scripts/validate.py

#!/usr/bin/env python3
"""Independently validate every Privacy Context Pairs v1.0.0 construction record.

Only the new, external report is written. This program uses the standard library,
never imports the builder, and performs no network, model, or tokenizer calls.
Passing establishes the frozen mechanical contract, not semantic privacy truth.
"""

import argparse
import hashlib
import itertools
import json
import os
import re
import sys
import traceback
import unicodedata
from collections import Counter, defaultdict
from dataclasses import dataclass
from datetime import date, datetime, timezone
from pathlib import Path, PurePosixPath
from urllib.parse import urlsplit

NAME = "Privacy Context Pairs"
VERSION = "1.0.0"
PROTOCOL_SHA256 = "2319f14b79f755d9b7e3a0ed6ddf397fc16c1dac784f21b6dd6fee30956d80f1"
SEED = "privacy-context-pairs-v1.0.0"
TYPES = (
    "person",
    "address",
    "email",
    "phone",
    "url",
    "date",
    "account_number",
    "secret",
)
LANGUAGES = ("en", "pt-BR")
SPLITS = ("train", "validation", "test")
ROLES = ("alternative", "attribute")
SIDES = ("left", "right")
ORDINARY = ("a", "b")
KINDS = ("role", "ordinary", "cue_side")
CONTEXT_FIELDS = (
    "attribute_cue",
    "alternative_cue",
    "control",
    "ordinary_a",
    "ordinary_b",
)
SOURCE_FIELDS = (
    "case_id",
    "family_id",
    "language",
    "target_type",
    "target",
    "scenario_domain",
    "attribute_reference",
    "alternative_reference",
    "attribute_cue",
    "alternative_cue",
    "control",
    "ordinary_a",
    "ordinary_b",
    "synthetic_target_notes",
)
SOURCE_SCHEMA = dict.fromkeys(SOURCE_FIELDS, str)
CASE_SCHEMA = {**SOURCE_SCHEMA, "split": str, "case_sha256": str}
SPAN_SCHEMA = {"start": int, "end": int}
TEXT_SCHEMA = {
    "id": str,
    "case_id": str,
    "family_id": str,
    "language": str,
    "target_type": str,
    "split": str,
    "scenario_domain": str,
    "text": str,
    "target": {"text": str, "start": int, "end": int},
    "construction": {
        "role": str,
        "cue_side": str,
        "ordinary": str,
        "intended_reference": str,
    },
    "slots": {"left": SPAN_SCHEMA, "right": SPAN_SCHEMA, "ordinary": SPAN_SCHEMA},
    "provenance": {
        "source_type": str,
        "authoring": str,
        "privacy_annotation_status": str,
        "classifier_screening": str,
    },
    "case_sha256": str,
    "text_sha256": str,
    "version": str,
}
PAIR_SCHEMA = {
    "id": str,
    "kind": str,
    "case_id": str,
    "family_id": str,
    "language": str,
    "target_type": str,
    "split": str,
    "base_id": str,
    "changed_id": str,
    "direction": str,
}
PROVENANCE = {
    "source_type": "synthetic",
    "authoring": "ai_assisted",
    "privacy_annotation_status": "not_human_adjudicated",
    "classifier_screening": "none",
}
TOTALS = {"records": 2048, "cases": 256, "families": 128, "contrasts": 3072}
SPLIT_TOTALS = {
    "train": {"records": 1280, "cases": 160, "families": 80, "contrasts": 1920},
    "validation": {"records": 256, "cases": 32, "families": 16, "contrasts": 384},
    "test": {"records": 512, "cases": 64, "families": 32, "contrasts": 768},
}
TYPE_TOTALS = {"records": 256, "cases": 32, "families": 16, "contrasts": 384}
LANGUAGE_TOTALS = {"records": 1024, "cases": 128, "contrasts": 1536}
INPUT_PATHS = tuple(
    sorted(
        [".construction/protocol.json", "scripts/build.py", "scripts/validate.py"]
        + [f"source/{kind}.json" for kind in TYPES]
    )
)
PAYLOAD_PATHS = tuple(
    sorted(
        [
            f"{directory}/{split}.jsonl"
            for directory in ("data", "cases", "pairs")
            for split in SPLITS
        ]
        + ["stats.json"]
    )
)
RELEASE_PATHS = set(PAYLOAD_PATHS) | {"manifest.json", "manifest.sha256"}
EXCLUDED = (
    "May",
    "Amber",
    "Rosa",
    "Violeta",
    "2001-02-03",
    "1987-06-14",
    "2025550183",
    "24681013570",
)
EN_MONTHS = (
    "january",
    "february",
    "march",
    "april",
    "may",
    "june",
    "july",
    "august",
    "september",
    "october",
    "november",
    "december",
)
PT_MONTHS = (
    "janeiro",
    "fevereiro",
    "março",
    "abril",
    "maio",
    "junho",
    "julho",
    "agosto",
    "setembro",
    "outubro",
    "novembro",
    "dezembro",
)
LIMITATIONS = [
    "Structural validation cannot certify fluent language, semantic scenario distinctness, bilingual equivalence, intended referents, or ordinary-sentence relevance.",
    "Construction roles and references are AI-assisted authoring intentions, not human annotation, human privacy adjudication, privacy gold, masking decisions, or expected classifier predictions.",
    "Mechanical category checks establish permitted string forms and reserved-domain/phone-pool constraints; they cannot certify that a name, date, account-like string, address, or credential-like fixture never coincides with real data or has never been issued.",
    "Synthetic provenance, absence of copied credentials or recovery mnemonics, no consultation of prior outputs or external records, and no model-guided selection are authoring-process declarations, not facts provable from these files.",
    "No tokenizer was run; tokenizer matching, token lengths, token positions, and model-specific comparability are not certified.",
    "No classifier or model was run; this is not classifier screening or evaluation and does not establish policy exclusion for alternative contexts.",
    "This structured, family-disjoint construction is not representative sampling, a natural corpus, a template-disjoint corpus, or proof of statistical independence or generalizability.",
    "This validator does not establish publication approval, human construction review, license ownership, or reproducible rebuild execution; those require separately retained evidence and authorization.",
]


def digest(blob):
    return hashlib.sha256(blob).hexdigest()


def compact(value):
    return json.dumps(
        value,
        sort_keys=True,
        ensure_ascii=False,
        separators=(",", ":"),
        allow_nan=False,
    )


def pretty_bytes(value):
    return (
        json.dumps(value, indent=2, sort_keys=True, ensure_ascii=False, allow_nan=False)
        + "\n"
    ).encode("utf-8")


def normalized(value):
    return unicodedata.normalize("NFC", unicodedata.normalize("NFC", value).casefold())


def one_line(value):
    return (
        bool(value)
        and value == value.strip()
        and not any(
            unicodedata.category(char) in {"Cc", "Cs", "Zl", "Zp"} for char in value
        )
    )


def phone_digits(value):
    digits = "".join(char for char in value if "0" <= char <= "9")
    return digits[1:] if len(digits) == 11 and digits.startswith("1") else digits


