scripts/build.py

"""Build Privacy Context Pairs from authored cases; no model or network access."""

from __future__ import annotations

import argparse
import hashlib
import json
import sys
from collections import Counter
from itertools import product
from pathlib import Path

SPLITS = ("train", "validation", "test")
ROLES = ("alternative", "attribute")
SIDES = ("left", "right")
ORDINARY = ("a", "b")


def _canonical(value: object) -> str:
    return json.dumps(value, ensure_ascii=False, sort_keys=True, separators=(",", ":"))


def _sha(payload: bytes) -> str:
    return hashlib.sha256(payload).hexdigest()


def _unique_object(pairs: list[tuple[str, object]]) -> dict:
    result = {}
    for key, value in pairs:
        if key in result:
            raise ValueError(f"Duplicate JSON key: {key}")
        result[key] = value
    return result


def _entry(root: Path, relative: str) -> dict:
    path = root / relative
    if not path.resolve().is_relative_to(root) or path.is_symlink():
        raise ValueError(f"Input must be an ordinary file inside its root: {relative}")
    payload = path.read_bytes()
    return {"path": relative, "bytes": len(payload), "sha256": _sha(payload)}


def _load(root: Path, relative: str) -> tuple[object, dict]:
    path = root / relative
    if not path.resolve().is_relative_to(root) or path.is_symlink():
        raise ValueError(f"Input must be an ordinary file inside its root: {relative}")
    payload = path.read_bytes()
    value = json.loads(payload, object_pairs_hook=_unique_object)
    return value, {"path": relative, "bytes": len(payload), "sha256": _sha(payload)}


def _write_json(path: Path, value: object) -> None:
    payload = json.dumps(value, indent=2, ensure_ascii=False, sort_keys=True) + "\n"
    with path.open("x", encoding="utf-8", newline="\n") as handle:
        handle.write(payload)


def _write_rows(path: Path, rows: list[dict], key: str) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open("x", encoding="utf-8", newline="\n") as handle:
        for row in sorted(rows, key=lambda item: item[key]):
            handle.write(_canonical(row) + "\n")


def _load_cases(root: Path, protocol: dict) -> tuple[list[dict], list[dict]]:
    cases = []
    inputs = []
    required = set(protocol["source_case_keys"])
    languages = set(protocol["languages"])
    per_type = protocol["case_inventory"]["families_per_type"]
    seen = set()
    for target_type in protocol["target_types"]:
        values, entry = _load(root, f"source/{target_type}.json")
        inputs.append(entry)
        if not isinstance(values, list) or len(values) != per_type * len(languages):
            raise ValueError(
                f"{entry['path']}: expected {per_type * len(languages)} cases"
            )
        expected = {
            f"{target_type}-{number:02d}-{language}"
            for number in range(1, per_type + 1)
            for language in languages
        }
        observed = set()
        for case in values:
            if not isinstance(case, dict) or set(case) != required:
                raise ValueError(f"{entry['path']}: source case has incorrect keys")
            if any(
                not isinstance(value, str) or not value.strip()
                for value in case.values()
            ):
                raise ValueError(
                    f"{entry['path']}: all case values must be nonempty strings"
                )
            case_id = case["case_id"]
            if case_id in seen:
                raise ValueError(f"Duplicate case ID: {case_id}")
            seen.add(case_id)
            observed.add(case_id)
            if case["target_type"] != target_type or case["language"] not in languages:
                raise ValueError(f"{case_id}: source type or language mismatch")
            if case_id != f"{case['family_id']}-{case['language']}":
                raise ValueError(f"{case_id}: family/language identifier mismatch")
            cases.append(case)
        if observed != expected:
            raise ValueError(f"{entry['path']}: incomplete family/language inventory")
    return sorted(cases, key=lambda case: case["case_id"]), inputs


def _family_splits(cases: list[dict], protocol: dict) -> dict[str, str]:
    seed = protocol["split_recipe"]["seed"]
    sizes = protocol["split_recipe"]["family_counts_per_type"]
    assignments = {}
    for target_type in protocol["target_types"]:
        families = {
            case["family_id"] for case in cases if case["target_type"] == target_type
        }
        ordered = sorted(families, key=lambda family: _sha(f"{seed}|{family}".encode()))
        offset = 0
        for split in SPLITS:
            for family in ordered[offset : offset + sizes[split]]:
                assignments[family] = split
            offset += sizes[split]
        if offset != len(ordered):
            raise ValueError(f"{target_type}: split sizes do not cover all families")
    return assignments


