Final-R6 is Brain's accepted cortex baseline. It combines compact neural state with explicit memory, evidence, semantic reasoning, and exact QSA computation. Cortex repair is complete; the current research phase develops learning, governed recall, knowledge, and conversation.
Post-cortex training and intelligence development. Cortex repair is closed; anti-forgetting and Brain-native runtime authority remain active.
Mutable language state
29,930,048 scalars / 54 tensors
Persisted Final-R6 neural state
49,324,558 scalars / limit ≤50M
Protected systems
Final-R6 / V22 / Query192
Public deployment
Conversational early access in preparation
Final-R6 architecture
A compact cortex inside a larger cognitive system
Brain does not ask one neural weight matrix to store, retrieve, reason, verify, remember, and govern everything. The cortex works inside a controlled architecture whose other parts remain explicit and testable.
Final-R6 / accepted baseline
Cortex repair is closed. Final-R6 is the accepted baseline.
Final-R6 brings the completed cortex repair into a fixed architecture for post-cortex learning. Its persisted neural state contains 49,324,558 scalars, including 29,930,048 mutable language scalars across 54 tensors. Training develops language and intelligence under Brain-native authority while established routes remain protected.
01Cortex repair
02Final-R6 acceptance
03Governed training
04Fresh evaluation
Verified or implemented
49,324,558 persisted neural scalars within a 50M limit
29,930,048 mutable language scalars / 54 tensors
Final-R6, V22, and Query192 protected
Anti-forgetting and route preservation active
Memory + evidenceRetrieve with source identity
Semantic stateBind entities, roles, and constraints
HypothesesMaintain competing explanations
World + proceduresConstruct candidate outcomes
QSA computationExecute exact admitted state work
Brain authorityDecide, record, or reject
Benchmark observatory
Research results, with their lineage intact
August benchmark results and the September 2 evidence study retain their original cohorts and controls. Fresh Final-R6 evaluation follows the current post-cortex training sequence.
pending fresh evaluation
Final-R6 fresh benchmark: Pending
ARC-Easy / PIQA / WinoGrande. The historical ~30M benchmark record is preserved independently. Experimental intermediate scores do not replace that record or supply a fresh Final-R6 result.
Historical correct-candidate availability / includes ARC-Challenge
Correct evidence was available in 2,354 of 2,378 cases.
Across the frozen domain, at least one legal source proposed the correct candidate in all but 24 cases. This measures information coverage inside Brain's evidence state, not final-answer accuracy. Subsequent cortex development addressed the reconstruction, relation, authority, and final-selection problems exposed by this study.
012,378
Frozen domain
Target-free internal validation cases
022,354
Correct evidence present
98.99% aggregate / 98.14% macro
032,101
V3 candidate + parent union
88.35% aggregate / 82.63% macro
04253
Recoverable evidence gap
Correct evidence not preserved by V3
ARC-Challenge52 / 5594.55%
BoolQ625 / 625100.00%
HellaSwag523 / 54496.14%
PIQA575 / 575100.00%
WinoGrande579 / 579100.00%
Historical fixed-parameter records / August 2026measured
200 fresh ARC questions
Grounded evidence doubled the fresh ARC result.
Inherited Brain43 / 200
Measured result92 / 200
Measured change+49 correct / +24.5 points
What changed
Candidate-specific evidence was evaluated through question anchoring, candidate relation, and grounded support before the answer was selected.
No retraining
No target labels in the reasoning path
Protected Brain state unchanged
No donor model at runtime
QSA and the information-identical classical product selected the same answers. This proves a strong architecture gain, not a uniquely quantum accuracy advantage.
Historical ~30M Benchmark Lineage
ARC-Easy / PIQA / WinoGrande
Historical profile~30M Brain
Overlaps published performance ranges of several larger reference models. Protocols are not identical.
Task identityARC-Easy
This table uses ARC-Easy. The September 2 five-task evidence study uses ARC-Challenge.
Current architectureFinal-R6
49.32M persisted neural scalars, including 29.93M mutable language scalars.
Fresh Final-R6 resultsPending
Evaluation follows the current training sequence. Historical scores do not transfer to a new build.