def target_keys(kind, value):
    keys = {("literal", normalized(value))}
    if kind == "phone":
        keys.add(("numeric", phone_digits(value)))
    elif kind == "account_number" and re.fullmatch(r"[0-9\s()./+-]+", value):
        keys.add(
            (
                "numeric",
                "".join(char for char in value if char.isascii() and char.isdigit()),
            )
        )
    return keys


def occurrences(haystack, needle):
    if not needle:
        return 0
    return sum(1 for _ in re.finditer("(?=" + re.escape(needle) + ")", haystack))


def example_host(host):
    if not isinstance(host, str) or not host or len(host) > 253:
        return False
    host = host.lower()
    if any(
        not re.fullmatch(r"[a-z0-9](?:[a-z0-9-]{0,61}[a-z0-9])?", label)
        for label in host.split(".")
    ):
        return False
    return any(
        host == root or host.endswith("." + root)
        for root in ("example.com", "example.net", "example.org")
    )


def calendar_date(value, language):
    """Parse explicitly unambiguous formats without locale or third-party parsers."""
    match = re.fullmatch(r"([0-9]{4})([-/.])([0-9]{2})\2([0-9]{2})", value)
    if match:
        return date(int(match[1]), int(match[3]), int(match[4]))
    folded = normalized(value)
    months = EN_MONTHS if language == "en" else PT_MONTHS
    month_pattern = "(" + "|".join(months) + ")"
    if language == "en":
        match = re.fullmatch(month_pattern + r" ([0-9]{1,2}), ([0-9]{4})", folded)
        if match:
            return date(int(match[3]), months.index(match[1]) + 1, int(match[2]))
        match = re.fullmatch(r"([0-9]{1,2}) " + month_pattern + r" ([0-9]{4})", folded)
    else:
        match = re.fullmatch(
            r"([0-9]{1,2}) de " + month_pattern + r" de ([0-9]{4})", folded
        )
    if match:
        return date(int(match[3]), months.index(match[2]) + 1, int(match[1]))
    raise ValueError(
        "use YYYY-MM-DD (or another year-first numeric format), a full English month form, or D de mês de YYYY"
    )


def no_duplicate_keys(pairs):
    result = {}
    for key, value in pairs:
        if key in result:
            raise ValueError(f"duplicate JSON object key {key!r}")
        result[key] = value
    return result


def reject_constant(value):
    raise ValueError(f"non-JSON numeric constant {value}")


def decode_json(text):
    return json.loads(
        text, object_pairs_hook=no_duplicate_keys, parse_constant=reject_constant
    )


def preview(value):
    rendered = repr(value)
    return rendered if len(rendered) <= 220 else rendered[:217] + "..."


@dataclass
class Record:
    value: object
    split: str
    location: str


class Validator:
    def __init__(self, root, release, producer):
        self.root = root
        self.release = release
        self.producer = producer
        self.errors = []
        self.checks = {}
        self.inputs = {}
        self.payloads = {}
        self.sources = {}
        self.records = {"cases": [], "data": [], "pairs": []}
        self.expected = {"cases": {}, "data": {}, "pairs": {}}
        self.split_for = {}
        self.safety = {kind: [] for kind in TYPES}
        self.observed = {}
        self.manifest_blob = None

    def check(self, condition, check, location, message):
        summary = self.checks.setdefault(check, {"evaluated": 0, "failed": 0})
        summary["evaluated"] += 1
        if not condition:
            summary["failed"] += 1
            self.errors.append(
                {"check": check, "location": location, "message": message}
            )
        return bool(condition)

    def equal(self, actual, expected, location, check="exact_values"):
        """Compare all fields strictly, including bool versus int and unexpected keys."""
        if type(actual) is not type(expected):
            return self.check(
                False,
                check,
                location,
                f"expected {type(expected).__name__}, got {type(actual).__name__}: {preview(actual)}",
            )
        valid = True
        if isinstance(expected, dict):
            valid &= self.check(
                set(actual) == set(expected),
                check,
                location,
                f"keys differ: missing={sorted(set(expected) - set(actual))}, extra={sorted(set(actual) - set(expected))}",
            )
            for key in sorted(set(actual) & set(expected)):
                valid &= self.equal(
                    actual[key], expected[key], f"{location}.{key}", check
                )
        elif isinstance(expected, list):
            valid &= self.check(
                len(actual) == len(expected),
                check,
                location,
                f"expected {len(expected)} entries, got {len(actual)}",
            )
            for index, (left, right) in enumerate(zip(actual, expected)):
                valid &= self.equal(left, right, f"{location}[{index}]", check)
        else:
            valid &= self.check(
                actual == expected,
                check,
                location,
                f"expected {preview(expected)}, got {preview(actual)}",
            )
        return bool(valid)

    def shape(self, value, schema, location):
        if isinstance(schema, dict):
            if not self.check(
                type(value) is dict, "schema", location, "expected a JSON object"
            ):
                return False
            valid = self.check(
                set(value) == set(schema),
                "schema",
                location,
                f"keys differ: missing={sorted(set(schema) - set(value))}, extra={sorted(set(value) - set(schema))}",
            )
            for key in sorted(set(value) & set(schema)):
                valid &= self.shape(value[key], schema[key], f"{location}.{key}")
            return bool(valid)
        valid = self.check(
            type(value) is schema,
            "schema",
            location,
            f"expected {schema.__name__}, got {type(value).__name__}",
        )
        if valid and schema is str:
            valid &= self.check(
                bool(value.strip()), "schema", location, "string must be nonempty"
            )
            valid &= self.check(
                not any(unicodedata.category(char) == "Cs" for char in value),
                "unicode",
                location,
                "unpaired Unicode surrogate is not valid UTF-8 corpus content",
            )
        return bool(valid)

    def read_blob(self, scope, relative, check):
        path = scope / relative
        try:
            resolved = path.resolve(strict=True)
            if not self.check(
                resolved.is_relative_to(scope),
                check,
                relative,
                "path resolves outside its declared root",
            ):
                return None
            if not self.check(
                resolved.is_file(), check, relative, "expected a regular file"
            ):
                return None
            return resolved.read_bytes()
        except (OSError, RuntimeError, ValueError) as exc:
            self.check(False, check, relative, f"cannot read file: {exc}")
            return None

    def parse(self, blob, location):
        if blob is None:
            return None
        try:
            return decode_json(blob.decode("utf-8"))
        except (UnicodeError, ValueError, RecursionError) as exc:
            self.check(False, "json_parse", location, f"invalid UTF-8 JSON: {exc}")
            return None

    def inventories(self):
        self.check(
            self.root.is_dir(),
            "input_inventory",
            str(self.root),
            "root must be an existing directory",
        )
        self.check(
            self.release.is_dir(),
            "release_inventory",
            str(self.release),
            "release must be an existing directory",
        )
        source_dir = self.root / "source"
        try:
            resolved_source = source_dir.resolve(strict=True)
            if self.check(
                resolved_source.is_relative_to(self.root),
                "input_inventory",
                "source",
                "source directory escapes root",
            ):
                names = {entry.name for entry in source_dir.iterdir()}
                self.equal(
                    sorted(names),
                    sorted(f"{kind}.json" for kind in TYPES),
                    "source",
                    "input_inventory",
                )
        except (OSError, RuntimeError) as exc:
            self.check(
                False,
                "input_inventory",
                "source",
                f"cannot enumerate source directory: {exc}",
            )
        found = set()

