File size: 10,971 Bytes
c9761c3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 | # License: Apache-2.0 License
# Created by: Patrick Lumbantobing, VertoX-AI
# Copyright (c) 2026 VertoX-AI. All rights reserved.
#
# This work is licensed under the Apache-2.0 License.
# To view a copy of this license, visit
# https://www.apache.org/licenses/LICENSE-2.0
"""Persist and verify backend-neutral full-word evidence with explicit provenance.
Native parsing remains in the existing backend adapters. This module accepts only
canonical alignment capsules and immutable file/member references, never native words.
"""
from __future__ import annotations
import hashlib
import json
import zipfile
from collections.abc import Mapping
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Any
from speech_text_data_aligners.core.errors import ConfigurationError
from speech_text_data_aligners.core.models import (
AlignmentProvenance,
AlignmentRequest,
AlignmentResult,
Diagnostic,
DiagnosticSeverity,
JsonValue,
LexicalWord,
WordAlignment,
)
from speech_text_data_aligners.core.validation import validate_alignment_result
from speech_text_data_aligners.runtime.receipts import sha256_file
@dataclass(frozen=True, slots=True)
class EvidenceCandidate:
"""Locate one previously audited canonical full-word capsule.
Parameters
----------
path: Local canonical JSON file to verify.
sha256: Expected capsule SHA-256; missing/mismatched hashes reject reuse.
"""
path: Path
sha256: str
def artifact_sha256(reference: str) -> str:
"""Hash a file or one exact ``archive.zip::member`` without extracting it.
Parameters
----------
reference: Existing local path, optionally with an explicit ZIP member.
Returns
-------
SHA-256 of the referenced bytes.
Raises
------
ConfigurationError: If the source/member is missing, duplicated, or unreadable.
Side Effects:
Reads source bytes only, including retained WAVs when verifying prior audio.
"""
if "::" not in reference:
return sha256_file(Path(reference))
archive, member = reference.split("::", 1)
try:
with zipfile.ZipFile(archive) as source:
if source.namelist().count(member) != 1:
raise ConfigurationError(
"archive evidence member is missing or ambiguous"
)
with source.open(member) as stream:
digest = hashlib.sha256()
for chunk in iter(lambda: stream.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
except (OSError, ValueError, zipfile.BadZipFile) as error:
raise ConfigurationError(
f"cannot hash archived evidence: {reference}"
) from error
def publish_json(path: Path, payload: object, *, jsonl: bool = False) -> None:
"""Publish a small deterministic artifact without replacing an existing path.
Parameters
----------
path: New destination, whose parent is created as needed.
payload: JSON-safe value, or a list/tuple of JSON-safe rows for JSONL.
jsonl: Write one compact JSON value per row when true.
Raises
------
ConfigurationError: If serialization fails or publication would overwrite data.
Side Effects:
Creates the destination after completing serialization with NaN forbidden.
"""
try:
if jsonl:
if not isinstance(payload, (list, tuple)):
raise ValueError("JSONL requires a row sequence")
text = "".join(
json.dumps(row, sort_keys=True, ensure_ascii=False, allow_nan=False)
+ "\n"
for row in payload
)
else:
text = (
json.dumps(
payload,
sort_keys=True,
indent=2,
ensure_ascii=False,
allow_nan=False,
)
+ "\n"
)
path.parent.mkdir(parents=True, exist_ok=True)
with path.open("x", encoding="utf-8", newline="\n") as stream:
stream.write(text)
except (OSError, ValueError, TypeError) as error:
raise ConfigurationError(f"cannot publish evidence: {path}: {error}") from error
def _freeze(value: Any) -> JsonValue:
"""Restore immutable attribute tuples from JSON ``value``; reject other types."""
if isinstance(value, list):
return tuple(_freeze(item) for item in value)
if value is None or isinstance(value, (str, int, float, bool)):
return value
raise ConfigurationError("alignment attributes must contain JSON scalars/tuples")
def _result(payload: dict[str, Any]) -> AlignmentResult:
"""Decode a canonical ``payload`` result, preserving provenance and diagnostics.
Returns:
Typed immutable alignment; validation is performed by the caller.
Raises:
KeyError, TypeError, ValueError: If the canonical schema is malformed.
ConfigurationError: If attribute serialization is invalid.
