analysis/code/diag.py

"""§6–§7: diagnosi per domanda (tutte le politiche), con attenzione alle discordanti D/B a 128k.
Uso: python3 analysis/diag.py  -> analysis/diag.json + analysis/diag.md

Convenzione del banco (verificata su tutte le 100 domande, vedi check_convention): il "tool result for round N" è
l'uscita dello strumento che PRECEDE il messaggio utente "Round N: ..." (indice = idx(Round N) − 2). La lettura
alternativa naturale, l'uscita che SEGUE "Round N" (idx(Round N) + 2), dà un "gold alternativo".
"""
import json
import re
import sqlite3
from collections import Counter, defaultdict
from pathlib import Path

import load as LD

OUT = Path(__file__).resolve().parent
rows = LD.load_results()
by = {(r["L"], r["pol"], r["qid"]): r for r in rows}


def ids_of(content):
    try:
        v = json.loads(content)
        if isinstance(v, dict) and "intersection" in v:
            return list(v["intersection"])
    except (ValueError, TypeError):
        pass
    return re.findall(r"\b\d{4}\b", content or "")


def round_user_idx(M, r):
    for n, m in enumerate(M):
        if m["role"] == "user" and isinstance(m["content"], str) and m["content"].startswith("Round %d:" % r):
            return n
    return None


def evidence(s):
    """Indici dei messaggi-prova (convenzione del banco) e alternativi, gold e gold alternativo."""
    M = s["messages"]
    t = s.get("pokemon_name") or s.get("target_name")
    rounds = [s["i_round"], s["j_round"]] if "i_round" in s else [s["round"]]
    ev, alt = [], []
    for r in rounds:
        k = round_user_idx(M, r)
        ev.append(k - 2)
        alt.append(k + 2 if k + 2 < len(M) and M[k + 2]["role"] == "tool" else None)
    if "i_round" in s:
        g = all(t in ids_of(M[i]["content"]) for i in ev)
        ga = all(i is not None and t in ids_of(M[i]["content"]) for i in alt)
    else:
        g = ids_of(M[ev[0]]["content"]).count(t)
        ga = ids_of(M[alt[0]]["content"]).count(t) if alt[0] is not None else None
    assert g == s["answer"], (s["id"], g, s["answer"])
    return {"target": t, "rounds": rounds, "ev": ev, "alt": alt, "gold": s["answer"], "gold_alt": ga,
            "n_msgs": len(M)}


def lcs_run(a, b):
    """Lunghezza della più lunga sequenza contigua comune (a = prova, b = testo)."""
    if not a or not b:
        return 0
    best = 0
    pos = defaultdict(list)
    for j, x in enumerate(b):
        pos[x].append(j)
    prev = {}
    for i, x in enumerate(a):
        cur = {}
        for j in pos.get(x, ()):
            cur[j] = prev.get(j - 1, 0) + 1
            if cur[j] > best:
                best = cur[j]
        prev = cur
    return best


def pred_norm(r):
    p = r["pred"]
    if isinstance(p, bool):
        return p
    if isinstance(p, str) and p.strip().lower() in ("yes", "true", "no", "false"):
        return p.strip().lower() in ("yes", "true")
    try:
        return int(p)
    except (TypeError, ValueError):
        return p


def _arg(rc, k):
    try:
        return json.loads(rc.get("args") or "{}").get(k)
    except ValueError:
        return None


def _kind(rc):
    try:
        a = json.loads(rc.get("args") or "{}")
    except ValueError:
        return "?"
    if "query" in a:
        return "id+query"
    if "offset" in a:
        return "id+offset"
    return "id+" + str(a.get("mode", "intero"))