        def walk_error(exc):
            self.check(
                False,
                "release_inventory",
                str(self.release),
                f"cannot enumerate release: {exc}",
            )

        if self.release.is_dir():
            for directory, dirs, files in os.walk(
                self.release, followlinks=False, onerror=walk_error
            ):
                base = Path(directory)
                for name in dirs:
                    path = base / name
                    relative = path.relative_to(self.release).as_posix()
                    self.check(
                        not path.is_symlink(),
                        "release_inventory",
                        relative,
                        "release directories must not be symlinks",
                    )
                    self.check(
                        relative in {"cases", "data", "pairs"},
                        "release_inventory",
                        relative,
                        "unexpected release directory",
                    )
                for name in files:
                    path = base / name
                    relative = path.relative_to(self.release).as_posix()
                    found.add(relative)
                    self.check(
                        not path.is_symlink(),
                        "release_inventory",
                        relative,
                        "release files must not be symlinks",
                    )
        self.equal(sorted(found), sorted(RELEASE_PATHS), "release", "release_inventory")
        for relative in INPUT_PATHS:
            blob = self.read_blob(self.root, relative, "input_files")
            if blob is not None:
                self.inputs[relative] = blob
        for relative in PAYLOAD_PATHS:
            blob = self.read_blob(self.release, relative, "release_files")
            if blob is not None:
                self.payloads[relative] = blob
        protocol = self.inputs.get(".construction/protocol.json")
        if protocol is not None:
            self.equal(
                digest(protocol),
                PROTOCOL_SHA256,
                ".construction/protocol.json.sha256",
                "frozen_protocol",
            )
            self.check(
                type(self.parse(protocol, ".construction/protocol.json")) is dict,
                "frozen_protocol",
                ".construction/protocol.json",
                "protocol must be a JSON object",
            )
        validator_blob = self.inputs.get("scripts/validate.py")
        if validator_blob is not None:
            self.equal(
                digest(validator_blob),
                self.producer["sha256"],
                "scripts/validate.py.sha256",
                "producer_identity",
            )

    def category(self, case, location):
        kind, target, language = case["target_type"], case["target"], case["language"]
        item = {
            "case_id": case["case_id"],
            "family_id": case["family_id"],
            "target": target,
            "normalized_target": normalized(target),
            "synthetic_target_notes": case["synthetic_target_notes"],
        }
        if kind not in TYPES:
            return
        if kind == "person":
            self.check(
                any(char.isalpha() for char in target)
                and all(
                    unicodedata.category(char)[0] in {"L", "M"} or char in " .-'’"
                    for char in target
                ),
                "person_form",
                location,
                "person target must be a name-like letter string, not an identifier or record",
            )
        elif kind == "address":
            marker = "Example" if language == "en" else "Exemplo"
            item["required_fictional_marker"] = marker
            self.check(
                marker.casefold() in target.casefold(),
                "address_safety",
                location,
                f"invented address must contain {marker!r}",
            )
        elif kind == "email":
            match = re.fullmatch(
                r"([A-Za-z0-9!#$%&'*+/=?^_`{|}~.-]+)@([^@\s]+)", target
            )
            host = match[2] if match else ""
            item["host"] = host.lower()
            self.check(
                bool(match)
                and not match[1].startswith(".")
                and not match[1].endswith(".")
                and ".." not in match[1]
                and example_host(host),
                "email_safety",
                location,
                "email must have a valid unquoted local part and an example.com/.net/.org host or subdomain",
            )
        elif kind == "phone":
            canonical = phone_digits(target)
            item["canonical_digits"] = canonical
            self.check(
                bool(re.fullmatch(r"[+0-9(). /-]+", target))
                and bool(re.fullmatch(r"20255501[0-9]{2}", canonical)),
                "phone_safety",
                location,
                "phone must contain only permitted formatting and 20255501xx, optionally prefixed by country digit 1",
            )
            self.check(
                canonical != "2025550183",
                "excluded_targets",
                location,
                "earlier phone fixture is excluded in every punctuation form",
            )
        elif kind == "url":
            try:
                parsed = urlsplit(target)
                host = parsed.hostname or ""
                item["host"] = host.lower()
                port = parsed.port
                self.check(
                    parsed.scheme == "https"
                    and example_host(host)
                    and parsed.username is None
                    and parsed.password is None
                    and not any(char.isspace() or ord(char) < 32 for char in target)
                    and "\\" not in target,
                    "url_safety",
                    location,
                    "URL must use https on an example.com/.net/.org host or subdomain, without userinfo, backslashes, or whitespace",
                )
                if port is not None:
                    self.check(
                        0 < port <= 65535,
                        "url_safety",
                        location,
                        "URL port must be between 1 and 65535",
                    )
            except ValueError as exc:
                self.check(False, "url_safety", location, f"invalid URL: {exc}")
        elif kind == "date":
            try:
                parsed_date = calendar_date(target, language)
                item["iso_date"] = parsed_date.isoformat()
                self.check(
                    parsed_date.isoformat() not in {"2001-02-03", "1987-06-14"},
                    "excluded_targets",
                    location,
                    "earlier calendar date is excluded in every supported display form",
                )
            except ValueError as exc:
                self.check(
                    False,
                    "date_safety",
                    location,
                    f"invalid or ambiguous calendar date: {exc}",
                )
        elif kind == "account_number":
            self.check(
                any("0" <= char <= "9" for char in target)
                and bool(re.fullmatch(r"[A-Za-z0-9 ._/#()+-]+", target)),
                "account_form",
                location,
                "account-like fixture must contain ASCII digits and only letters or ordinary identifier separators",
            )
            item["numeric_format_normalization"] = sorted(
                value for key, value in target_keys(kind, target) if key == "numeric"
            )
            self.check(
                not any(
                    value == "24681013570"
                    for key, value in target_keys(kind, target)
                    if key == "numeric"
                ),
                "excluded_targets",
                location,
                "earlier numeric account fixture is excluded in every punctuation form",
            )
        elif kind == "secret":
            item["ascii_length"] = len(target)
            self.check(
                12 <= len(target) <= 64
                and all(32 <= ord(char) <= 126 for char in target),
                "secret_safety",
                location,
                "credential-like fixture must be 12–64 printable ASCII characters",
            )
            self.check(
                "-----BEGIN" not in target.upper()
                and "PRIVATE KEY" not in target.upper(),
                "secret_safety",
                location,
                "private-key block markers are forbidden",
            )
        self.safety[kind].append(item)

