{
  "document": "Brain public research brief",
  "organization": "R&D BioTech Alaska",
  "published_date": "2026-09-02",
  "system": {
    "name": "Brain",
    "classification": "protected local quantum-enhanced cognitive research system",
    "brain_owned_mutable_scalars": 29930048,
    "changed_scalars_in_accepted_training": 29929950,
    "accepted_optimizer_steps": 1024,
    "whole_brain_acceptance": {
      "passed": 64,
      "total": 64
    },
    "retention": {
      "passed": 106,
      "total": 106
    },
    "donor_runtime_required": false,
    "production_access": "protected"
  },
  "accepted_public_performance": {
    "interpretation": "approximately 500M conventional-model performance class on the previously measured ARC-Easy, PIQA, and WinoGrande profile",
    "brain_parameter_estimate": 30000000,
    "boundary": "This remains a performance-class estimate assembled from Brain evidence, not a protocol-identical head-to-head or universal model equivalence.",
    "prior_public_record": "/brain/brain-benchmark-brief-2026-08.json"
  },
  "target_free_evidence_frontier": {
    "status": "frozen internal validation measurement",
    "validation_rows": 2378,
    "evidence_sources": 17,
    "correct_candidate_present": 2354,
    "aggregate_coverage_percent": 98.99,
    "macro_coverage_percent": 98.14,
    "rows_without_a_correct_source_proposal": 24,
    "tasks": [
      {
        "benchmark_family": "ARC-Challenge",
        "correct_candidate_present": 52,
        "rows": 55,
        "coverage_percent": 94.55
      },
      {
        "benchmark_family": "BoolQ",
        "correct_candidate_present": 625,
        "rows": 625,
        "coverage_percent": 100.0
      },
      {
        "benchmark_family": "HellaSwag",
        "correct_candidate_present": 523,
        "rows": 544,
        "coverage_percent": 96.14
      },
      {
        "benchmark_family": "PIQA",
        "correct_candidate_present": 575,
        "rows": 575,
        "coverage_percent": 100.0
      },
      {
        "benchmark_family": "WinoGrande",
        "correct_candidate_present": 579,
        "rows": 579,
        "coverage_percent": 100.0
      }
    ],
    "measurement_boundary": "Evidence coverage measures whether at least one legal target-free source proposed the correct candidate. It is not final-answer accuracy."
  },
  "v3_reconstruction_run": {
    "status": "terminated and rejected",
    "candidate_parent_union_correct": 2101,
    "validation_rows": 2378,
    "aggregate_percent": 88.35,
    "macro_percent": 82.63,
    "recoverable_correct_evidence_not_preserved": 253,
    "interpretation": "The available evidence exceeds the present reconstruction and arbitration system's ability to preserve the correct candidate."
  },
  "engineering_milestone": {
    "target": ">=90% reproducible target-free executable performance",
    "status": "not yet achieved",
    "controls": [
      "frozen five-task domain",
      "no protected target access for tuning",
      "exact runtime identity",
      "QSA identity and control records",
      "reload and checkpoint authority preserved"
    ],
    "conditional_comparison": "If reproduced, 70B+ performance class on this measured five-task profile",
    "comparison_boundary": "The conditional comparison would apply only to the measured task suite. It would not establish universal equivalence to a 70B model."
  },
  "parameter_efficiency": {
    "brain_learned_cortex_parameters": 29930048,
    "brain_raw_fp16_storage": "approximately 60 MB",
    "comparisons": [
      {
        "conventional_parameter_scale": "500M",
        "relative_parameter_count": "16.7x larger",
        "raw_fp16_storage": "approximately 1 GB"
      },
      {
        "conventional_parameter_scale": "20B",
        "relative_parameter_count": "approximately 668x larger",
        "raw_fp16_storage": "approximately 40 GB"
      },
      {
        "conventional_parameter_scale": "27B",
        "relative_parameter_count": "approximately 902x larger",
        "raw_fp16_storage": "approximately 54 GB"
      },
      {
        "conventional_parameter_scale": "32B",
        "relative_parameter_count": "approximately 1,069x larger",
        "raw_fp16_storage": "approximately 64 GB"
      },
      {
        "conventional_parameter_scale": "70B",
        "relative_parameter_count": "approximately 2,339x larger",
        "raw_fp16_storage": "approximately 140 GB"
      },
      {
        "conventional_parameter_scale": "405B",
        "relative_parameter_count": "approximately 13,532x larger"
      }
    ],
    "runtime_boundary": "Raw weight storage is not whole-system latency or memory use. Brain also performs retrieval, structured cognition, evidence evaluation, and QSA computation."
  },
  "published_larger_model_context": [
    {
      "model": "GPT-NeoX-20B",
      "parameters": "20B",
      "arc_challenge_percent": 41.0,
      "boolq_percent": 69.4,
      "hellaswag_percent": 71.4,
      "piqa_percent": 77.7,
      "winogrande_percent": 66.0,
      "source": "https://proceedings.neurips.cc/paper_files/paper/2022/file/adf7fa39d65e2983d724ff7da57f00ac-Supplemental-Conference.pdf"
    },
    {
      "model": "Gemma 2 27B PT",
      "parameters": "27B",
      "arc_challenge_percent": 71.4,
      "boolq_percent": 84.8,
      "hellaswag_percent": 86.4,
      "piqa_percent": 83.2,
      "winogrande_percent": 83.7,
      "simple_five_task_mean_percent": 81.9,
      "source": "https://ai.google.dev/gemma/docs/core/model_card_2"
    }
  ],
  "architecture_public_description": [
    "compact learned cortex",
    "memory",
    "evidence",
    "semantic state",
    "competing hypotheses",
    "world state",
    "procedures",
    "provenance",
    "exact QSA computation",
    "final Brain authority"
  ],
  "qsa_boundary": "QSA is Brain's exact quantum-state computation layer. Brain is described as quantum-enhanced. Current matched controls do not establish physical quantum speedup or universal quantum advantage.",
  "biomedical_research_targets": [
    "drug-target discovery",
    "pathway analysis",
    "multi-omics integration",
    "biomarker discovery",
    "treatment-response modeling",
    "rare-disease hypothesis generation",
    "protein-function relationships",
    "mechanistic biomedical research"
  ],
  "claim_boundary": [
    "The 98.99% result is target-free evidence coverage, not final-answer accuracy.",
    "The V3 run was rejected and Brain has not yet achieved the >=90% executable milestone.",
    "70B+ is a conditional measured-suite performance-class comparison, not a current result or universal model equivalence.",
    "No current result establishes physical quantum speedup, universal quantum advantage, sentience, general intelligence, or clinical validity.",
    "Private weights, corpora, memory contents, model topology, thresholds, credentials, and security-sensitive protocols are not included."
  ],
  "public_foundations": {
    "brain": "https://github.com/R-D-BioTech-Alaska/Brain",
    "qelm": "https://github.com/R-D-BioTech-Alaska/Qelm",
    "qsa": "https://github.com/R-D-BioTech-Alaska/QSA"
  }
}