{ "document": "Brain public benchmark brief", "organization": "R&D BioTech Alaska", "published_date": "2026-08-22", "system": { "name": "Brain", "classification": "protected local quantum-classical intelligence research system", "brain_owned_mutable_scalars": 29930048, "accepted_optimizer_steps": 1024, "changed_scalars": 29929950, "whole_brain_acceptance": { "passed": 64, "total": 64 }, "retention": { "passed": 106, "total": 106 }, "donor_runtime_required": false, "production_access": "protected" }, "benchmark_records": [ { "record": "ARC grounded evidence V33", "cohort": "200 fresh ARC questions", "inherited_correct": 43, "measured_correct": 92, "inherited_accuracy_percent": 21.5, "measured_accuracy_percent": 46.0, "correct_answer_delta": 49, "percentage_point_delta": 24.5, "controls": [ "no retraining", "no target labels available to the reasoning path", "protected Brain state unchanged", "no donor runtime", "information-identical classical product control" ], "attribution": "The QSA route and information-identical classical product selected the same answers. The result supports the cognitive architecture, not a uniquely quantum accuracy advantage.", "evidence_sha256": "978d35794ae4a19b24f7e842d5ddedff7397dd2f0f1222d8d4e9aaf6ced29f72" }, { "record": "WinoGrande object-state V35", "cohort": "1,024 never-used train-split holdout rows", "inherited_correct": 523, "measured_correct": 550, "inherited_accuracy_percent": 51.0742, "measured_accuracy_percent": 53.7109, "correct_answer_delta": 27, "percentage_point_delta": 2.6367, "object_route": { "resolved": 28, "correct": 26, "actual_override_gains": 14, "actual_override_losses": 2 }, "controls": [ "prediction frozen before scoring", "1,024 of 1,024 inherited choices independently reproduced", "public QSA 0.2.0", "private fallback disabled", "no neural parameter increase" ], "boundary": "This is a fresh train-split holdout result, not an official WinoGrande validation score.", "prediction_sha256": "80218f169d93dd6e22950605f386d7a8904aafdc86146e9cead926eed52d276e", "score_sha256": "0d1989ef3b5fafd1e7e15cd886a496d3b53726e7981c330ce9bed6cdb150952a" }, { "record": "HellaSwag event-state V31", "cohort": "384 untouched event rows", "inherited_correct": 107, "measured_correct": 131, "inherited_accuracy_percent": 27.8646, "measured_accuracy_percent": 34.1146, "correct_answer_delta": 24, "percentage_point_delta": 6.25, "controls": [ "299 unique Pareto branches independently confirmed", "1,537 QSA registers", "public QSA 0.2.0", "no neural parameter increase" ], "boundary": "This result applies to the recorded untouched event cohort and is not a universal HellaSwag or model-equivalence claim." }, { "record": "PIQA symbolic state V26", "cohort": "128 fresh PIQA rows", "inherited_correct": 52, "measured_correct": 57, "inherited_accuracy_percent": 40.625, "measured_accuracy_percent": 44.5313, "correct_answer_delta": 5, "percentage_point_delta": 3.9063, "actual_gains": 5, "actual_losses": 0, "boundary": "The measured route was high precision and low coverage. A later lexical expansion did not generalize and was rejected." } ], "performance_class_comparison": { "comparison_type": "current Brain estimate versus published conventional-model evaluations", "brain_parameter_estimate": 30000000, "current_interpretation": "Brain's three-task profile is presently consistent with the several-hundred-million-parameter conventional-model class, with Qwen1.5-500M the closest listed reference.", "rows": [ { "model": "Brain", "parameters": "approximately 30M", "arc_easy_percent": "approximately 54-55", "piqa_percent": "approximately 66-67", "winogrande_percent": "approximately 55-56", "result_type": "current performance-class estimate from Brain evidence" }, { "model": "OPT", "parameters": "125M", "arc_easy_percent": 41.3, "piqa_percent": 62.0, "winogrande_percent": 50.8, "result_type": "published zero-shot evaluation" }, { "model": "GPT-Neo", "parameters": "125M", "arc_easy_percent": 40.7, "piqa_percent": 62.5, "winogrande_percent": 50.7, "result_type": "published zero-shot evaluation" }, { "model": "MobileLLM", "parameters": "125M", "arc_easy_percent": 43.9, "piqa_percent": 65.3, "winogrande_percent": 53.1, "result_type": "published model evaluation" }, { "model": "Pythia", "parameters": "410M", "arc_easy_percent": 47.1, "piqa_percent": 67.2, "winogrande_percent": 53.4, "result_type": "published model evaluation" }, { "model": "MobileLLM", "parameters": "350M", "arc_easy_percent": 53.8, "piqa_percent": 68.6, "winogrande_percent": 57.6, "result_type": "published model evaluation" }, { "model": "Qwen1.5", "parameters": "500M", "arc_easy_percent": 54.7, "piqa_percent": 68.9, "winogrande_percent": 55.0, "result_type": "published model evaluation", "parameter_ratio_to_brain": 16.7 }, { "model": "MobileLLM", "parameters": "600M", "arc_easy_percent": 58.1, "piqa_percent": 72.3, "winogrande_percent": 58.6, "result_type": "published model evaluation" }, { "model": "TinyLlama", "parameters": "1.1B", "arc_easy_percent": "55.2-59.2", "piqa_percent": "72.9-73.6", "winogrande_percent": "58.8-59.4", "result_type": "published checkpoint range" } ], "additional_reference": { "model": "SmolLM2-135M", "parameters": "135M", "arc_average_percent": 43.9, "piqa_percent": 68.4, "winogrande_percent": 51.3, "hellaswag_percent": 42.1, "parameter_ratio_to_brain": 4.5 }, "protocol_note": "Protocols, checkpoints, prompts, and evaluation harnesses are not perfectly identical. This is a performance-class comparison, not a publication-grade head-to-head or a claim of universal model equivalence." }, "external_sources": [ { "title": "MobileLLM published evaluation table", "url": "https://huggingface.co/facebook/MobileLLM-125M/blob/main/README.md" }, { "title": "SmolLM2-135M model card", "url": "https://huggingface.co/HuggingFaceTB/SmolLM2-135M" }, { "title": "TinyLlama checkpoint evaluations", "url": "https://huggingface.co/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T" }, { "title": "Qwen1.5-0.5B model card", "url": "https://huggingface.co/Qwen/Qwen1.5-0.5B" }, { "title": "Pythia model suite", "url": "https://github.com/EleutherAI/pythia" } ], "claim_boundary": [ "The results demonstrate benchmark-specific fixed-parameter gains.", "The current three-task estimate places Brain in a several-hundred-million-parameter performance class, but it does not establish protocol-matched model equivalence.", "No current result establishes universal quantum advantage, physical-QPU advantage, sentience, general intelligence, or broad superiority to conventional language models.", "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" } }