"""
provenance = dict(payload["provenance"])
provenance["attributes"] = tuple(
(key, _freeze(value)) for key, value in provenance.get("attributes", [])
)
provenance["artifacts"] = tuple(provenance.get("artifacts", []))
words = tuple(
WordAlignment(
LexicalWord(**word["lexical_word"]),
word["start_s"],
word["end_s"],
word.get("confidence"),
tuple((key, _freeze(value)) for key, value in word.get("attributes", [])),
)
for word in payload["words"]
)
diagnostics = tuple(
Diagnostic(
d["code"],
DiagnosticSeverity(d["severity"]),
d["message"],
d["stage"],
tuple((k, _freeze(v)) for k, v in d.get("details", [])),
)
for d in payload.get("diagnostics", [])
)
return AlignmentResult(
payload["request_id"],
words,
AlignmentProvenance(**provenance),
diagnostics,
payload.get("runtime_s", 0.0),
)
def save_alignment_evidence(
path: Path,
request: AlignmentRequest,
result: AlignmentResult,
*,
audio_sha256: str,
transcript_sha256: str,
source_artifacts: Mapping[str, str],
) -> EvidenceCandidate:
"""Persist a validated reference-text result as a reusable full-word capsule.
Parameters
----------
path: New capsule destination.
request: Original source request to validate against the result.
result: Full neutral result returned/reconstructed by a repository adapter.
audio_sha256: Previously verified exact WAV digest.
transcript_sha256: Previously verified exact transcript digest.
source_artifacts: Immutable reference-to-SHA mapping proving source/provenance.
Returns
-------
Hashed candidate; reuse independently verifies its contents and references.
Raises
------
SpeechTextAlignerError: If alignment validation or publication fails.
Side Effects:
Publishes only canonical words and provenance; no WAV/model bytes are copied.
"""
validate_alignment_result(request, result)
publish_json(
path,
dict(
schema_version="w1t.alignment_evidence.v1",
audio_id=request.request_id,
audio_sha256=audio_sha256,
original_en_transcript_sha256=transcript_sha256,
reference_text_authoritative=True,
result=asdict(result),
source_artifacts=dict(source_artifacts),
),
)
return EvidenceCandidate(path, sha256_file(path))
def verify_alignment_evidence(
candidate: EvidenceCandidate,
request: AlignmentRequest,
*,
audio_sha256: str,
transcript_sha256: str,
expected_identity: Mapping[str, str],
) -> AlignmentResult:
"""Require exact source, backend, model/configuration, and full-word evidence.
Parameters
----------
candidate: Canonical capsule and its expected immutable hash.
request: Current original transcript/audio request.
audio_sha256: Verified current WAV digest.
transcript_sha256: Verified current exact transcript digest.
expected_identity: Required provenance fields, including backend_name.
Returns
-------
Validated full result, unchanged.
Raises
------
SpeechTextAlignerError: If a required proof is absent/mismatched or times fail.
Side Effects:
Hashes the capsule and every cited source artifact; no model is invoked.
"""
if not candidate.sha256 or sha256_file(candidate.path) != candidate.sha256:
raise ConfigurationError("capsule artifact SHA-256 mismatch")
try:
value = json.loads(candidate.path.read_text(encoding="utf-8"))
if (
value["schema_version"] != "w1t.alignment_evidence.v1"
or value["audio_id"] != request.request_id
):
raise ConfigurationError("capsule schema/audio identity mismatch")
if value["audio_sha256"] != audio_sha256:
raise ConfigurationError("audio SHA-256 mismatch")
if value["original_en_transcript_sha256"] != transcript_sha256:
raise ConfigurationError("transcript SHA-256 mismatch")
if value["reference_text_authoritative"] is not True:
raise ConfigurationError("reference-text authority is not proved")
result = _result(value["result"])
provenance = asdict(result.provenance)
required = (
"backend_name",
"backend_version",
"model_id",
"config_sha256",
"python_version",
"device",
)
if any(not provenance[k] for k in required) or not (
provenance["model_revision"] or provenance["model_sha256"]
):
raise ConfigurationError(
"insufficient backend/model/configuration provenance"
)
if not expected_identity.get("backend_name") or any(
provenance.get(k) != v for k, v in expected_identity.items()
):
raise ConfigurationError("backend identity mismatch")
sources = value["source_artifacts"]
if not isinstance(sources, dict) or not sources:
raise ConfigurationError("source artifact proof is missing")
for reference, digest in sources.items():
if artifact_sha256(reference) != digest:
raise ConfigurationError("source artifact SHA-256 mismatch")
return validate_alignment_result(request, result)
except (
OSError,
UnicodeError,
ValueError,
KeyError,
TypeError,
AttributeError,
) as error:
raise ConfigurationError(f"malformed full-word evidence: {error}") from error
|