source/tools/select_corpus.py

#!/usr/bin/env python3
"""Choose which tasks ship, so the corpus hits a target resolved rate.

Difficulty is set by composition, not by any one task. Given a measured resolved
rate per tier, this picks how many of each tier to include so the aggregate
lands in the target band, while keeping the easy tier non-empty -- a corpus that
scores a model at zero discriminates no better than one that scores it at 100%.

Rates come from measurement (see CALIBRATION.md); they are inputs here, never
assumptions baked into the selection.
"""
import argparse, collections, json, sys
from pathlib import Path

HERE = Path(__file__).parent
HAND_COMPOUND = {
    "incident-triage", "scheduler-regressions", "retry-subsystem-broken",
    "api-review-findings", "routing-regressions", "negotiation-and-headers",
    "audit-findings",
}


def tier_of(task_id: str) -> str:
    if any(k in task_id for k in ("triage3", "review3", "audit3")):
        return "d3"
    if any(k in task_id for k in ("triage2", "review2", "audit2")):
        return "d2"
    if task_id.split("-", 1)[1] in HAND_COMPOUND:
        return "d3"
    return "single"


def load(path: Path):
    return [json.loads(line) for line in path.read_text().splitlines() if line.strip()]


def choose(rows, rates, target, total, floor_single, easy_repo="ledger"):
    """Pick counts per tier whose weighted rate is closest to `target`."""
    pools = collections.defaultdict(list)
    for row in rows:
        tier = tier_of(row["task_id"])
        # A compound inherits the difficulty of the repo it sits in. The small
        # repo can be read end to end, which removes the orientation cost that
        # makes compounds hard, so its tasks count as easy however many defects
        # they carry.
        if tier != "single" and row.get("repo") == easy_repo:
            tier = "single"
        pools[tier].append(row)
    for tasks in pools.values():
        tasks.sort(key=lambda r: r["task_id"])

    best = None
    for n_single in range(floor_single, min(len(pools["single"]), total) + 1):
        for n_d2 in range(0, min(len(pools["d2"]), total - n_single) + 1):
            n_d3 = total - n_single - n_d2
            if n_d3 < 0 or n_d3 > len(pools["d3"]):
                continue
            solved = (n_single * rates["single"] + n_d2 * rates["d2"]
                      + n_d3 * rates["d3"])
            rate = solved / total
            score = abs(rate - target)
            if best is None or score < best[0]:
                best = (score, rate, n_single, n_d2, n_d3)
    return best, pools


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--corpus", default=str(HERE / "agentic-corpus.jsonl"))
    ap.add_argument("--out", default=str(HERE / "agentic-corpus-selected.jsonl"))
    ap.add_argument("--total", type=int, default=60)
    ap.add_argument("--target", type=float, default=0.175,
                    help="midpoint of the wanted resolved band")
    ap.add_argument("--floor-single", type=int, default=4,
                    help="minimum easy-tier tasks, so the floor is not zero")
    ap.add_argument("--rate-single", type=float, required=True)
    ap.add_argument("--rate-d2", type=float, required=True)
    ap.add_argument("--rate-d3", type=float, required=True)
    ap.add_argument("--easy-repo", default="ledger",
                    help="repo whose tasks are treated as the easy tier at every "
                         "defect count, and excluded from the hard tiers")
    args = ap.parse_args()

    rows = load(Path(args.corpus))
    rates = {"single": args.rate_single, "d2": args.rate_d2, "d3": args.rate_d3}
    best, pools = choose(rows, rates, args.target, args.total,
                         args.floor_single, args.easy_repo)
    if best is None:
        print("no selection satisfies the constraints")
        return 1

    _, rate, n_single, n_d2, n_d3 = best
    picked = (pools["single"][:n_single] + pools["d2"][:n_d2] + pools["d3"][:n_d3])

    print(f"measured rates: single={rates['single']:.0%} "
          f"d2={rates['d2']:.0%} d3={rates['d3']:.0%}")
    print(f"selection ({args.total} tasks): "
          f"{n_single} single + {n_d2} two-defect + {n_d3} three-defect")
    print(f"projected resolved rate: {rate:.1%}")
    langs = collections.Counter(r["lang"] for r in picked)
    repos = collections.Counter(r["repo"] for r in picked)
    print(f"languages: {dict(langs)}")
    print(f"repos:     {dict(repos)}")

    Path(args.out).write_text("".join(json.dumps(r) + "\n" for r in picked))
    print(f"\nwrote {len(picked)} tasks -> {args.out}")
    return 0


if __name__ == "__main__":
    sys.exit(main())