- {i} · {cons[i]} · J1 {j1.get(i, ('—', ''))[0]}: {j1.get(i, ('', ''))[1][:140]} · J2 {j2.get(i, ('—', ''))[0]}: {j2.get(i, ('', ''))[1][:140]}") p("\n## Descartadas con consenso SÍ") for i in de: if cons[i] == "SI": k = clave[i]["par"] p(f"- {i} · A {A[k]['etiqueta']} ({A[k]['motivo'][:90]}) · B {B[k]['etiqueta']} ({B[k]['motivo'][:90]})") with open(J / "jueces_resultado.csv", "w", newline="", encoding="utf-8") as fh: w = csv.writer(fh, delimiter=";") w.writerow(["id", "grupo", "par", "e0b", "juez_1", "juez_2", "consenso", "motivo_1", "motivo_2", "A", "B", "conf_alta_ambas"]) for i in ids: k = clave[i]["par"] w.writerow([i, clave[i]["grupo"], k, clave[i]["e0b"], j1.get(i, ("", ""))[0], j2.get(i, ("", ""))[0], cons[i], j1.get(i, ("", ""))[1], j2.get(i, ("", ""))[1], A[k]["etiqueta"], B[k]["etiqueta"], int(alta(i))]) (W / "salidas/60_jueces.md").write_text("# E-M2b · jueces\n\n" + "\n".join(out) + "\n", encoding="utf-8")