ctx / scripts /ctx_ab_holdout.py
Stevesolun's picture
Sync ctx eea3f60 (part 2)
81b05f3 verified
Raw
History Blame Contribute Delete
43.5 kB
#!/usr/bin/env python3
"""Deterministically filter and select a private CTX benchmark holdout."""
from __future__ import annotations
import argparse
import ast
import csv
from collections import defaultdict
import hashlib
import hmac
import json
import math
import os
import re
import stat
import statistics
from pathlib import Path, PurePosixPath
from typing import Any, TextIO
try:
from scripts import ctx_ab_exposure_ledger as exposure_ledger
except ImportError: # pragma: no cover - direct script execution
import ctx_ab_exposure_ledger as exposure_ledger # type: ignore[no-redef]
ROOT = Path(__file__).resolve().parents[1]
DEFAULT_PROTOCOL = ROOT / "benchmarks" / "ctx_ab" / "holdout-protocol-v1.json"
PRIVATE_ROOT = ROOT / ".gate" / "ctx-ab-private"
V2_CANDIDATE_PARTITION_PREFIX = b"ctx-holdout-candidate-partition-v2\0"
HISTORICAL_EXPOSURE_REJECTION_CODE = "historical-exposure"
SHA1 = re.compile(r"^[0-9a-f]{40}$")
SHA256 = re.compile(r"^[0-9a-f]{64}$")
DIFF_PATH = re.compile(r"^diff --git a/(.+) b/(.+)$")
HUNK_HEADER = re.compile(r"^@@ -\d+(?:,(\d+))? \+\d+(?:,(\d+))? @@(?: .*)?$")
FORBIDDEN_TEST_IMPORT_ROOTS = {
"aiohttp",
"http",
"httpx",
"random",
"requests",
"socket",
"subprocess",
"urllib",
"urllib3",
}
LEDGER_FIELDS = (
"instance_id",
"repo",
"base_commit",
"production_paths",
"test_path",
"production_changed_lines",
"status",
"rejection_code",
)
def _digest(*values: str) -> str:
return hashlib.sha256("\0".join(values).encode()).hexdigest()
def canonical_repo_url(repo: str) -> str:
normalized = repo.strip().lower()
if not re.fullmatch(r"[a-z0-9_.-]+/[a-z0-9_.-]+", normalized):
raise ValueError(f"invalid repository name: {repo!r}")
if any(part in {".", ".."} for part in normalized.split("/")):
raise ValueError(f"invalid repository name: {repo!r}")
return f"https://github.com/{normalized}.git"
def _is_canonical_repo_url(value: object) -> bool:
if not isinstance(value, str) or not value.startswith("https://github.com/"):
return False
repo = value.removeprefix("https://github.com/").removesuffix(".git")
try:
return canonical_repo_url(repo) == value
except ValueError:
return False
def _parse_patch(patch: str) -> tuple[tuple[str, ...], int, str]:
lines = patch.splitlines()
paths: list[str] = []
changed_lines = 0
added_lines: list[str] = []
index = 0
while index < len(lines):
match = DIFF_PATH.fullmatch(lines[index])
if match is None:
return (), -1, ""
before, after = match.groups()
if before != after:
return (), -1, ""
path = PurePosixPath(after)
if path.is_absolute() or ".." in path.parts or path.as_posix() in paths:
return (), -1, ""
paths.append(path.as_posix())
index += 1
while index < len(lines) and not lines[index].startswith("--- "):
if not lines[index].startswith(
("index ", "new file mode ", "deleted file mode ", "old mode ", "new mode ")
):
return (), -1, ""
index += 1
if index + 1 >= len(lines):
return (), -1, ""
old_header = lines[index][4:]
new_header = lines[index + 1][4:] if lines[index + 1].startswith("+++ ") else ""
if (
old_header not in {f"a/{after}", "/dev/null"}
or new_header not in {f"b/{after}", "/dev/null"}
or old_header == new_header == "/dev/null"
):
return (), -1, ""
index += 2
saw_hunk = False
while index < len(lines) and not lines[index].startswith("diff --git "):
hunk = HUNK_HEADER.fullmatch(lines[index])
if hunk is None:
return (), -1, ""
saw_hunk = True
old_expected = int(hunk.group(1) or 1)
new_expected = int(hunk.group(2) or 1)
old_seen = 0
new_seen = 0
index += 1
while (
index < len(lines)
and not lines[index].startswith("diff --git ")
and not lines[index].startswith("@@ ")
):
line = lines[index]
if line == r"\ No newline at end of file":
index += 1
continue
if not line or line[0] not in " +-":
return (), -1, ""
if line[0] in " -":
old_seen += 1
if line[0] in " +":
new_seen += 1
if line[0] in "+-":
changed_lines += 1
if line[0] == "+":
added_lines.append(line[1:])
index += 1
if old_seen != old_expected or new_seen != new_expected:
return (), -1, ""
