{ "document": "Brain Final-R6 public research record", "organization": "R&D BioTech Alaska", "published_date": "2026-09-14", "source": "R&D BioTech Alaska research-state update, September 2026", "current_state": { "accepted_architecture": "Final-R6", "status": "accepted", "phase": "Post-cortex training and intelligence development", "mutable_language_scalars": 29930048, "mutable_language_tensors": 54, "persisted_neural_scalars": 49324558, "persisted_neural_scalar_limit": 50000000, "protected_systems": ["Final-R6", "V22", "Query192"], "anti_forgetting": "active", "runtime_authority": "Brain-native", "cortex_repair": "closed", "public_deployment": "Early-access conversational release in preparation", "active_work": ["Post-cortex training", "Governed recall", "Conversational integration", "Learning Center and knowledge integration", "Anti-forgetting and route preservation"], "fresh_evaluation": { "build": "Final-R6", "status": "pending fresh evaluation", "tasks": ["ARC-Easy", "PIQA", "WinoGrande"], "timing": "After the current post-cortex training sequence", "scores": null } }, "count_boundary": "29,930,048 describes the mutable language state across 54 tensors. 49,324,558 describes the persisted Final-R6 neural state. Neither count measures total runtime memory, retrieval storage, or latency.", "historical_records": { "august_benchmarks": "/brain/brain-benchmark-brief-2026-08.json", "september_2_frontier": "/brain/brain-research-brief-2026-09.json", "interpretation": "The historical approximately 30M three-task profile overlaps published results from several larger reference models. Protocols are not identical. It is not a Final-R6 benchmark.", "task_boundary": "The September 2 evidence study includes ARC-Challenge. The historical three-task comparison and planned Final-R6 evaluation use ARC-Easy. These are distinct experiments." }, "evidence_records": [ { "id": "FINAL-R6", "date": "September 2026", "status": "accepted", "build": "Final-R6", "title": "Fixed cortex baseline accepted", "question": "Which architecture is the current accepted baseline?", "method": "Whole-cortex repair and acceptance established the fixed Final-R6 baseline for subsequent training.", "result": "Cortex repair is closed. Final-R6 persists 49,324,558 neural scalars within the 50M limit; mutable language state remains 29,930,048 scalars across 54 tensors.", "interpretation": "Work has moved into post-cortex learning, recall, conversation, and knowledge integration.", "limitation": "Architecture acceptance does not supply a fresh Final-R6 ARC-Easy, PIQA, or WinoGrande score.", "artifact": "/brain/brain-final-r6-2026-09-14.json", "artifact_label": "Public Final-R6 summary", "artifact_boundary": "Operator-reported research state. Detailed acceptance artifacts remain protected; no public run hash is supplied." }, { "id": "FINAL-R6-LEARNING", "date": "September 2026", "status": "in progress", "build": "Final-R6", "title": "Post-cortex Brain-native learning", "question": "How does accepted Brain state develop while established routes are retained?", "method": "Exact Final-R6 training execution, anti-forgetting, route preservation, governed recall, and Learning Center integration.", "result": "Post-cortex learning is established. Full training, conversational integration, and governed-recall development are active.", "interpretation": "Stable milestones describe the research phase; intermediate training steps are not public capability scores.", "limitation": "No completion claim or fresh benchmark score is attached to the ongoing sequence.", "artifact": "/brain/brain-final-r6-2026-09-14.json", "artifact_label": "Public training-state summary", "artifact_boundary": "Training state and private knowledge remain protected." }, { "id": "FINAL-R6-EVALUATION", "date": "Next evaluation", "status": "pending fresh evaluation", "build": "Final-R6 after training", "title": "Fresh three-task benchmark", "question": "What does Final-R6 achieve after the current training sequence?", "method": "Fresh ARC-Easy, PIQA, and WinoGrande evaluation following post-cortex training, with build and protocol recorded.", "result": "Pending. No Final-R6 scores are published in this update.", "interpretation": "Historical approximately 30M scores remain independent of this evaluation.", "limitation": "Intermediate experimental scores cannot substitute for the fresh evaluation.", "artifact": null, "artifact_label": "Result pending", "artifact_boundary": "A completed evaluation artifact is not yet available." }, { "id": "AUG21-TRAINING", "date": "August 21, 2026", "status": "superseded", "build": "Historical approximately 30M Brain", "title": "End-to-end training and exact reload", "question": "Could Brain complete its learning and acceptance cycle?", "method": "A fixed 29,930,048-scalar budget, 1,024 optimizer steps, checkpoint reload, whole-Brain acceptance, and retention checks.", "result": "29,929,950 scalars changed; 64 / 64 acceptance cases and 106 / 106 retention checks passed.", "interpretation": "An accepted historical milestone. Final-R6 is now the architecture baseline.", "limitation": "These acceptance counts