{
  "schema_version": 3,
  "recorded_at": "2026-08-17",
  "claim": "Aray was evaluated end to end in two stages connected by a frozen handoff: its normalization workflow accepted all 416 selected source entries, all final original-derivative pairs passed the supported deterministic checks, and all 406 entries admitted by preflight produced artifacts matching their upstream original rules.",
  "evaluation_design": {
    "type": "staged_end_to_end_with_frozen_handoff",
    "lineage": "public source file -> Aray source reduction and normalization -> deterministic normalized-rule validation -> frozen accepted predicate -> preflight -> extraction -> realization -> upstream-original YARA scan",
    "acceptance_oracle": "Identifier-restricted YARA match against the derivative-independent upstream original rule",
    "source_association": "matching filename, trailing directory path, and first public rule identifier; missing or ambiguous associations fail explicitly",
    "scope_note": "The original-rule oracle independently validates each concrete artifact against a different predicate, but shares Aray's source-selection and dependency-closure machinery and the same YARA implementation. Its independence is predicate-level, not implementation-level, and it does not prove universal implication."
  },
  "corpus": {
    "source": "https://github.com/Yara-Rules/rules",
    "source_commit": "0f93570194a80d2f2032869055808b0ddcdfb360",
    "local_path": "evaluation/normalized-glm-5.2-stable",
    "local_path_role": "frozen handoff from Aray normalization to deterministic realization",
    "frozen_corpus_is_versioned": false,
    "availability_note": "The repository versions this aggregate summary and digest, not the frozen corpus or canonical per-rule reports.",
    "rule_count": 416,
    "sha256_of_sorted_file_hash_manifest": "2b7d09de6a0f152218a471ad018dd50fe9af4a4aae08d5ac7fccfa9cd2cb985d",
    "digest_command": "rg --files evaluation/normalized-glm-5.2-stable -g '*.yar' -0 | sort -z | xargs -0 sha256sum | sha256sum"
  },
  "normalization_stage": {
    "owner": "Aray",
    "interface": "aray-normalize",
    "implementation": "Aray source selection, aray.nodes.normalize_rule, deterministic canonicalization and candidate validation, contextual regex witness checks, and batch-only passed-verdict admission",
    "configuration": {
      "normalize_model": "glm-5.2:cloud",
      "judge_model": "glm-5.2:cloud",
      "workers": 1
    },
    "result": {
      "source_entries": 416,
      "accepted": 416,
      "failed": 0,
      "accepted_without_model_normalization": 182,
      "accepted_after_model_normalization": 234,
      "admission_rate_percent": 100.0
    },
    "evidence": {
      "report_path": "evaluation/reports/2026-08-12T19-30-54-glm-5.2/norm_report.json",
      "report_is_versioned": false,
      "scope_note": "This report records the initial normalization stage and predates the two anchored-regex repairs applied before the final freeze."
    }
  },
  "normalized_rule_validation": {
    "implementation": "yara-python compilation, selected-rule identity, retained declarations and modifiers, supported count expansions, fixed witnesses derived from regex and hex declarations, and a narrow deterministic proof for supported conjunction/count transformations",
    "validated_at": "2026-08-17",
    "validation_method": "Pair each final frozen derivative with its selected upstream original and invoke aray.yara_validation.validate_normalization.",
    "final_original_derivative_pairs": 416,
    "passed_supported_deterministic_checks": 416,
    "pairwise_report_is_versioned": false,
    "fully_proven_by_narrow_implication_fragment_not_reported": true,
    "universal_implication_claimed": false,
    "regressions_repaired_before_final_run": [
      "MALW_AlMashreq",
      "RAT_PoetRATDoc"
    ]
  },
  "canonical_validation": {
    "command": "uv run aray-eval --model deterministic-control --base-url http://127.0.0.1:9/v1 --validate-original --original-directory evaluation/yara-repos/rules --workers 1 --output /tmp/aray-original-deterministic-control.json",
    "mode": "full_compile",
    "scan_only": false,
    "workers": 1,
    "duration_seconds": 11.418,
    "canonical_per_rule_report_is_versioned": false,
    "model_control": {
      "configured_model": "deterministic-control",
      "base_url": "http://127.0.0.1:9/v1",
      "endpoint_expected_to_be_unreachable": true,
      "purpose": "Any attempted LLM invocation would make the run fail."