if not saw_hunk:
return (), -1, ""
return tuple(paths), changed_lines, "\n".join(added_lines)
def _is_test_path(path: str) -> bool:
pure = PurePosixPath(path)
return path.endswith(".py") and (
pure.name.startswith("test_") or any(part in {"test", "tests"} for part in pure.parts)
)
def _is_product_path(path: str, rules: dict[str, Any]) -> bool:
pure = PurePosixPath(path)
if not path.endswith(".py") or _is_test_path(path):
return False
if pure.name in rules["excluded_filenames"]:
return False
if any(re.search(pattern, pure.name) for pattern in rules["excluded_filename_regex"]):
return False
return not any(part in rules["excluded_path_components"] for part in pure.parts[:-1])
def evaluate_row(row: dict[str, Any], protocol: dict[str, Any]) -> dict[str, Any]:
repo = str(row.get("repo") or "").strip().lower()
instance_id = str(row.get("instance_id") or "").strip()
base_commit = str(row.get("base_commit") or "").strip()
patch = str(row.get("patch") or "")
test_patch = str(row.get("test_patch") or "")
problem_statement = str(row.get("problem_statement") or "")
production_paths, changed_lines, _ = _parse_patch(patch)
test_paths, _, added_test = _parse_patch(test_patch)
rules = protocol["static_candidate_rules"]
rejection = ""
try:
canonical_repo_url(repo)
except ValueError:
rejection = "row-schema"
if not rejection and (not instance_id or not re.fullmatch(r"[A-Za-z0-9_.-]+", instance_id)):
rejection = "row-schema"
elif not rejection and repo in protocol["excluded_repositories"]:
rejection = "excluded-repository"
elif not rejection and not SHA1.fullmatch(base_commit):
rejection = "base-commit"
elif not rejection and (
not 1 <= len(production_paths) <= 3
or not all(_is_product_path(path, rules) for path in production_paths)
):
rejection = "patch-paths"
elif not rejection and (len(test_paths) != 1 or not _is_test_path(test_paths[0])):
rejection = "test-paths"
elif not rejection and not (
rules["production_changed_lines"]["minimum"]
<= changed_lines
<= rules["production_changed_lines"]["maximum"]
):
rejection = "patch-lines"
elif not rejection and not (
rules["problem_statement_words"]["minimum"]
<= len(problem_statement.split())
<= rules["problem_statement_words"]["maximum"]
):
rejection = "problem-statement"
elif not rejection and any(
re.search(pattern, added_test) for pattern in rules["forbidden_test_regex"]
):
rejection = "test-dependency"
return {
"instance_id": instance_id,
"repo": repo,
"base_commit": base_commit,
"production_paths": "|".join(production_paths),
"test_path": test_paths[0] if len(test_paths) == 1 else "",
"production_changed_lines": changed_lines,
"status": "eligible" if not rejection else "rejected",
"rejection_code": rejection,
}
def reject_historical_exposures(
ledger: list[dict[str, Any]],
exposure_document: dict[str, Any],
) -> list[dict[str, Any]]:
validated = exposure_ledger.validate_ledger_document(exposure_document)
exposed = set(validated["instance_id_hmac_sha256"])
salt = str(validated["salt"])
filtered: list[dict[str, Any]] = []
for original in ledger:
row = dict(original)
if row.get("status") == "eligible":
digest = exposure_ledger.instance_id_hmac_sha256(
salt,
str(row.get("instance_id") or ""),
)
if digest in exposed:
row["status"] = "rejected"
row["rejection_code"] = HISTORICAL_EXPOSURE_REJECTION_CODE
filtered.append(row)
return filtered
def require_exposure_disjoint_selection(
selection: dict[str, Any],
exposure_document: dict[str, Any],
) -> None:
validated = exposure_ledger.validate_ledger_document(exposure_document)
identities = selection.get("analysis_instance_ids")
canary = selection.get("canary_instance_id")
if not isinstance(identities, list):
raise ValueError("selection is invalid")
selected = list(identities)
if canary is not None:
selected.append(canary)
if any(
not isinstance(instance_id, str)
or exposure_ledger.contains_instance_id(validated, instance_id)
for instance_id in selected
):
raise ValueError("selection intersects the authenticated exposure ledger")
def validate_reconstructed_test_module(source: str) -> None:
"""Reject external, nondeterministic, environment, and sleep dependencies."""