    def source_cases(self):
        family_cases = defaultdict(list)
        target_owners = defaultdict(set)
        tuple_owners = defaultdict(set)
        excluded = {normalized(value) for value in EXCLUDED}
        for kind in TYPES:
            relative = f"source/{kind}.json"
            values = self.parse(self.inputs.get(relative), relative)
            if not self.check(
                type(values) is list,
                "source_inventory",
                relative,
                "expected a source JSON array",
            ):
                continue
            self.equal(len(values), 32, relative + ".length", "source_inventory")
            expected_ids = {
                f"{kind}-{number:02d}-{language}"
                for number in range(1, 17)
                for language in LANGUAGES
            }
            observed_ids = []
            for index, case in enumerate(values):
                location = f"{relative}[{index}]"
                if not self.shape(case, SOURCE_SCHEMA, location):
                    continue
                case_id, family, language = (
                    case["case_id"],
                    case["family_id"],
                    case["language"],
                )
                observed_ids.append(case_id)
                identity_ok = self.check(
                    case["target_type"] == kind,
                    "source_identity",
                    location,
                    f"target_type must be {kind!r}",
                )
                identity_ok &= self.check(
                    language in LANGUAGES,
                    "source_identity",
                    location,
                    f"unsupported language {language!r}",
                )
                identity_ok &= self.check(
                    family in {f"{kind}-{number:02d}" for number in range(1, 17)},
                    "source_identity",
                    location,
                    f"invalid family_id {family!r}",
                )
                identity_ok &= self.check(
                    case_id == f"{family}-{language}",
                    "source_identity",
                    location,
                    "case_id must be family_id + '-' + language",
                )
                self.check(
                    bool(
                        re.fullmatch(
                            r"[a-z][a-z0-9]*(?:_[a-z0-9]+)*", case["scenario_domain"]
                        )
                    ),
                    "source_text_rules",
                    location + ".scenario_domain",
                    "expected an English ASCII snake_case domain",
                )
                target = case["target"]
                self.check(
                    one_line(target),
                    "source_text_rules",
                    location + ".target",
                    "target must be a nonempty single line without outer whitespace or control characters",
                )
                self.check(
                    normalized(target) not in excluded,
                    "excluded_targets",
                    location + ".target",
                    "target reuses an excluded earlier fixture",
                )
                for field in CONTEXT_FIELDS:
                    sentence = case[field]
                    self.check(
                        one_line(sentence) and sentence.endswith("."),
                        "source_text_rules",
                        location + "." + field,
                        "context slot must be a single-line sentence ending in a period, with no outer whitespace",
                    )
                    self.check(
                        normalized(target) not in normalized(sentence),
                        "target_exclusivity",
                        location + "." + field,
                        f"context contains target {target!r} under NFC/casefold, including substrings",
                    )
                self.check(
                    len(
                        {
                            normalized(case[field])
                            for field in ("attribute_cue", "alternative_cue", "control")
                        }
                    )
                    == 3,
                    "source_distinctions",
                    location,
                    "attribute cue, alternative cue, and control must differ",
                )
                self.check(
                    normalized(case["ordinary_a"]) != normalized(case["ordinary_b"]),
                    "source_distinctions",
                    location,
                    "ordinary_a and ordinary_b must differ",
                )
                for cue in ("attribute_cue", "alternative_cue"):
                    for ordinary in ("ordinary_a", "ordinary_b"):
                        self.check(
                            normalized(case[ordinary]) not in normalized(case[cue]),
                            "source_distinctions",
                            location + "." + cue,
                            f"role cue must not contain the entire {ordinary} sentence",
                        )
                self.category(case, location + ".target")
                if not identity_ok:
                    continue
                self.check(
                    case_id not in self.sources,
                    "source_uniqueness",
                    location,
                    f"duplicate case_id {case_id!r}",
                )
                if case_id not in self.sources:
                    self.sources[case_id] = case
                family_cases[family].append(case)
                for key in target_keys(kind, target):
                    target_owners[key].add(family)
                tuple_owners[
                    tuple(normalized(case[field]) for field in CONTEXT_FIELDS)
                ].add(family)
            self.equal(
                sorted(observed_ids),
                sorted(expected_ids),
                relative + ".case_ids",
                "source_inventory",
            )
        self.equal(len(self.sources), 256, "source.cases", "source_inventory")
        self.equal(len(family_cases), 128, "source.families", "source_inventory")
        for family, cases in sorted(family_cases.items()):
            self.equal(
                sorted(case["language"] for case in cases),
                sorted(LANGUAGES),
                family + ".languages",
                "bilingual_families",
            )
            self.check(
                len({case["scenario_domain"] for case in cases}) == 1,
                "bilingual_families",
                family,
                "paired languages must share scenario_domain",
            )
        for key, families in sorted(target_owners.items()):
            self.check(
                len(families) == 1,
                "source_target_reuse",
                preview(key),
                f"target reused by different families: {sorted(families)}",
            )
        for families in tuple_owners.values():
            self.check(
                len(families) == 1,
                "source_scenario_duplicates",
                "source.context_tuples",
                f"distinct families share an identical NFC/casefold cue/control/ordinary tuple: {sorted(families)}",
            )
        for kind in TYPES:
            families = [f"{kind}-{number:02d}" for number in range(1, 17)]
            ranked = sorted(
                families, key=lambda family: digest(f"{SEED}|{family}".encode("utf-8"))
            )
            for rank, family in enumerate(ranked):
                self.split_for[family] = (
                    "train" if rank < 10 else "validation" if rank < 12 else "test"
                )
        for key, families in sorted(target_owners.items()):
            splits = {self.split_for[family] for family in families}
            self.check(
                len(splits) == 1,
                "source_target_leakage",
                preview(key),
                f"normalized target crosses splits: {sorted(splits)}; families={sorted(families)}",
            )

