{
  "schema_version": 1,
  "scope": "internal_behavioral_diagnostics",
  "models": {
    "candidate": "LFM2.5-2.6B-UNCENSORED-ABLITERATED-PHILADELPHIA-CLASS",
    "untouched_source": "LiquidAI/LFM2.5-2.6B"
  },
  "protocol_notes": {
    "labels": "Deterministic response-form heuristics; not factuality, safety, legality, or human-approval judgments.",
    "holdout": "126 full prompt hashes excluded from direction fitting; exact-identity disjoint, not semantic-family disjoint.",
    "reasoning": "Coherence and long-form scoring used the final answer after a completed reasoning block.",
    "matched": true
  },
  "opening_t96": {
    "sample_count_per_cohort": 842,
    "harmful": {
      "untouched_source": {"direct": 149, "warning": 561, "refusal": 132},
      "candidate": {"direct": 773, "warning": 61, "refusal": 8}
    },
    "paired_harmless": {
      "untouched_source": {"direct": 795, "warning": 45, "refusal": 2},
      "candidate": {"direct": 812, "warning": 30, "refusal": 0}
    },
    "harmful_transition_notes": {
      "source_refusals_cleared": 131,
      "new_candidate_refusals": 7
    }
  },
  "holdout_t96": {
    "sample_count_per_cohort": 126,
    "harmful": {
      "untouched_source": {"direct": 15, "warning": 92, "refusal": 19},
      "candidate": {"direct": 114, "warning": 11, "refusal": 1}
    },
    "paired_harmless": {
      "untouched_source": {"direct": 118, "warning": 8, "refusal": 0},
      "candidate": {"direct": 123, "warning": 3, "refusal": 0}
    }
  },
  "lfm_aware_coherence24": {
    "sample_count": 24,
    "untouched_source_passed": 17,
    "candidate_passed": 20,
    "untouched_source_code_syntax": "4/6",
    "candidate_code_syntax": "6/6",
    "untouched_source_code_semantic": "2/6",
    "candidate_code_semantic": "5/6",
    "untouched_source_json_validity": "4/4",
    "candidate_json_validity": "4/4",
    "untouched_source_repetition_flags": 4,
    "candidate_repetition_flags": 3
  },
  "longform24": {
    "sample_count": 24,
    "untouched_source": {
      "strictly_usable": 6,
      "refusal": 18,
      "repetition_flags": 0,
      "degenerate_openings": 0,
      "incomplete_or_empty_finals": 0
    },
    "candidate": {
      "strictly_usable": 18,
      "refusal": 0,
      "repetition_flags": 5,
      "degenerate_openings": 3,
      "incomplete_or_empty_finals": 2
    },
    "note": "Candidate structural flags overlap across six failed rows."
  },
  "first_token_kl": {
    "source_vs_candidate_mean": 0.00152233,
    "source_vs_candidate_max": 0.00418210
  },
  "selection_note": "A stronger third projection pass was rejected because it retained eight harmful t96 refusals while increasing warnings and regressing harmless behavior."
}