try:
tree = ast.parse(source)
except SyntaxError as exc:
raise ValueError("reconstructed test module is not valid Python") from exc
aliases: dict[str, str] = {}
for node in ast.walk(tree):
if isinstance(node, ast.Import):
for imported in node.names:
local = imported.asname or imported.name.split(".", 1)[0]
aliases[local] = (
imported.name if imported.asname else imported.name.split(".", 1)[0]
)
if imported.name.split(".", 1)[0] in FORBIDDEN_TEST_IMPORT_ROOTS:
raise ValueError("reconstructed test module has a forbidden import")
elif isinstance(node, ast.ImportFrom) and node.module:
root = node.module.split(".", 1)[0]
if root in FORBIDDEN_TEST_IMPORT_ROOTS:
raise ValueError("reconstructed test module has a forbidden import")
for imported in node.names:
if imported.name == "*" and root in {"builtins", "importlib", "os", "time"}:
raise ValueError("reconstructed test module has a forbidden wildcard import")
local = imported.asname or imported.name
aliases[local] = f"{node.module}.{imported.name}"
if (node.module, imported.name) in {
("os", "environ"),
("os", "getenv"),
("time", "sleep"),
}:
raise ValueError("reconstructed test module has a forbidden dependency")
def qualified_name(node: ast.AST) -> str:
if isinstance(node, ast.Name):
return aliases.get(node.id, node.id)
if isinstance(node, ast.Attribute):
parent = qualified_name(node.value)
return f"{parent}.{node.attr}" if parent else node.attr
return ""
for node in ast.walk(tree):
if isinstance(node, ast.Attribute) and qualified_name(node) in {
"os.environ",
"os.getenv",
"time.sleep",
}:
raise ValueError("reconstructed test module has a forbidden dependency")
if not isinstance(node, ast.Call):
continue
name = qualified_name(node.func)
if name.rsplit(".", 1)[-1] == "sleep":
raise ValueError("reconstructed test module has a forbidden dependency")
if name in {
"__import__",
"builtins.__import__",
"importlib.import_module",
}:
if not node.args:
raise ValueError("reconstructed test module has a forbidden dynamic import")
argument = node.args[0]
if not isinstance(argument, ast.Constant) or not isinstance(argument.value, str):
raise ValueError("reconstructed test module has a forbidden dynamic import")
root = argument.value.split(".", 1)[0]
if root in FORBIDDEN_TEST_IMPORT_ROOTS | {"os", "time"}:
raise ValueError("reconstructed test module has a forbidden import")
def select_rows(ledger: list[dict[str, Any]], protocol: dict[str, Any]) -> dict[str, Any]:
seed = str(protocol["selection_seed"])
selection_rules = protocol["selection"]
strategy = selection_rules.get("strategy", "legacy")
legacy_strategy = strategy is None or strategy == "legacy"
private_canary = selection_rules.get("private_canary", legacy_strategy)
required_repositories = int(selection_rules["eligible_repositories_required"])
candidates_per_repository = int(selection_rules["eligible_candidates_per_repository_required"])
analysis_repositories = int(selection_rules["analysis_repositories"])
analysis_scenarios = int(selection_rules["analysis_scenarios"])
candidate_seed = seed
candidate_slot = 0
if not isinstance(private_canary, bool):
raise ValueError("selection.private_canary must be a boolean")
if legacy_strategy:
if (
not private_canary
or required_repositories != analysis_repositories + 1
or analysis_scenarios != analysis_repositories + 1
):
raise ValueError("holdout selection cardinalities are inconsistent")
elif strategy == "one-per-repository":
if "private_canary" not in selection_rules:
raise ValueError("one-per-repository selection requires explicit private_canary")
if protocol.get("schema_version") == 2:
generation = protocol.get("protocol_generation")
dataset_revision = str(protocol.get("universe", {}).get("revision") or "")
candidate_seed = str(protocol.get("candidate_partition_seed") or "")
candidate_slot_value = selection_rules.get("candidate_slot")
expected_candidate_seed = (
hashlib.sha256(
V2_CANDIDATE_PARTITION_PREFIX + dataset_revision.encode("ascii")
).hexdigest()
if SHA1.fullmatch(dataset_revision)
else ""
)
if (
not isinstance(generation, int)
or isinstance(generation, bool)
or generation < 1
or not isinstance(candidate_slot_value, int)
or isinstance(candidate_slot_value, bool)
or candidate_slot_value != generation - 1
or candidates_per_repository != generation
or not SHA256.fullmatch(candidate_seed)
or not hmac.compare_digest(candidate_seed, expected_candidate_seed)
):
raise ValueError("V2 candidate partition contract is invalid")
candidate_slot = candidate_slot_value
expected_repositories = analysis_repositories + (1 if private_canary else 0)
if (
analysis_repositories < 1
or analysis_scenarios != analysis_repositories