    def reconstruct(self):
        for case_id, source in sorted(self.sources.items()):
            split = self.split_for[source["family_id"]]
            case_hash = digest(compact(source).encode("utf-8"))
            self.expected["cases"][case_id] = {
                **source,
                "split": split,
                "case_sha256": case_hash,
            }
            shared = {
                field: source[field]
                for field in (
                    "case_id",
                    "family_id",
                    "language",
                    "target_type",
                    "scenario_domain",
                )
            }
            shared["split"] = split
            cell_ids = {}
            for role, side, ordinary in itertools.product(ROLES, SIDES, ORDINARY):
                active = source[role + "_cue"]
                left, right = (
                    (active, source["control"])
                    if side == "left"
                    else (source["control"], active)
                )
                ending = source["ordinary_" + ordinary]
                target = source["target"]
                start = len(left) + 2
                end = start + len(target)
                right_start = end + 3
                ordinary_start = right_start + len(right) + 2
                text = "\n\n".join((left, target + ".", right, ending))
                identifier = f"pcp-v1-{case_id}-{role}-{side}-{ordinary}"
                cell_ids[role, side, ordinary] = identifier
                self.expected["data"][identifier] = {
                    **shared,
                    "id": identifier,
                    "text": text,
                    "target": {"text": target, "start": start, "end": end},
                    "construction": {
                        "role": role,
                        "cue_side": side,
                        "ordinary": ordinary,
                        "intended_reference": source[role + "_reference"],
                    },
                    "slots": {
                        "left": {"start": 0, "end": len(left)},
                        "right": {
                            "start": right_start,
                            "end": right_start + len(right),
                        },
                        "ordinary": {"start": ordinary_start, "end": len(text)},
                    },
                    "provenance": dict(PROVENANCE),
                    "case_sha256": case_hash,
                    "text_sha256": digest(text.encode("utf-8")),
                    "version": VERSION,
                }
            pair_shared = {
                key: value for key, value in shared.items() if key != "scenario_domain"
            }
            specs = []
            for side, ordinary in itertools.product(SIDES, ORDINARY):
                specs.append(
                    (
                        f"{case_id}:role:{side}:{ordinary}",
                        "role",
                        ("alternative", side, ordinary),
                        ("attribute", side, ordinary),
                        "attribute - alternative",
                    )
                )
            for role, side in itertools.product(ROLES, SIDES):
                specs.append(
                    (
                        f"{case_id}:ordinary:{role}:{side}",
                        "ordinary",
                        (role, side, "a"),
                        (role, side, "b"),
                        "b - a",
                    )
                )
            for role, ordinary in itertools.product(ROLES, ORDINARY):
                specs.append(
                    (
                        f"{case_id}:cue_side:{role}:{ordinary}",
                        "cue_side",
                        (role, "left", ordinary),
                        (role, "right", ordinary),
                        "right - left",
                    )
                )
            for identifier, kind, base, changed, direction in specs:
                self.expected["pairs"][identifier] = {
                    **pair_shared,
                    "id": identifier,
                    "kind": kind,
                    "base_id": cell_ids[base],
                    "changed_id": cell_ids[changed],
                    "direction": direction,
                }
        for directory, count in (("cases", 256), ("data", 2048), ("pairs", 3072)):
            self.equal(
                len(self.expected[directory]),
                count,
                "reconstructed." + directory,
                "complete_reconstruction",
            )

    def load_records(self):
        schemas = {"cases": CASE_SCHEMA, "data": TEXT_SCHEMA, "pairs": PAIR_SCHEMA}
        for directory in ("cases", "data", "pairs"):
            key = "case_id" if directory == "cases" else "id"
            seen = {}
            for split in SPLITS:
                relative = f"{directory}/{split}.jsonl"
                blob = self.payloads.get(relative)
                if blob is None:
                    continue
                self.check(
                    bool(blob) and blob.endswith(b"\n") and b"\r" not in blob,
                    "jsonl_encoding",
                    relative,
                    "JSONL must be nonempty UTF-8 with LF after every row and no CR bytes",
                )
                try:
                    text = blob.decode("utf-8")
                except UnicodeError as exc:
                    self.check(
                        False, "jsonl_encoding", relative, f"invalid UTF-8: {exc}"
                    )
                    continue
                lines = text.split("\n")
                if lines[-1] == "":
                    lines.pop()
                identifiers = []
                for number, line in enumerate(lines, 1):
                    location = f"{relative}:{number}"
                    try:
                        value = decode_json(line)
                    except (ValueError, RecursionError) as exc:
                        self.check(
                            False, "jsonl_parse", location, f"invalid JSON row: {exc}"
                        )
                        continue
                    record = Record(value, split, location)
                    self.records[directory].append(record)
                    self.shape(value, schemas[directory], location)
                    try:
                        self.equal(
                            line, compact(value), location, "jsonl_canonical_bytes"
                        )
                    except (ValueError, UnicodeError) as exc:
                        self.check(
                            False,
                            "jsonl_canonical_bytes",
                            location,
                            f"cannot serialize canonical UTF-8: {exc}",
                        )
                    if not isinstance(value, dict):
                        continue
                    self.equal(
                        value.get("split"),
                        split,
                        location + ".split",
                        "file_split_membership",
                    )
                    identifier = value.get(key)
                    if not self.check(
                        type(identifier) is str and bool(identifier),
                        "record_identity",
                        location,
                        f"missing or invalid {key}",
                    ):
                        continue
                    identifiers.append(identifier)
                    self.check(
                        identifier not in seen,
                        "record_uniqueness",
                        location,
                        f"duplicate {key} {identifier!r}; first seen at {seen.get(identifier)}",
                    )
                    seen.setdefault(identifier, location)
                    expected = self.expected[directory].get(identifier)
                    if self.check(
                        expected is not None,
                        "record_identity",
                        location,
                        f"unexpected {key} {identifier!r}",
                    ):
                        self.equal(value, expected, location, "reconstructed_records")
                self.equal(identifiers, sorted(identifiers), relative, "record_sorting")
                expected_ids = sorted(
                    identifier
                    for identifier, row in self.expected[directory].items()
                    if row["split"] == split
                )
                self.equal(
                    sorted(identifiers),
                    expected_ids,
                    relative + ".identifiers",
                    "complete_record_inventory",
                )
            self.equal(
                sorted(seen),
                sorted(self.expected[directory]),
                directory + ".identifiers",
                "complete_record_inventory",
            )

