tempscripts/story_remediation/unbundle/analyze_all.py

"""Analyze the full unbundle C3 run vs baseline.

Reports:
  - aggregate catch rate across all rows vs baseline 64.7%
  - T1-slot flips (same trie position as baseline): fixed / regressed / net
  - tail-turn catch rate, broken down by tail length and by position
  - the 512 untouched passthrough rows as a sanity control
Writes RESULT_all.md.
Run: python -u temp/story_remediation/unbundle/analyze_all.py
"""
from __future__ import annotations
import json
from collections import defaultdict, Counter
from pathlib import Path

HERE = Path(__file__).resolve().parent
ROOT = HERE.parents[2]
OUT = HERE / "out"
RES = OUT / "C3_all.jsonl"
BASELINE = ROOT / "datasetreview" / "results" / "new" / "C3.jsonl"
N100 = ROOT / "distractor_generation_2" / "datasets" / "n100" / "trajectories.jsonl"
MD = HERE / "RESULT_all.md"


def caught_of(d):
    if "caught" in d and d["caught"] is not None:
        return d["caught"]
    return d.get("guess") == d.get("answer_key")


def main():
    base = {}
    for l in open(BASELINE, encoding="utf-8"):
        if l.strip():
            d = json.loads(l); base[d["item_id"]] = d
    base_caught = {k: caught_of(v) for k, v in base.items()}
    nb = len(base); bc = sum(base_caught.values())

    res = [json.loads(l) for l in open(RES, encoding="utf-8") if l.strip()]
    ok = [r for r in res if r.get("error") is None and "caught" in r]
    errs = [r for r in res if r not in ok]

    # tail length per orig from n100
    orig = {json.loads(l)["example_id"]: json.loads(l)
            for l in open(N100, encoding="utf-8") if l.strip()}
    tail_len = {e: len(r["calls"]) - (r["metadata"]["anchor_depth"] + 1)
                for e, r in orig.items()}

    n = len(ok); c = sum(1 for r in ok if r["caught"])
    t1 = [r for r in ok if r.get("role") == "turn1"]
    tail = [r for r in ok if r.get("role") and r.get("role") != "turn1"]
    passth = [r for r in ok if not r.get("role")]

    # T1 flips vs baseline (same eid, same trie position)
    fixed = regress = kept_c = kept_f = 0
    for r in t1:
        bcaught = base_caught.get(r["item_id"])
        if bcaught is None:
            continue
        if bcaught and not r["caught"]:
            fixed += 1
        elif not bcaught and r["caught"]:
            regress += 1
        elif bcaught and r["caught"]:
            kept_c += 1
        else:
            kept_f += 1
    t1_caught = sum(1 for r in t1 if r["caught"])

    # tail by length and position
    tail_by_len = defaultdict(lambda: [0, 0])
    tail_by_pos = defaultdict(lambda: [0, 0])
    for r in tail:
        oe = r.get("orig_eid"); tl = tail_len.get(oe, 0)
        tail_by_len[tl][0] += 1; tail_by_len[tl][1] += int(r["caught"])
        pos = int(r["role"][4:])
        tail_by_pos[pos][0] += 1; tail_by_pos[pos][1] += int(r["caught"])

    pc = sum(1 for r in passth if r["caught"])

    L = []
    def p(s=""): L.append(s); print(s)

    p("# Full Unbundle: C3 Results\n")
    p(f"Baseline (n100, single-turn):  **{bc}/{nb} = {bc/nb:.1%}** caught\n")
    p(f"Unbundled dataset: {n} rows judged ({len(errs)} errors)\n")
    p(f"Overall caught: **{c}/{n} = {c/n:.1%}**\n")
    p("## Row composition")
    p(f"- turn1 (preserved trie divergence): {len(t1)}")
    p(f"- tail turns (one action each):      {len(tail)}")
    p(f"- untouched passthrough:             {len(passth)}\n")

    p("## T1 slots vs baseline (same 283 trie positions)")
    denom = fixed + regress + kept_c + kept_f
    p(f"- baseline caught here: {fixed+kept_c}/{denom}")
    p(f"- unbundled caught here: {t1_caught}/{len(t1)} = {t1_caught/max(len(t1),1):.1%}")
    p(f"- **fixed (caught -> fooled): {fixed}**")
    p(f"- regressed (fooled -> caught): {regress}")
    p(f"- net catch reduction on T1: {fixed-regress}\n")

    p("## Tail turns (the re-rooted follow-ups)")
    tc = sum(1 for r in tail if r["caught"])
    p(f"- tail caught: {tc}/{len(tail)} = {tc/max(len(tail),1):.1%}")
    p("- by original tail length:")
    for tl in sorted(tail_by_len):
        tot, cc = tail_by_len[tl]
        p(f"    tail_len={tl}: {cc}/{tot} = {cc/tot:.0%}")
    p("- by turn position:")
    for pos in sorted(tail_by_pos):
        tot, cc = tail_by_pos[pos]
        p(f"    turn{pos}: {cc}/{tot} = {cc/tot:.0%}")
    p("")

    p("## Passthrough control (should track baseline)")
    p(f"- caught: {pc}/{len(passth)} = {pc/max(len(passth),1):.1%}")
    bpc = sum(base_caught.get(r["item_id"], False) for r in passth)
    p(f"- same rows in baseline: {bpc}/{len(passth)} = {bpc/max(len(passth),1):.1%}\n")

    # effective per-scenario: does the ORIGINAL confuser (now split) get caught
    # in ANY of its turns? (a scenario is "detected" if any split turn is caught)
    by_orig = defaultdict(list)
    for r in t1 + tail:
        by_orig[r.get("orig_eid")].append(r["caught"])
    scen_any = sum(1 for e, v in by_orig.items() if any(v))
    scen_t1only = sum(1 for r in t1 if r["caught"])
    p("## Per-scenario view (283 split confusers)")
    p(f"- caught in >=1 of its turns: {scen_any}/{len(by_orig)} = {scen_any/len(by_orig):.1%}")
    p(f"- caught at the T1 (trie) slot: {scen_t1only}/{len(by_orig)} = {scen_t1only/len(by_orig):.1%}")
    baseline_on_split = sum(base_caught.get(e, False) for e in by_orig)
    p(f"- baseline caught these same confusers: {baseline_on_split}/{len(by_orig)} = {baseline_on_split/len(by_orig):.1%}")

    MD.write_text("\n".join(L) + "\n", encoding="utf-8")
    print(f"\nwrote {MD.relative_to(ROOT)}")


if __name__ == "__main__":
    main()