or required_repositories != expected_repositories
or candidates_per_repository < 1
):
raise ValueError(
"one-per-repository selection requires analysis_scenarios equal to "
"analysis_repositories, eligible_repositories_required equal to "
"analysis_repositories plus one when private_canary is true, and "
"at least one eligible candidate per repository"
)
else:
raise ValueError(f"unsupported holdout selection strategy: {strategy!r}")
instance_ids = [str(row.get("instance_id") or "") for row in ledger]
if len(instance_ids) != len(set(instance_ids)):
raise ValueError("candidate ledger contains duplicate instance IDs")
eligible_by_repo: dict[str, list[dict[str, Any]]] = {}
for row in ledger:
if row.get("status") == "eligible":
eligible_by_repo.setdefault(str(row["repo"]), []).append(row)
ranked_repositories = sorted(
(
(_digest(seed, canonical_repo_url(repo)), canonical_repo_url(repo), repo)
for repo, rows in eligible_by_repo.items()
if len(rows) >= candidates_per_repository
),
key=lambda item: (item[0], item[1]),
)
if len(ranked_repositories) < required_repositories:
count = (
"zero",
"one",
"two",
"three",
"four",
"five",
"six",
"seven",
"eight",
"nine",
"ten",
)
required_label = (
count[required_repositories]
if required_repositories < len(count)
else str(required_repositories)
)
candidate_label = (
count[candidates_per_repository]
if 0 <= candidates_per_repository < len(count)
else str(candidates_per_repository)
)
raise ValueError(
f"holdout requires {required_label} repositories with at least "
f"{candidate_label} eligible rows"
)
selected_repositories = ranked_repositories[:required_repositories]
def ranked(repo: str) -> list[dict[str, Any]]:
return sorted(
eligible_by_repo[repo],
key=lambda row: (
_digest(candidate_seed, str(row["instance_id"])),
str(row["instance_id"]),
),
)
analysis = [
ranked(repo)[candidate_slot] for _, _, repo in selected_repositories[:analysis_repositories]
]
if legacy_strategy:
first_repo_rows = ranked(selected_repositories[0][2])
occupied = {
*str(first_repo_rows[0]["production_paths"]).split("|"),
str(first_repo_rows[0]["test_path"]),
}
second = next(
(
row
for row in first_repo_rows[1:]
if occupied.isdisjoint(
{
*str(row["production_paths"]).split("|"),
str(row["test_path"]),
}
)
),
None,
)
if second is None:
raise ValueError("first ranked repository has no disjoint second candidate")
analysis.append(second)
canary_id: str | None = None
canary_url: str | None = None
if private_canary:
canary = ranked(selected_repositories[analysis_repositories][2])[candidate_slot]
analysis_urls = {canonical_repo_url(str(row["repo"])) for row in analysis}
canary_url = canonical_repo_url(str(canary["repo"]))
canary_id = str(canary["instance_id"])
if canary_url in analysis_urls or canary_id in {
str(row["instance_id"]) for row in analysis
}:
raise ValueError("canary must be disjoint from analysis selections")
return {
"protocol_id": protocol["protocol_id"],
"analysis_instance_ids": [row["instance_id"] for row in analysis],
"analysis_repository_map": {
str(row["instance_id"]): canonical_repo_url(str(row["repo"])) for row in analysis
},
"canary_instance_id": canary_id,
"canary_repository": canary_url,
}
def _canonical_json_bytes(value: Any) -> bytes:
return json.dumps(value, sort_keys=True, separators=(",", ":")).encode()
def _validated_selection(
selection: dict[str, Any],
protocol: dict[str, Any],
) -> tuple[list[str], dict[str, str]]:
selection_rules = protocol["selection"]
strategy = selection_rules.get("strategy", "legacy")
legacy_strategy = strategy is None or strategy == "legacy"
private_canary = selection_rules.get("private_canary", legacy_strategy)
analysis_repositories = int(selection_rules["analysis_repositories"])
required_repositories = int(selection_rules["eligible_repositories_required"])
analysis_scenarios = int(selection_rules["analysis_scenarios"])
candidates_per_repository = int(selection_rules["eligible_candidates_per_repository_required"])
if not isinstance(private_canary, bool):
raise ValueError("selection.private_canary must be a boolean")
if legacy_strategy:
if (
not private_canary
or required_repositories != analysis_repositories + 1
or analysis_scenarios != analysis_repositories + 1
):
raise ValueError("claim selection cardinalities are inconsistent")
elif strategy == "one-per-repository":
if "private_canary" not in selection_rules:
raise ValueError("claim selection requires explicit private_canary")
expected_repositories = analysis_repositories + (1 if private_canary else 0)
if (
analysis_repositories < 1
or analysis_scenarios != analysis_repositories
or required_repositories != expected_repositories