    },
    "toolchains": [
      "gcc",
      "x86_64-w64-mingw32-gcc",
      "i686-w64-mingw32-gcc",
      "yara"
    ],
    "result": {
      "upstream_original_matched": 406,
      "unsupported": 7,
      "infeasible": 2,
      "unsatisfiable": 1,
      "yara_mismatch": 0,
      "construction_failed": 0,
      "positive_fixture_yield_percent": 97.6,
      "artifact_routes": {
        "elf": 258,
        "pe": 101,
        "generic": 47
      }
    },
    "processing_paths": {
      "normalization": {
        "not_needed": 416
      },
      "strings_extraction": {
        "deterministic": 406,
        "not_reached": 10,
        "llm_fallback": 0
      },
      "constants_extraction": {
        "deterministic": 406,
        "not_reached": 10,
        "llm_fallback": 0
      },
      "llm_usage": {
        "rules_using_llm": 0,
        "rules_without_llm": 416,
        "scope": "deterministic realization stage only"
      }
    }
  },
  "end_to_end_result": {
    "source_entries_processed": 416,
    "terminal_dispositions": 416,
    "upstream_original_matching_artifacts": 406,
    "positive_fixture_yield_percent": 97.6,
    "conditional_realization": "406/406",
    "preflight_dispositions": 10
  },
  "configuration_cross_checks_scope": "Historical realization-stage controls used the normalized predicate oracle and are not evidence for the final upstream-original claim.",
  "configuration_cross_checks": [
    {
      "configured_extract_model": "glm-5.2:cloud",
      "matched": 406,
      "preflight_dispositions": 10,
      "llm_invoked": false
    },
    {
      "configured_extract_model": "qwen3.5:4b",
      "matched": 406,
      "preflight_dispositions": 10,
      "llm_invoked": false
    },
    {
      "configured_extract_model_family": "Phi-4",
      "configured_extract_model_identifier": null,
      "identifier_retained": false,
      "matched": 406,
      "preflight_dispositions": 10,
      "llm_invoked": false
    },
    {
      "configured_extract_model": "gpt-4.1",
      "matched": 406,
      "preflight_dispositions": 10,
      "llm_invoked": false
    }
  ],
  "failure_categories": {
    "unsupported_pe_module": [
      "APT_CrashOverride",
      "APT_Shamoon_StoneDrill",
      "MALW_Batel",
      "MALW_IcedID",
      "MALW_Pyinstaller",
      "RANSOM_Stampado",
      "peid"
    ],
    "whole_file_hash_preimage": [
      "APT_Grasshopper",
      "RAT_CrossRAT"
    ],
    "integer_value_out_of_range": [
      "APT_Derusbi"
    ]
  },
  "interpretation": {
    "measures": [
      "Aray normalization admission and path",
      "deterministic YARA constraint extraction",
      "capability preflight",
      "artifact construction",
      "compiler and linker behavior",
      "deterministic normalized-rule validation",
      "final YARA matching against associated upstream original rules"
    ],
    "does_not_measure": [
      "LLM extraction quality",
      "LLM latency",
      "LLM cost",
      "provider performance",
      "universal semantic implication of normalized predicates to upstream source rules",
      "coverage of arbitrary YARA"
    ],
    "note": "The four model names above were inactive controls in the realization stage, not model benchmarks. Model use occurred in the preceding Aray normalization stage for 234 entries. Two anchored-regex mismatches observed before the final run were repaired and retained as validator regressions, so the reported result is post-fix rather than held out."
  }
}