    def text_invariants(self):
        texts = {}
        cells = defaultdict(Counter)
        targets = defaultdict(lambda: {"families": set(), "splits": set()})
        for record in self.records["data"]:
            row, location = record.value, record.location
            if not isinstance(row, dict):
                continue
            text, target, slots, construction = (
                row.get(key) for key in ("text", "target", "slots", "construction")
            )
            if not isinstance(text, str):
                continue
            self.check(
                text not in texts,
                "text_uniqueness",
                location,
                f"duplicate text; first seen at {texts.get(text)}",
            )
            texts.setdefault(text, location)
            try:
                self.equal(
                    row.get("text_sha256"),
                    digest(text.encode("utf-8")),
                    location + ".text_sha256",
                    "text_hashes",
                )
            except UnicodeError as exc:
                self.check(
                    False,
                    "text_hashes",
                    location,
                    f"text is not encodable as UTF-8: {exc}",
                )
            if isinstance(target, dict) and isinstance(target.get("text"), str):
                literal = target["text"]
                self.check(
                    one_line(literal),
                    "target_spans",
                    location,
                    "target text must be a nonempty single line without outer whitespace",
                )
                self.equal(
                    occurrences(text, literal),
                    1,
                    location + ".literal_occurrences",
                    "target_exclusivity",
                )
                self.equal(
                    occurrences(normalized(text), normalized(literal)),
                    1,
                    location + ".normalized_occurrences",
                    "target_exclusivity",
                )
                start, end = target.get("start"), target.get("end")
                if self.check(
                    type(start) is int
                    and type(end) is int
                    and 0 <= start < end <= len(text),
                    "target_spans",
                    location,
                    "target start/end must be valid nonempty Unicode codepoint offsets",
                ):
                    self.equal(
                        text[start:end],
                        literal,
                        location + ".target_slice",
                        "target_spans",
                    )
                    if isinstance(slots, dict):
                        expected_boundaries = {
                            "left": (0, start - 2),
                            "right": (end + 3, None),
                            "ordinary": (None, len(text)),
                        }
                        valid_slots = {}
                        for name in ("left", "right", "ordinary"):
                            span = slots.get(name)
                            if not isinstance(span, dict):
                                continue
                            begin, finish = span.get("start"), span.get("end")
                            if self.check(
                                type(begin) is int
                                and type(finish) is int
                                and 0 <= begin < finish <= len(text),
                                "slot_spans",
                                location + ".slots." + name,
                                "slot must be a nonempty valid Unicode codepoint span",
                            ):
                                valid_slots[name] = (begin, finish)
                                self.check(
                                    one_line(text[begin:finish])
                                    and text[begin:finish].endswith("."),
                                    "slot_spans",
                                    location + ".slots." + name,
                                    "span must extract one complete period-terminated sentence",
                                )
                                first, last = expected_boundaries[name]
                                if first is not None:
                                    self.equal(
                                        begin,
                                        first,
                                        location + ".slots." + name + ".start",
                                        "carrier",
                                    )
                                if last is not None:
                                    self.equal(
                                        finish,
                                        last,
                                        location + ".slots." + name + ".end",
                                        "carrier",
                                    )
                        self.equal(
                            text[start - 2 : start],
                            "\n\n",
                            location + ".before_target",
                            "carrier",
                        )
                        self.equal(
                            text[end : end + 3],
                            ".\n\n",
                            location + ".after_target",
                            "carrier",
                        )
                        if "right" in valid_slots and "ordinary" in valid_slots:
                            right_end = valid_slots["right"][1]
                            ordinary_start = valid_slots["ordinary"][0]
                            self.equal(
                                ordinary_start,
                                right_end + 2,
                                location + ".ordinary_start",
                                "carrier",
                            )
                            self.equal(
                                text[right_end:ordinary_start],
                                "\n\n",
                                location + ".before_ordinary",
                                "carrier",
                            )
                family, kind = row.get("family_id"), row.get("target_type")
                if isinstance(family, str) and isinstance(kind, str):
                    for key in target_keys(kind, literal):
                        targets[key]["families"].add(family)
                        targets[key]["splits"].add(record.split)
            case_id = row.get("case_id")
            if isinstance(construction, dict) and isinstance(case_id, str):
                factors = tuple(
                    construction.get(key) for key in ("role", "cue_side", "ordinary")
                )
                if all(isinstance(value, str) for value in factors):
                    cells[case_id][factors] += 1
            if isinstance(case_id, str) and case_id in self.sources:
                expected_hash = digest(compact(self.sources[case_id]).encode("utf-8"))
                self.equal(
                    row.get("case_sha256"),
                    expected_hash,
                    location + ".case_sha256",
                    "case_hashes",
                )
            self.equal(
                row.get("provenance"),
                PROVENANCE,
                location + ".provenance",
                "construction_provenance",
            )
            self.equal(
                row.get("version"), VERSION, location + ".version", "release_version"
            )
        grid = Counter({cell: 1 for cell in itertools.product(ROLES, SIDES, ORDINARY)})
        for case_id in sorted(set(self.sources) | set(cells)):
            self.equal(
                dict(cells[case_id]),
                dict(grid),
                case_id + ".factorial",
                "eight_cell_factorial",
            )
        self.equal(len(texts), 2048, "data.unique_texts", "text_uniqueness")
        for key, ownership in sorted(targets.items()):
            self.check(
                len(ownership["splits"]) == 1,
                "release_target_leakage",
                preview(key),
                f"normalized target appears in multiple file splits: {sorted(ownership['splits'])}",
            )
            self.check(
                len(ownership["families"]) == 1,
                "release_target_reuse",
                preview(key),
                f"normalized target appears in multiple families: {sorted(ownership['families'])}",
            )

    def case_and_pair_invariants(self):
        family_splits = defaultdict(set)
        for directory, records in self.records.items():
            for record in records:
                row = record.value
                if not isinstance(row, dict):
                    continue
                family = row.get("family_id")
                if isinstance(family, str):
                    family_splits[family].add(record.split)
                    if family in self.split_for:
                        self.equal(
                            record.split,
                            self.split_for[family],
                            record.location,
                            "split_sha_ranking",
                        )
                    else:
                        self.check(
                            False,
                            "split_sha_ranking",
                            record.location,
                            f"unknown family {family!r}",
                        )
                if directory == "cases" and all(
                    type(row.get(key)) is str for key in SOURCE_FIELDS
                ):
                    source_view = {key: row[key] for key in SOURCE_FIELDS}
                    try:
                        self.equal(
                            row.get("case_sha256"),
                            digest(compact(source_view).encode("utf-8")),
                            record.location + ".case_sha256",
                            "case_hashes",
                        )
                    except UnicodeError as exc:
                        self.check(
                            False,
                            "case_hashes",
                            record.location,
                            f"case source view is not UTF-8: {exc}",
                        )
        for family, splits in sorted(family_splits.items()):
            self.check(
                len(splits) == 1,
                "family_split_isolation",
                family,
                f"family crosses file splits: {sorted(splits)}",
            )
        by_id = {}
        for record in self.records["data"]:
            row = record.value
            if isinstance(row, dict) and isinstance(row.get("id"), str):
                by_id.setdefault(row["id"], row)
        pair_counts = defaultdict(Counter)
        pair_edges = defaultdict(Counter)
        expected_order = {
            "role": ("alternative", "attribute"),
            "ordinary": ("a", "b"),
            "cue_side": ("left", "right"),
        }
        directions = {
            "role": "attribute - alternative",
            "ordinary": "b - a",
            "cue_side": "right - left",
        }
        for record in self.records["pairs"]:
            pair, location = record.value, record.location
            if not isinstance(pair, dict):
                continue
            kind, case_id = pair.get("kind"), pair.get("case_id")
            if isinstance(kind, str) and isinstance(case_id, str):
                pair_counts[case_id][kind] += 1
            base_id, changed_id = pair.get("base_id"), pair.get("changed_id")
            if not isinstance(base_id, str) or not isinstance(changed_id, str):
                continue
            if isinstance(case_id, str):
                pair_edges[case_id][(base_id, changed_id)] += 1
            base, changed = by_id.get(base_id), by_id.get(changed_id)
            if not self.check(
                base is not None and changed is not None,
                "pair_references",
                location,
                f"pair endpoints must exist: base={base_id!r}, changed={changed_id!r}",
            ):
                continue
            for field in ("case_id", "family_id", "language", "target_type", "split"):
                self.equal(
                    base.get(field),
                    pair.get(field),
                    location + ".base." + field,
                    "pair_identity",
                )
                self.equal(
                    changed.get(field),
                    pair.get(field),
                    location + ".changed." + field,
                    "pair_identity",
                )
            self.equal(
                base.get("target", {}).get("text")
                if isinstance(base.get("target"), dict)
                else None,
                changed.get("target", {}).get("text")
                if isinstance(changed.get("target"), dict)
                else None,
                location + ".target_literal",
                "pair_fixed_factors",
            )
            a, b = base.get("construction"), changed.get("construction")
            if (
                not isinstance(a, dict)
                or not isinstance(b, dict)
                or not isinstance(kind, str)
            ):
                continue
            if self.check(
                kind in KINDS, "pair_kind", location, f"unknown contrast kind {kind!r}"
            ):
                differences = [
                    factor
                    for factor in ("role", "cue_side", "ordinary")
                    if a.get(factor) != b.get(factor)
                ]
                self.equal(
                    differences,
                    [kind],
                    location + ".changed_factors",
                    "pair_fixed_factors",
                )
                self.equal(
                    [a.get(kind), b.get(kind)],
                    list(expected_order[kind]),
                    location + ".factor_order",
                    "pair_direction",
                )
                self.equal(
                    pair.get("direction"),
                    directions[kind],
                    location + ".direction",
                    "pair_direction",
                )
        for case_id in sorted(set(self.sources) | set(pair_counts)):
            self.equal(
                dict(pair_counts[case_id]),
                {kind: 4 for kind in KINDS},
                case_id + ".contrasts",
                "twelve_pair_factorial",
            )
            self.equal(
                len(pair_edges[case_id]),
                12,
                case_id + ".unique_edges",
                "twelve_pair_factorial",
            )
            for edge, count in pair_edges[case_id].items():
                self.equal(
                    count,
                    1,
                    case_id + ".edge." + preview(edge),
                    "twelve_pair_factorial",
                )