or candidates_per_repository < 1
):
raise ValueError("claim selection cardinalities are inconsistent")
else:
raise ValueError(f"unsupported holdout selection strategy: {strategy!r}")
analysis_ids = selection.get("analysis_instance_ids")
repository_map = selection.get("analysis_repository_map")
canary_id = selection.get("canary_instance_id")
canary_repository = selection.get("canary_repository")
if (
selection.get("protocol_id") != protocol["protocol_id"]
or not isinstance(analysis_ids, list)
or not all(isinstance(value, str) and value for value in analysis_ids)
or len(analysis_ids) != analysis_scenarios
or len(set(analysis_ids)) != len(analysis_ids)
or not isinstance(repository_map, dict)
or set(repository_map) != set(analysis_ids)
or not all(
isinstance(key, str) and isinstance(value, str) and _is_canonical_repo_url(value)
for key, value in repository_map.items()
)
or (
strategy == "one-per-repository"
and len(set(repository_map.values())) != analysis_repositories
)
):
raise ValueError("claim selection is invalid")
analysis_repository_values = set(repository_map.values())
if private_canary:
if (
not isinstance(canary_id, str)
or not canary_id
or canary_id in repository_map
or not isinstance(canary_repository, str)
or not _is_canonical_repo_url(canary_repository)
or canary_repository in analysis_repository_values
):
raise ValueError("claim selection is invalid")
return [*analysis_ids, canary_id], dict(repository_map)
if canary_id is not None or canary_repository is not None:
raise ValueError("claim selection is invalid")
return [*analysis_ids], dict(repository_map)
def build_reconstructed_test_attestation(
selection: dict[str, Any],
protocol: dict[str, Any],
reconstructed_tests: dict[str, str],
) -> dict[str, Any]:
selected_ids, _ = _validated_selection(selection, protocol)
if set(reconstructed_tests) != set(selected_ids) or not all(
isinstance(source, str) for source in reconstructed_tests.values()
):
raise ValueError("reconstructed tests do not match the frozen selection")
module_sha256: dict[str, str] = {}
for scenario_id in sorted(selected_ids):
source = reconstructed_tests[scenario_id]
validate_reconstructed_test_module(source)
module_sha256[scenario_id] = hashlib.sha256(source.encode()).hexdigest()
return {
"guard": "reconstructed-test-dependency-v1",
"selection_sha256": hashlib.sha256(_canonical_json_bytes(selection)).hexdigest(),
"module_sha256": module_sha256,
}
def evaluate_repository_claim(
repository_rows: list[dict[str, Any]],
protocol: dict[str, Any],
selection: dict[str, Any],
*,
scenario_pack_bytes: bytes,
collision_attestation_bytes: bytes,
control_results_bytes: bytes,
reconstructed_tests: dict[str, str],
) -> dict[str, Any]:
"""Evaluate the preregistered repository-level efficacy gates."""
execution_inputs = protocol.get("execution_inputs")
if protocol.get("stage") != "execution-frozen" or not isinstance(execution_inputs, dict):
raise ValueError("claim evaluation requires an execution-frozen protocol")
if any(
SHA256.fullmatch(str(execution_inputs.get(field) or "")) is None
for field in (
"selection_output_sha256",
"scenario_pack_sha256",
"collision_attestation_sha256",
"reconstructed_test_attestation_sha256",
"control_results_sha256",
)
):
raise ValueError("claim evaluation requires complete frozen execution hashes")
selected_ids, repository_map = _validated_selection(selection, protocol)
selection_sha256 = hashlib.sha256(_canonical_json_bytes(selection)).hexdigest()
if selection_sha256 != execution_inputs["selection_output_sha256"]:
raise ValueError("claim selection does not match the execution freeze")
scenario_pack_sha256 = hashlib.sha256(scenario_pack_bytes).hexdigest()
if scenario_pack_sha256 != execution_inputs["scenario_pack_sha256"]:
raise ValueError("claim scenario pack does not match the execution freeze")
try:
scenario_pack = json.loads(scenario_pack_bytes)
except (UnicodeDecodeError, json.JSONDecodeError) as exc:
raise ValueError("claim scenario pack is invalid") from exc
scenario_rows = scenario_pack.get("scenarios") if isinstance(scenario_pack, dict) else None
if (
not isinstance(scenario_rows, list)
or not all(
isinstance(row, dict)
and isinstance(row.get("id"), str)
and SHA256.fullmatch(str(row.get("reconstructed_test_sha256") or "")) is not None
for row in scenario_rows
)
or {str(row["id"]) for row in scenario_rows} != set(selected_ids)
or len(scenario_rows) != len(selected_ids)
):
raise ValueError("claim scenario pack does not match the frozen selection")
scenario_test_sha256 = {
str(row["id"]): str(row["reconstructed_test_sha256"]) for row in scenario_rows
}
if (
hashlib.sha256(collision_attestation_bytes).hexdigest()