def d_diag(r, e, sample):
    """Diagnostica D dal giornale e dall'archivio del proxy."""
    d = Path(r["dir"]) / "proxy"
    J = [json.loads(l) for l in open(d / "journal.jsonl", errors="replace") if l.strip()]
    ec = LD.load_engine_calls(r["dir"])
    con = sqlite3.connect("file:%s?mode=ro" % (d / "archive.sqlite"), uri=True)
    segs = con.execute("select seg, cut_idx, notes_msg from segments order by seg").fetchall()
    cut = segs[-1][1] if segs else 0
    notes = segs[-1][2] if segs else ""
    masked = {i for (i,) in con.execute("select idx from masks")}
    rid2idx = {rid: i for rid, i in con.execute("select rid, idx from archive")}
    arch_len = {i: n for i, n in con.execute("select idx, length(content) from archive where role='tool'")}
    inserts = []
    for (msgs,) in con.execute("select msgs from inserts"):
        for m in json.loads(msgs):
            if m.get("role") == "tool":
                inserts.append(m.get("content") or "")
    con.close()
    guards = [x for x in J if x.get("event") == "window_guard"]
    recalls = [x for x in J if x.get("event") == "recall"]
    notes_ev = [x for x in J if x.get("event") == "notes"]
    M = sample["messages"]
    out = {"switches": len(segs), "cut": cut, "rounds": sum(1 for c in ec if c["kind"] == "round"),
           "final_calls": sum(1 for c in ec if c["kind"] == "final"),
           "recall_n": len(recalls), "recall_args": [x.get("args") for x in recalls],
           "close": [(g.get("action"), g.get("reason"), g.get("where")) for g in guards],
           "removed_final": sum(len(g.get("removed_final") or []) for g in guards),
           "removed_search": sum(len(g.get("removed_search") or []) for g in guards),
           "notes_finish": [x.get("finish") for x in notes_ev],
           "final_prompt": max((c["prompt_tokens"] for c in ec if c["kind"] in ("final", "round")), default=0)}
    t = e["target"]
    for key, idxs in (("ev", e["ev"]), ("alt", e["alt"])):
        st = []
        for i in idxs:
            if i is None:
                st.append("n/a")
                continue
            ev_ids = ids_of(M[i]["content"])
            live = i >= cut and i not in masked
            rec = max((lcs_run(ev_ids, ids_of(x)) for x in inserts), default=0)
            res_i = [x for rc in recalls for x in (rc.get("results") or []) if x.get("idx") == i]
            by_id = [rc for rc in recalls if rid2idx.get(_arg(rc, "id")) == i]
            if live:
                st.append("vivo")
            elif rec >= max(1, int(0.9 * len(ev_ids))):
                st.append("letto intero")
            elif by_id or res_i:
                st.append("toccato (%s)" % ", ".join(sorted({_kind(rc) for rc in by_id} |
                                                         ({"risultato"} if res_i else set()))))
            else:
                st.append("assente")
        out[key + "_in_final"] = st
    # il modello ha cercato i round giusti? (recall che toccano gli indici-prova o i loro vicini)
    hit = set()
    for x in recalls:
        for res in x.get("results") or []:
            hit.add(res.get("idx"))
    out["touched_ev"] = [i in hit for i in e["ev"]]
    out["touched_alt"] = [i in hit for i in e["alt"] if i is not None]
    out["target_in_notes"] = t in (notes or "")
    out["reasoning_head"] = (json.loads((Path(r["dir"]) / "result.json").read_text()).get("reasoning") or "")[-600:]
    return out


def b_diag(r, e, sample):
    res = json.loads((Path(r["dir"]) / "result.json").read_text())
    evs = res.get("policy_events") or []
    comps = [x for x in evs if x.get("event") == "compact"]
    fc = [x for x in evs if x.get("event") == "final_context"]
    n_hist = e["n_msgs"] - 1                      # messaggi di storia (indici 1..n_hist)
    raw = (fc[-1]["messages"] - 3) if (fc and comps) else n_hist
    first_raw = n_hist - raw + 1
    summ = comps[-1]["summary"] if comps else ""
    t = e["target"]
    st = []
    for i in e["ev"]:
        if i >= first_raw:
            st.append("vivo")
        else:
            ev_ids = ids_of(sample["messages"][i]["content"])
            run = lcs_run(ev_ids, ids_of(summ))
            st.append("nel riassunto intero" if run >= len(ev_ids) else
                      ("nel riassunto in parte (%d/%d)" % (run, len(ev_ids)) if run else "assente"))
    return {"compactions": len(comps), "first_raw_idx": first_raw, "ev_in_final": st,
            "target_in_summary": t in summ,
            "rounds_in_summary": [("Round %d" % k) in summ or ("round %d" % k) in summ.lower() for k in e["rounds"]],
            "summary_finish": [x.get("finish_reason") for x in comps][-1:] if comps else [],
            "summary_tail": summ[-500:]}


def main():
    diag = {}
    for L in LD.LENGTHS:
        S = {**LD.samples(L, "CF"), **LD.samples(L, "FD")}
        for q in sorted({r["qid"] for r in rows if r["L"] == L}):
            s = S[q]
            e = evidence(s)
            rec = {"L": L, "qid": q, "qt": LD.qtype_short(s["question_type"]), **{k: v for k, v in e.items()}}
            rec["neg"] = (e["gold"] is False) or (e["gold"] == 0)
            rec["conv_distinguishes"] = e["gold"] != e["gold_alt"]
            for pol in LD.POLS:
                r = by[(L, pol, q)]
                p = pred_norm(r)
                rec[pol] = {"ok": int(r["off"] >= 1), "pred": p if not isinstance(p, str) else p[:40],
                            "trunc": r["trunc"], "matches_alt": (p == e["gold_alt"]) and e["gold"] != e["gold_alt"],
                            "neg_answer": p in (False, 0)}
            rec["D_diag"] = d_diag(by[(L, "D", q)], e, s)
            rec["B_diag"] = b_diag(by[(L, "B", q)], e, s)
            diag[f"{L}/{q}"] = rec
    (OUT / "diag.json").write_text(json.dumps(diag, indent=1, ensure_ascii=False, default=str))
    return diag


if __name__ == "__main__":
    d = main()
    print(len(d), "domande")