    def statistics(self):
        def dimension(record, key):
            value = record.value.get(key) if isinstance(record.value, dict) else None
            return value if isinstance(value, str) else "<invalid>"

        def measures(texts, cases, pairs, families=True):
            result = {
                "records": len(texts),
                "cases": len(cases),
                "contrasts": len(pairs),
            }
            if families:
                result["families"] = len(
                    {dimension(record, "family_id") for record in cases}
                )
            return result

        texts, cases, pairs = (self.records[key] for key in ("data", "cases", "pairs"))
        by_split = {}
        for split in SPLITS:
            by_split[split] = measures(
                [record for record in texts if record.split == split],
                [record for record in cases if record.split == split],
                [record for record in pairs if record.split == split],
            )
        grouped = {}
        for field, expected_values in (("target_type", TYPES), ("language", LANGUAGES)):
            values = set(expected_values)
            values.update(dimension(record, field) for record in texts + cases + pairs)
            grouped[field] = {
                value: measures(
                    [record for record in texts if dimension(record, field) == value],
                    [record for record in cases if dimension(record, field) == value],
                    [record for record in pairs if dimension(record, field) == value],
                    families=field == "target_type",
                )
                for value in sorted(values)
            }
        kinds = Counter(dimension(record, "kind") for record in pairs)
        contrast_counts = {
            kind: kinds[kind] for kind in sorted(set(KINDS) | set(kinds))
        }
        lengths = [
            len(record.value["text"])
            for record in texts
            if isinstance(record.value, dict)
            and isinstance(record.value.get("text"), str)
        ]
        totals = measures(texts, cases, pairs)
        self.observed = {
            "schema_version": 1,
            "name": NAME,
            "version": VERSION,
            **totals,
            "gold_privacy_labels": 0,
            "by_split": by_split,
            "by_target_type": grouped["target_type"],
            "by_language": grouped["language"],
            "by_contrast_kind": contrast_counts,
            "text_codepoint_length": {
                "min": min(lengths) if lengths else None,
                "max": max(lengths) if lengths else None,
            },
        }
        self.equal(totals, TOTALS, "observed.totals", "counts")
        self.equal(by_split, SPLIT_TOTALS, "observed.by_split", "counts")
        self.equal(
            grouped["target_type"],
            {kind: TYPE_TOTALS for kind in TYPES},
            "observed.by_target_type",
            "counts",
        )
        self.equal(
            grouped["language"],
            {language: LANGUAGE_TOTALS for language in LANGUAGES},
            "observed.by_language",
            "counts",
        )
        self.equal(
            contrast_counts,
            {kind: 1024 for kind in KINDS},
            "observed.by_contrast_kind",
            "counts",
        )
        blob = self.payloads.get("stats.json")
        document = self.parse(blob, "stats.json")
        if document is not None:
            self.equal(document, self.observed, "stats.json", "recomputed_statistics")
            self.canonical_document(blob, document, "stats.json")

    def canonical_document(self, blob, document, location):
        try:
            expected = pretty_bytes(document)
            self.check(
                blob == expected,
                "pretty_json_bytes",
                location,
                "JSON must use indent=2, sorted keys, ensure_ascii=False, UTF-8, and one trailing LF",
            )
        except (ValueError, UnicodeError, RecursionError) as exc:
            self.check(
                False,
                "pretty_json_bytes",
                location,
                f"cannot serialize canonical UTF-8 JSON: {exc}",
            )

    def file_entry(self, relative, blob):
        return {"path": relative, "bytes": len(blob), "sha256": digest(blob)}

    def manifest_entries(self, entries, scope, expected_paths, blobs, location):
        if not self.check(
            type(entries) is list,
            "manifest_inventory",
            location,
            "expected a list of file entries",
        ):
            return
        paths = []
        for index, entry in enumerate(entries):
            item_location = f"{location}[{index}]"
            if not self.shape(
                entry, {"path": str, "bytes": int, "sha256": str}, item_location
            ):
                continue
            relative = entry["path"]
            paths.append(relative)
            posix = PurePosixPath(relative)
            valid_path = (
                not posix.is_absolute()
                and relative == posix.as_posix()
                and "\\" not in relative
                and "\x00" not in relative
                and all(part not in {"", ".", ".."} for part in posix.parts)
            )
            if not self.check(
                valid_path,
                "manifest_paths",
                item_location,
                "path must be a normalized relative POSIX path without traversal",
            ):
                continue
            try:
                self.check(
                    (scope / relative).resolve().is_relative_to(scope),
                    "manifest_paths",
                    item_location,
                    "declared path resolves outside its declared root",
                )
            except (OSError, RuntimeError, ValueError) as exc:
                self.check(
                    False,
                    "manifest_paths",
                    item_location,
                    f"cannot resolve declared path: {exc}",
                )
            self.check(
                entry["bytes"] >= 0,
                "manifest_hashes",
                item_location + ".bytes",
                "byte count cannot be negative",
            )
            self.check(
                bool(re.fullmatch(r"[0-9a-f]{64}", entry["sha256"])),
                "manifest_hashes",
                item_location + ".sha256",
                "SHA256 must be 64 lowercase hexadecimal characters",
            )
            if self.check(
                relative in expected_paths,
                "manifest_inventory",
                item_location,
                f"undeclared-contract path {relative!r} is not permitted",
            ):
                if relative in blobs:
                    self.equal(
                        entry,
                        self.file_entry(relative, blobs[relative]),
                        item_location,
                        "manifest_hashes",
                    )
                else:
                    self.check(
                        False,
                        "manifest_hashes",
                        item_location,
                        "expected file was unavailable; cannot verify bytes/hash",
                    )
        self.equal(
            paths, sorted(expected_paths), location + ".paths", "manifest_inventory"
        )