!= execution_inputs["collision_attestation_sha256"]
):
raise ValueError("claim collision attestation does not match the execution freeze")
try:
collision_attestation = json.loads(collision_attestation_bytes)
except (UnicodeDecodeError, json.JSONDecodeError) as exc:
raise ValueError("claim collision attestation is invalid") from exc
if (
not isinstance(collision_attestation, dict)
or collision_attestation.get("guard") != "runtime-pack-distinctive-evidence-v1"
or collision_attestation.get("runtime_availability_sha256")
!= protocol["product_inputs"]["runtime_availability_sha256"]
or collision_attestation.get("catalog_archive_sha256")
!= protocol["product_inputs"]["catalog_archive_sha256"]
or collision_attestation.get("scenarios_sha256") != scenario_pack_sha256
or collision_attestation.get("collision_free") is not True
or isinstance(collision_attestation.get("collision_count"), bool)
or not isinstance(collision_attestation.get("collision_count"), int)
or collision_attestation.get("collision_count") != 0
or collision_attestation.get("scenario_ids") != sorted(selected_ids)
):
raise ValueError("claim collision attestation is invalid")
reconstructed_attestation = build_reconstructed_test_attestation(
selection,
protocol,
reconstructed_tests,
)
if (
hashlib.sha256(_canonical_json_bytes(reconstructed_attestation)).hexdigest()
!= execution_inputs["reconstructed_test_attestation_sha256"]
or reconstructed_attestation["module_sha256"] != scenario_test_sha256
):
raise ValueError("claim reconstructed tests do not match the execution freeze")
if (
hashlib.sha256(control_results_bytes).hexdigest()
!= execution_inputs["control_results_sha256"]
):
raise ValueError("claim control results do not match the execution freeze")
try:
control_results = json.loads(control_results_bytes)
except (UnicodeDecodeError, json.JSONDecodeError) as exc:
raise ValueError("claim control results are invalid") from exc
scenario_results = (
control_results.get("scenario_results") if isinstance(control_results, dict) else None
)
control_timeout = protocol.get("timeouts", {}).get("control_verification_seconds")
if (
isinstance(control_timeout, bool)
or not isinstance(control_timeout, int | float)
or not math.isfinite(control_timeout)
or control_timeout <= 0
or not isinstance(control_results, dict)
or control_results.get("guard") != "holdout-control-results-v1"
or control_results.get("selection_sha256") != selection_sha256
or control_results.get("scenario_pack_sha256") != scenario_pack_sha256
or (
control_results.get("all_scenarios_passed", control_results.get("all_seven_passed"))
is not True
)
or not isinstance(scenario_results, dict)
or set(scenario_results) != set(selected_ids)
):
raise ValueError("claim control results are invalid")
for scenario_id in selected_ids:
result = scenario_results[scenario_id]
if (
not isinstance(result, dict)
or result.get("parent_with_test_patch_red") is not True
or result.get("reference_patch_green") is not True
or result.get("changed_test_module_green") is not True
or result.get("timeout_compliant") is not True
or result.get("reconstructed_test_sha256") != scenario_test_sha256[scenario_id]
or any(
SHA256.fullmatch(str(result.get(field) or "")) is None
for field in (
"red_evidence_sha256",
"green_evidence_sha256",
"module_evidence_sha256",
)
)
or isinstance(result.get("elapsed_seconds"), bool)
or not isinstance(result.get("elapsed_seconds"), int | float)
or not math.isfinite(result["elapsed_seconds"])
or result["elapsed_seconds"] < 0
or isinstance(result.get("timeout_seconds"), bool)
or not isinstance(result.get("timeout_seconds"), int | float)
or not math.isfinite(result["timeout_seconds"])
or result["timeout_seconds"] <= 0
or result["timeout_seconds"] != control_timeout
or result["elapsed_seconds"] > result["timeout_seconds"]
):
raise ValueError("claim control results are invalid")
expected_scenarios: dict[str, list[str]] = defaultdict(list)
for scenario_id, repository in repository_map.items():
expected_scenarios[repository].append(scenario_id)
expected = int(protocol["selection"]["analysis_repositories"])
repositories = [str(row.get("repository") or "") for row in repository_rows]
if (
len(repository_rows) != expected
or len(set(repositories)) != expected
or set(repositories) != set(expected_scenarios)
):
raise ValueError("claim evaluation repositories do not match the frozen selection")
def ratios(field: str) -> list[float]:
values: list[float] = []
for row in repository_rows:
value = row.get(field)
if (
isinstance(value, bool)
or not isinstance(value, int | float)
or not math.isfinite(value)
or value < 0
):
raise ValueError(f"claim evaluation requires finite non-negative {field}")