def _record_id(case_id: str, role: str, side: str, ordinary: str) -> str:
    return f"pcp-v1-{case_id}-{role}-{side}-{ordinary}"


def _text_row(
    case: dict, split: str, case_sha: str, role: str, side: str, ordinary: str
) -> dict:
    cue = case[f"{role}_cue"]
    left, right = (cue, case["control"]) if side == "left" else (case["control"], cue)
    scene = case[f"ordinary_{ordinary}"]
    target = case["target"]
    text = f"{left}\n\n{target}.\n\n{right}\n\n{scene}"
    start = len(left) + 2
    end = start + len(target)
    right_start = end + 3
    right_end = right_start + len(right)
    ordinary_start = right_end + 2
    return {
        "id": _record_id(case["case_id"], role, side, ordinary),
        "case_id": case["case_id"],
        "family_id": case["family_id"],
        "language": case["language"],
        "target_type": case["target_type"],
        "split": split,
        "scenario_domain": case["scenario_domain"],
        "text": text,
        "target": {"text": target, "start": start, "end": end},
        "construction": {
            "role": role,
            "cue_side": side,
            "ordinary": ordinary,
            "intended_reference": case[f"{role}_reference"],
        },
        "slots": {
            "left": {"start": 0, "end": len(left)},
            "right": {"start": right_start, "end": right_end},
            "ordinary": {"start": ordinary_start, "end": len(text)},
        },
        "provenance": {
            "source_type": "synthetic",
            "authoring": "ai_assisted",
            "privacy_annotation_status": "not_human_adjudicated",
            "classifier_screening": "none",
        },
        "case_sha256": case_sha,
        "text_sha256": _sha(text.encode("utf-8")),
        "version": "1.0.0",
    }


def _pairs(case: dict, split: str) -> list[dict]:
    case_id = case["case_id"]
    common = {
        "case_id": case_id,
        "family_id": case["family_id"],
        "language": case["language"],
        "target_type": case["target_type"],
        "split": split,
    }
    result = []
    for side, ordinary in product(SIDES, ORDINARY):
        result.append(
            {
                **common,
                "id": f"{case_id}:role:{side}:{ordinary}",
                "kind": "role",
                "base_id": _record_id(case_id, "alternative", side, ordinary),
                "changed_id": _record_id(case_id, "attribute", side, ordinary),
                "direction": "attribute - alternative",
            }
        )
    for role, side in product(ROLES, SIDES):
        result.append(
            {
                **common,
                "id": f"{case_id}:ordinary:{role}:{side}",
                "kind": "ordinary",
                "base_id": _record_id(case_id, role, side, "a"),
                "changed_id": _record_id(case_id, role, side, "b"),
                "direction": "b - a",
            }
        )
    for role, ordinary in product(ROLES, ORDINARY):
        result.append(
            {
                **common,
                "id": f"{case_id}:cue_side:{role}:{ordinary}",
                "kind": "cue_side",
                "base_id": _record_id(case_id, role, "left", ordinary),
                "changed_id": _record_id(case_id, role, "right", ordinary),
                "direction": "right - left",
            }
        )
    return result


def _stats(
    cases: dict[str, list[dict]],
    texts: dict[str, list[dict]],
    pairs: dict[str, list[dict]],
    protocol: dict,
) -> dict:
    by_split = {}
    by_type = {
        name: {"records": 0, "cases": 0, "families": 0, "contrasts": 0}
        for name in protocol["target_types"]
    }
    by_language = {
        name: {"records": 0, "cases": 0, "contrasts": 0}
        for name in protocol["languages"]
    }
    seen_families = set()
    by_kind = Counter()
    shortest = None
    longest = 0
    for split in SPLITS:
        split_families = set()
        for case in cases[split]:
            target_type = case["target_type"]
            by_type[target_type]["cases"] += 1
            by_language[case["language"]]["cases"] += 1
            family = case["family_id"]
            split_families.add(family)
            if family not in seen_families:
                by_type[target_type]["families"] += 1
                seen_families.add(family)
        for row in texts[split]:
            by_type[row["target_type"]]["records"] += 1
            by_language[row["language"]]["records"] += 1
            length = len(row["text"])
            shortest = length if shortest is None else min(shortest, length)
            longest = max(longest, length)
        for row in pairs[split]:
            by_type[row["target_type"]]["contrasts"] += 1
            by_language[row["language"]]["contrasts"] += 1
            by_kind[row["kind"]] += 1
        by_split[split] = {
            "records": len(texts[split]),
            "cases": len(cases[split]),
            "families": len(split_families),
            "contrasts": len(pairs[split]),
        }
    return {
        "schema_version": 1,
        "name": protocol["name"],
        "version": protocol["version"],
        "records": sum(len(rows) for rows in texts.values()),
        "cases": sum(len(rows) for rows in cases.values()),
        "families": len(seen_families),
        "contrasts": sum(len(rows) for rows in pairs.values()),
        "gold_privacy_labels": 0,
        "by_split": by_split,
        "by_target_type": by_type,
        "by_language": by_language,
        "by_contrast_kind": dict(by_kind),
        "text_codepoint_length": {"min": shortest, "max": longest},
    }