    def manifests(self):
        self.manifest_blob = self.read_blob(
            self.release, "manifest.json", "release_files"
        )
        checksum_blob = self.read_blob(self.release, "manifest.sha256", "release_files")
        manifest = self.parse(self.manifest_blob, "manifest.json")
        if self.manifest_blob is not None and checksum_blob is not None:
            expected_checksum = (
                digest(self.manifest_blob) + "  manifest.json\n"
            ).encode("ascii")
            self.check(
                checksum_blob == expected_checksum,
                "manifest_checksum",
                "manifest.sha256",
                "expected exactly SHA256(manifest.json), two spaces, manifest.json, and LF",
            )
        if manifest is None:
            return
        self.shape(
            manifest,
            {
                "schema_version": int,
                "name": str,
                "version": str,
                "protocol_sha256": str,
                "inputs": list,
                "files": list,
                "counts": {
                    "records": int,
                    "cases": int,
                    "families": int,
                    "contrasts": int,
                },
            },
            "manifest.json",
        )
        self.canonical_document(self.manifest_blob, manifest, "manifest.json")
        if not isinstance(manifest, dict):
            return
        self.manifest_entries(
            manifest.get("inputs"),
            self.root,
            INPUT_PATHS,
            self.inputs,
            "manifest.json.inputs",
        )
        self.manifest_entries(
            manifest.get("files"),
            self.release,
            PAYLOAD_PATHS,
            self.payloads,
            "manifest.json.files",
        )
        expected = {
            "schema_version": 1,
            "name": NAME,
            "version": VERSION,
            "protocol_sha256": PROTOCOL_SHA256,
            "inputs": [
                self.file_entry(path, self.inputs[path])
                for path in INPUT_PATHS
                if path in self.inputs
            ],
            "files": [
                self.file_entry(path, self.payloads[path])
                for path in PAYLOAD_PATHS
                if path in self.payloads
            ],
            "counts": dict(TOTALS),
        }
        self.equal(manifest, expected, "manifest.json", "recomputed_manifest")
        if self.observed:
            self.equal(
                manifest.get("counts"),
                {key: self.observed[key] for key in TOTALS},
                "manifest.json.counts",
                "manifest_observed_counts",
            )

    def run(self):
        self.inventories()
        self.source_cases()
        self.reconstruct()
        self.load_records()
        self.text_invariants()
        self.case_and_pair_invariants()
        self.statistics()
        self.manifests()

    def report(self):
        manifest_identity = {
            "path": str(self.release / "manifest.json"),
            "bytes": None,
            "sha256": None,
        }
        if self.manifest_blob is not None:
            manifest_identity.update(
                bytes=len(self.manifest_blob), sha256=digest(self.manifest_blob)
            )
        protocol_blob = self.inputs.get(".construction/protocol.json")
        return {
            "schema_version": 1,
            "name": NAME,
            "version": VERSION,
            "timestamp": datetime.now(timezone.utc).isoformat(),
            "producer": self.producer,
            "root": {
                "path": str(self.root),
                "protocol_sha256": digest(protocol_blob)
                if protocol_blob is not None
                else None,
            },
            "release": {
                "path": str(self.release),
                "manifest_sha256": manifest_identity["sha256"],
            },
            "manifest": manifest_identity,
            "status": "failure" if self.errors else "pass",
            "counts": {
                "expected": dict(TOTALS),
                "source_cases_reconstructed": len(self.sources),
                "parsed_release_rows": {
                    key: len(value) for key, value in self.records.items()
                },
                "observed_statistics": self.observed,
                "errors": len(self.errors),
            },
            "checks": {
                name: {**summary, "status": "failure" if summary["failed"] else "pass"}
                for name, summary in sorted(self.checks.items())
            },
            "safety_inventories": {
                kind: sorted(items, key=lambda item: item["case_id"])
                for kind, items in self.safety.items()
            },
            "errors": self.errors,
            "limitations": list(LIMITATIONS),
        }


def main(argv=None):
    parser = argparse.ArgumentParser(
        description="Validate the complete frozen construction corpus without executing a builder, model, or tokenizer.",
        epilog="A pass is mechanical validation only: not semantic/privacy adjudication, representative sampling, tokenizer matching, or classifier evaluation.",
    )
    parser.add_argument(
        "--root",
        type=Path,
        default=Path(__file__).resolve().parent.parent,
        help="source repository root (default: this script's repository)",
    )
    parser.add_argument(
        "--release",
        type=Path,
        required=True,
        help="existing generated release directory; read-only",
    )
    parser.add_argument(
        "--report",
        type=Path,
        required=True,
        help="new report JSON path outside the release; parent must exist",
    )
    args = parser.parse_args(argv)
    try:
        root = args.root.expanduser().resolve()
        release = args.release.expanduser().resolve()
        requested_report = args.report.expanduser()
        if os.path.lexists(requested_report):
            raise ValueError(
                f"report already exists; refusing to overwrite: {requested_report}"
            )
        report_path = requested_report.resolve()
        if report_path.is_relative_to(release):
            raise ValueError(f"report must be outside release: {report_path}")
        if not report_path.parent.is_dir():
            raise ValueError(
                f"report parent directory does not exist: {report_path.parent}"
            )
        producer_path = Path(__file__).resolve(strict=True)
        producer_blob = producer_path.read_bytes()
        producer = {
            "path": str(producer_path),
            "bytes": len(producer_blob),
            "sha256": digest(producer_blob),
        }
        # Exclusive creation closes the overwrite race. No release or input file
        # is opened for writing, and no report directory is implicitly created.
        report_stream = report_path.open("x", encoding="utf-8", newline="\n")
    except (OSError, RuntimeError, ValueError) as exc:
        print(f"validation refused: {exc}", file=sys.stderr)
        return 1
    validator = Validator(root, release, producer)
    with report_stream:
        try:
            validator.run()
        except Exception as exc:
            validator.check(
                False,
                "validator_internal_error",
                "validator",
                f"{type(exc).__name__}: {exc}\n{traceback.format_exc()}",
            )
        report = validator.report()
        try:
            # Escaping report strings also preserves failures containing an
            # invalid surrogate from corrupt input. Release serialization is
            # checked separately against ensure_ascii=False.
            json.dump(
                report,
                report_stream,
                indent=2,
                sort_keys=True,
                ensure_ascii=True,
                allow_nan=False,
            )
            report_stream.write("\n")
            report_stream.flush()
            os.fsync(report_stream.fileno())
        except (OSError, UnicodeError, ValueError) as exc:
            print(
                f"cannot persist complete validation report at {report_path}: {exc}",
                file=sys.stderr,
            )
            return 1
    if validator.errors:
        print(
            f"FAIL: {len(validator.errors)} validation errors. Complete retained report: {report_path}",
            file=sys.stderr,
        )
        for error in validator.errors[:20]:
            print(
                f"  [{error['check']}] {error['location']}: {error['message']}",
                file=sys.stderr,
            )
        if len(validator.errors) > 20:
            print(
                f"  {len(validator.errors) - 20} additional errors retained in the report.",
                file=sys.stderr,
            )
        return 1
    print(
        f"PASS: 256 cases, 128 families, 2048 texts, 3072 contrasts. Report: {report_path}"
    )
    print(
        "Mechanical validation only; no semantic/privacy gold, representativeness, tokenizer matching, or classifier evaluation."
    )
    return 0


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