belong to the August build and are not relabeled as Final-R6 validation.", "artifact": "/brain/brain-benchmark-brief-2026-08.json", "artifact_label": "August benchmark brief", "artifact_boundary": "Public summary; private checkpoint contents are excluded." }, { "id": "V33", "date": "August 2026", "status": "measured", "build": "Historical ARC grounded-evidence route", "title": "Grounded evidence on 200 fresh ARC questions", "question": "Can grounded candidate evidence improve the inherited answer?", "method": "Frozen candidate-specific evidence, question anchoring, relation checks, no retraining, and an information-identical classical control.", "result": "43 / 200 became 92 / 200: +49 correct answers and +24.5 percentage points.", "interpretation": "QSA and the classical control tied. The measured gain supports the cognitive architecture.", "limitation": "The source labels this cohort ARC without identifying Easy or Challenge. It is not equated with either comparison table.", "artifact": "/brain/brain-benchmark-brief-2026-08.json", "artifact_label": "V33 public record and recorded hash", "evidence_sha256": "978d35794ae4a19b24f7e842d5ddedff7397dd2f0f1222d8d4e9aaf6ced29f72", "artifact_boundary": "Hash retained from the August public brief; underlying evaluation artifacts are not distributed here." }, { "id": "V35", "date": "August 2026", "status": "measured", "build": "Historical WinoGrande object-state route", "title": "Fresh WinoGrande train-holdout gain", "question": "Does object-state reasoning improve a fresh holdout?", "method": "Frozen public-knowledge fallback and exact object-transition route on 1,024 never-used train-split rows.", "result": "523 / 1,024 became 550 / 1,024. The object route answered 26 / 28 resolved rows correctly; actual overrides produced 14 gains and 2 losses.", "interpretation": "A measured improvement on the named historical holdout.", "limitation": "This is not an official WinoGrande validation score or a protocol-matched comparison with external models.", "artifact": "/brain/brain-benchmark-brief-2026-08.json", "artifact_label": "V35 public benchmark record", "artifact_boundary": "Cohort, controls, and recorded hashes are retained in the August brief." }, { "id": "FRONTIER-20260902", "date": "September 2, 2026", "status": "measured", "build": "Pre-Final-R6 / frozen five-task domain", "title": "Correct-candidate evidence frontier", "question": "Does the legal evidence pool contain the correct candidate?", "method": "17 target-free evidence sources across 2,378 frozen ARC-Challenge, BoolQ, HellaSwag, PIQA, and WinoGrande cases.", "result": "2,354 / 2,378 correct candidates present: 98.99% aggregate and 98.14% macro coverage; 24 cases lacked a correct source proposal.", "interpretation": "A strong information-availability result that informed subsequent cortex development.", "limitation": "Evidence coverage is not final-answer accuracy. ARC-Challenge here is distinct from the historical ARC-Easy comparison.", "artifact": "/brain/brain-research-brief-2026-09.json", "artifact_label": "September 2 research brief", "artifact_boundary": "Dated historical summary; it does not describe the current architecture." }, { "id": "V3", "date": "September 2, 2026", "status": "rejected", "build": "Pre-Final-R6 reconstruction and arbitration", "title": "Candidate-plus-parent union", "question": "How much correct evidence did V3 preserve?", "method": "Compare the candidate-plus-parent union with the legal evidence pool on the same frozen domain.", "result": "2,101 / 2,378: 88.35% aggregate and 82.63% macro, with a 253-case recoverable-evidence gap. The run was terminated and rejected.", "interpretation": "Historically informative. Later cortex work addressed reconstruction, relation, authority, and final selection.", "limitation": "The rejected run is not the current Final-R6 baseline or a fresh Final-R6 score.", "artifact": "/brain/brain-research-brief-2026-09.json", "artifact_label": "V3 historical research record", "artifact_boundary": "The rejected result remains preserved in its original record." } ], "access": { "status": "in preparation", "request_email": "contact@rdbiotech.org", "proposed_chat_host": "chat.brain.rdbiotech.org", "public_chat_available": false, "evaluation_focus": ["Factual recall", "Reasoning", "Conversational continuity", "Memory behavior", "Explanation quality", "Failure modes"], "learning_policy": "Public conversations are not automatically admitted into Brain's persistent knowledge or training state. Learning remains governed separately." }, "claim_boundaries": [ "Benchmark results apply to the stated build and protocol.", "Final-R6 fresh evaluation is pending; historical scores remain a separate lineage.", "Evidence coverage is not equivalent to final-answer accuracy.", "Parameter-count comparisons do not establish protocol-identical model equivalence.", "QSA participation does not by itself establish quantum advantage.", "Scientific and biomedical applications are research directions, not clinical capabilities.", "Private weights, memory, topology, thresholds, credentials, and security protocols are not included." ] }