def build(root: Path, out: Path) -> dict:
    root = root.resolve()
    out = out.resolve()
    if out.exists():
        raise ValueError(f"Output already exists; choose a new directory: {out}")
    protocol, protocol_entry = _load(root, ".construction/protocol.json")
    if (
        protocol["version"] != "1.0.0"
        or protocol["classifier_evaluation_allowed"] is not False
    ):
        raise ValueError("Expected the frozen, construction-only v1.0.0 protocol")
    source_cases, source_entries = _load_cases(root, protocol)
    inputs = sorted(
        [
            protocol_entry,
            *source_entries,
            _entry(root, "scripts/build.py"),
            _entry(root, "scripts/validate.py"),
        ],
        key=lambda entry: entry["path"],
    )
    assignments = _family_splits(source_cases, protocol)
    cases = {split: [] for split in SPLITS}
    texts = {split: [] for split in SPLITS}
    pairs = {split: [] for split in SPLITS}
    for source_case in source_cases:
        split = assignments[source_case["family_id"]]
        case_sha = _sha(_canonical(source_case).encode("utf-8"))
        cases[split].append({**source_case, "split": split, "case_sha256": case_sha})
        for role, side, ordinary in product(ROLES, SIDES, ORDINARY):
            texts[split].append(
                _text_row(source_case, split, case_sha, role, side, ordinary)
            )
        pairs[split].extend(_pairs(source_case, split))
    stats = _stats(cases, texts, pairs, protocol)
    counts = {key: stats[key] for key in ("records", "cases", "families", "contrasts")}
    expected = {"records": 2048, "cases": 256, "families": 128, "contrasts": 3072}
    if counts != expected:
        raise ValueError(f"Unexpected construction counts: {counts}")
    for entry in inputs:
        if _entry(root, entry["path"]) != entry:
            raise ValueError(f"Input changed during construction: {entry['path']}")
    out.mkdir(parents=True)
    generated = []
    for split in SPLITS:
        for directory, grouped, key in (
            ("data", texts, "id"),
            ("cases", cases, "case_id"),
            ("pairs", pairs, "id"),
        ):
            relative = f"{directory}/{split}.jsonl"
            _write_rows(out / relative, grouped[split], key)
            generated.append(relative)
    _write_json(out / "stats.json", stats)
    generated.append("stats.json")
    files = [_entry(out, relative) for relative in sorted(generated)]
    manifest = {
        "schema_version": 1,
        "name": protocol["name"],
        "version": protocol["version"],
        "protocol_sha256": protocol_entry["sha256"],
        "inputs": inputs,
        "files": files,
        "counts": counts,
    }
    _write_json(out / "manifest.json", manifest)
    with (out / "manifest.sha256").open("x", encoding="utf-8", newline="\n") as handle:
        handle.write(f"{_sha((out / 'manifest.json').read_bytes())}  manifest.json\n")
    return counts


def main() -> int:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--root", type=Path, default=Path(__file__).resolve().parents[1]
    )
    parser.add_argument(
        "--out",
        type=Path,
        required=True,
        help="New output directory; never overwritten",
    )
    args = parser.parse_args()
    try:
        counts = build(args.root, args.out)
    except (ValueError, OSError, KeyError, TypeError) as error:
        print(f"Build failed: {error}", file=sys.stderr)
        return 1
    print(
        json.dumps(
            {"status": "built", "output": str(args.out.resolve()), **counts},
            sort_keys=True,
        )
    )
    return 0


if __name__ == "__main__":
    raise SystemExit(main())