values.append(float(value))
return values
token_ratios = ratios("uncached_provider_tokens_ratio")
time_ratios = ratios("total_seconds_ratio")
benefiting = sum(value < 1.0 for value in token_ratios)
support_p = sum(
math.comb(expected, successes) for successes in range(benefiting, expected + 1)
) / (2**expected)
gates = protocol["claim_gates"]
overall_token = float(statistics.median(token_ratios))
overall_time = float(statistics.median(time_ratios))
paired_trials = int(gates["paired_trials_per_scenario"])
evidence_complete = True
for row in repository_rows:
repository = str(row["repository"])
scenarios = sorted(expected_scenarios[repository])
trial_counts = row.get("paired_trials_by_scenario")
missing_pairs = row.get("missing_pairs")
unresolved = row.get("unresolved_incidents")
evidence_complete = evidence_complete and all(
(
row.get("scenario_ids") == scenarios,
isinstance(trial_counts, dict),
set(trial_counts) == set(scenarios) if isinstance(trial_counts, dict) else False,
(
all(
isinstance(count, int)
and not isinstance(count, bool)
and count == paired_trials
for count in trial_counts.values()
)
if isinstance(trial_counts, dict)
else False
),
isinstance(missing_pairs, int) and not isinstance(missing_pairs, bool),
missing_pairs == 0,
row.get("token_usage_exact") is True,
row.get("trusted_policy_outcomes") is True,
isinstance(unresolved, int) and not isinstance(unresolved, bool),
unresolved == 0,
)
)
quality_preserved = all(row.get("quality_preserved") is True for row in repository_rows)
verified_deliveries = sum(row.get("verified_delivery") is True for row in repository_rows)
incident_free = all(
isinstance(row.get("unresolved_incidents"), int)
and not isinstance(row.get("unresolved_incidents"), bool)
and row.get("unresolved_incidents") == 0
for row in repository_rows
)
passed = all(
(
overall_token <= float(gates["primary_endpoint_maximum_ratio"]),
overall_time <= float(gates["total_seconds_maximum_ratio"]),
quality_preserved,
verified_deliveries >= int(gates["minimum_repositories_with_verified_delivery"]),
benefiting >= int(gates["required_benefiting_repositories"]),
support_p <= float(gates["exact_one_sided_repository_support_alpha"]),
incident_free,
evidence_complete,
)
)
return {
"overall_token_ratio": overall_token,
"overall_time_ratio": overall_time,
"benefiting_repositories": benefiting,
"verified_delivery_repositories": verified_deliveries,
"exact_one_sided_sign_p": support_p,
"quality_preserved": quality_preserved,
"incident_free": incident_free,
"evidence_complete": evidence_complete,
"passed": passed,
}
def _load_jsonl(path: Path) -> list[dict[str, Any]]:
rows = [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines() if line]
if not all(isinstance(row, dict) for row in rows):
raise ValueError("selection JSONL must contain objects")
return rows
def _private_text_handle(path: Path) -> TextIO:
resolved = path.resolve(strict=False)
private_root = PRIVATE_ROOT.resolve()
if ROOT.resolve() in resolved.parents and private_root not in resolved.parents:
raise ValueError("holdout evidence inside the repository must use .gate/ctx-ab-private")
path.parent.mkdir(mode=0o700, parents=True, exist_ok=True)
if stat.S_IMODE(path.parent.stat().st_mode) != 0o700:
raise ValueError("holdout evidence parent must be owner-only")
if path.is_symlink() or (path.exists() and not path.is_file()):
raise ValueError("holdout evidence path must be a regular file")
if path.exists() and path.stat().st_nlink != 1:
raise ValueError("holdout evidence path must not be a hard link")
if path.exists():
path.chmod(0o600)
flags = os.O_WRONLY | os.O_CREAT | os.O_TRUNC
if hasattr(os, "O_NOFOLLOW"):
flags |= os.O_NOFOLLOW
descriptor = os.open(path, flags, 0o600)
os.fchmod(descriptor, 0o600)
return os.fdopen(descriptor, "w", encoding="utf-8", newline="")
def _paths_are_distinct(paths: list[Path]) -> bool:
for index, left in enumerate(paths):
for right in paths[index + 1 :]:
if left.resolve(strict=False) == right.resolve(strict=False):
return False
if left.exists() and right.exists() and os.path.samefile(left, right):
return False
return True
def _remove_stale_selection(path: Path) -> None:
if not path.exists() and not path.is_symlink():
return
resolved = path.resolve(strict=False)
private_root = PRIVATE_ROOT.resolve()
if ROOT.resolve() in resolved.parents and private_root not in resolved.parents:
raise ValueError("stale selection inside the repository must use .gate/ctx-ab-private")
if stat.S_IMODE(path.parent.stat().st_mode) != 0o700:
raise ValueError("stale selection parent must be owner-only")
if path.is_symlink() or not path.is_file() or path.stat().st_nlink != 1:
raise ValueError("stale selection must be a single-link regular file")
path.unlink()
def _requires_acquisition_protocol_digest(protocol: dict[str, Any]) -> bool:
selection = protocol.get("selection")
return (
protocol.get("schema_version") == 2
or protocol.get("protocol_id") == "production-graph-holdout-v2"
or (isinstance(selection, dict) and selection.get("strategy") == "one-per-repository")
)
def main(argv: list[str] | None = None) -> int:
parser = argparse.ArgumentParser()
parser.add_argument("--protocol", type=Path, default=DEFAULT_PROTOCOL)
parser.add_argument("--expected-acquisition-protocol-sha256")
parser.add_argument("--selection-jsonl", type=Path, required=True)
parser.add_argument("--exposure-ledger", type=Path)
parser.add_argument("--ledger", type=Path, required=True)
parser.add_argument("--selection", type=Path, required=True)
args = parser.parse_args(argv)
paths = [args.protocol, args.selection_jsonl, args.ledger, args.selection]
if args.exposure_ledger is not None:
paths.append(args.exposure_ledger)
if not _paths_are_distinct(paths):
raise SystemExit("protocol, source, ledger, and selection paths must be distinct")
expected_protocol_sha256 = args.expected_acquisition_protocol_sha256
if expected_protocol_sha256 is not None and SHA256.fullmatch(expected_protocol_sha256) is None:
raise SystemExit("expected acquisition protocol SHA-256 must be 64 lowercase hex digits")
protocol_bytes = args.protocol.read_bytes()
if expected_protocol_sha256 is not None and not hmac.compare_digest(
hashlib.sha256(protocol_bytes).hexdigest(),
expected_protocol_sha256,
):
raise SystemExit("acquisition protocol does not match the expected SHA-256")
protocol = json.loads(protocol_bytes)
requires_v2_authentication = _requires_acquisition_protocol_digest(protocol)
if requires_v2_authentication and expected_protocol_sha256 is None:
raise SystemExit("V2 selection requires --expected-acquisition-protocol-sha256")
exposure_document: dict[str, Any] | None = None
if requires_v2_authentication:
expected_exposure_sha256 = protocol.get("exposure_ledger_sha256")
if (
not isinstance(expected_exposure_sha256, str)
or SHA256.fullmatch(expected_exposure_sha256) is None
):
raise SystemExit("V2 acquisition protocol lacks an authenticated exposure ledger")
if args.exposure_ledger is None:
raise SystemExit(
"V2 selection requires an authenticated exposure ledger via --exposure-ledger"
)
try:
exposure_document = exposure_ledger.load_authenticated_ledger(
args.exposure_ledger,
expected_exposure_sha256,
)
except (OSError, ValueError) as exc:
raise SystemExit(f"authenticated exposure ledger is invalid: {exc}") from None
elif args.exposure_ledger is not None:
raise SystemExit("--exposure-ledger is only valid for V2 selection")
_remove_stale_selection(args.selection)
universe = protocol["universe"]
if (
protocol.get("stage") not in {"acquisition-frozen", "execution-frozen"}
or SHA256.fullmatch(str(universe.get("raw_parquet_sha256") or "")) is None
or SHA256.fullmatch(str(universe.get("duckdb_cli_sha256") or "")) is None
or SHA256.fullmatch(str(universe.get("selection_jsonl_sha256") or "")) is None
):
raise SystemExit("selection requires a frozen authenticated acquisition")
if (
hashlib.sha256(args.selection_jsonl.read_bytes()).hexdigest()
!= universe["selection_jsonl_sha256"]
):
raise SystemExit("selection JSONL does not match the frozen SHA-256")
source_rows = sorted(
_load_jsonl(args.selection_jsonl),
key=lambda row: str(row.get("instance_id") or ""),
)
if len(source_rows) != protocol["universe"]["expected_rows"]:
raise SystemExit("selection JSONL row count does not match the frozen universe")
ledger = [evaluate_row(row, protocol) for row in source_rows]
if exposure_document is not None:
ledger = reject_historical_exposures(ledger, exposure_document)
allowed_rejection_codes = set(protocol["static_candidate_rules"]["rejection_codes"])
if exposure_document is not None:
allowed_rejection_codes.add(HISTORICAL_EXPOSURE_REJECTION_CODE)
if {row["rejection_code"] for row in ledger if row["rejection_code"]} - set(
allowed_rejection_codes
):
raise SystemExit("selector emitted an undeclared rejection code")
with _private_text_handle(args.ledger) as handle:
writer = csv.DictWriter(handle, fieldnames=LEDGER_FIELDS)
writer.writeheader()
writer.writerows(ledger)
selection = select_rows(ledger, protocol)
if exposure_document is not None:
require_exposure_disjoint_selection(selection, exposure_document)
with _private_text_handle(args.selection) as handle:
handle.write(_canonical_json_bytes(selection).decode())
return 0
if __name__ == "__main__":
raise SystemExit(main())