GPT-NeoX-20B's published table reports 41.0% ARC-Challenge, 69.4% BoolQ, 71.4% HellaSwag, 77.7% PIQA, and 66.0% WinoGrande. Google's Gemma 2 27B model card reports 71.4%, 84.8%, 86.4%, 83.2%, and 83.7% respectively under its stated task protocols, a simple 81.9% mean. These are scale markers, not a protocol-identical comparison.
Final-R6 is the accepted architecture. Earlier accepted, measured, and rejected experiments remain part of its history. Each record identifies the question, method, outcome, and limits of the available evidence.
September 2026 / FINAL-R6Fixed cortex baseline acceptedaccepted
Build: Final-R6
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.
The rejected result remains preserved in its original record.
Research chronology / August to September 2026
From accepted training to Final-R6
The August build established end-to-end learning. The September evidence study informed cortex repair and Final-R6 acceptance. Post-cortex training and governed-recall development are now active.
Aug 2001
Structured procedure layer sealed
A single fresh 2,120-row blind closed the accepted structured cognition fallback with its own immutable evidence record.
Aug 2102
Brain completed accepted training
The Brain-owned state completed 1,024 optimizer steps, reloaded exactly, passed 64 / 64 whole-Brain acceptance, and retained 106 / 106 prior behaviors.
Aug 2203
Fixed-parameter benchmark gains replicated
Fresh ARC, PIQA, HellaSwag, and WinoGrande routes improved without increasing the neural parameter budget or reconnecting the teacher runtime.
Sep 0204
Five-task evidence frontier measured
The frozen 2,378-row domain reached 98.99% correct-candidate evidence coverage. Rejected V3 preserved an 88.35% union, exposing the reconstruction and arbitration gap at that stage.
September05
Five-task capability training
Capability training extended the research program across five benchmark families and informed the subsequent cortex investigation.
September06
Whole-cortex repair completed
Failure isolation and repair addressed reconstruction, relations, authority, and final selection. Cortex repair closed.
September07
Final-R6 accepted
The fixed cortex baseline was accepted with 49,324,558 persisted neural scalars inside the 50M architecture limit.
September08
Brain-native learning established
Post-cortex learning and exact Final-R6 training execution were developed under Brain-native authority.
Current09
Training, recall, and route preservation
Full training, anti-forgetting, governed recall, conversational integration, and Learning Center knowledge integration are active.
Next10
Fresh Final-R6 benchmark
ARC-Easy, PIQA, and WinoGrande will receive a fresh evaluation after the current training sequence.
Next11
Conversational early access
A limited outside-testing release is in preparation, with a public build identifier and separately governed learning.
Longer-term evaluation target
The earlier research roadmap proposed at least 90% executable performance on a frozen five-task suite and comparison with 70B+ reference models. These are research targets. Fresh results, equivalent protocols, and matched controls are required before assigning any performance class to Final-R6.
Brain / QELM / QSA
Connected systems, separate authorities
QELM is the language-model research foundation. QSA supplies exact quantum-state execution. Brain owns language, memory, evidence, learning, checkpoint adoption, and the final decision path.
The Final-R6 cortex baseline is accepted. Governed post-cortex learning develops language, recall, and knowledge while anti-forgetting and route-preservation controls protect established behavior.
AreaEstablished capabilityActive frontier
Brain-owned model
Final-R6 accepted; 49,324,558 persisted neural scalars, including 29,930,048 mutable language scalars across 54 tensors
Complete post-cortex training within the 50M persisted-neural-scalar limit
Language ingress
Compositional task, role, relation, constraint, evidence, option, and output binding
Develop conversational continuity, explanation quality, and post-cortex language behavior
Quantum-state work
QSA 0.2 structured execution, branch-state computation, state transport, and matched controls
Prove when QSA adds runtime, representation, or cognitive value beyond an identical classical control
Memory + evidence
Working, episodic, semantic, and procedural memory with provenance, consolidation, and rollback boundaries
Develop governed recall, Learning Center integration, and knowledge use while preserving established routes
Evaluation
Historical benchmark lineage and the September 2 evidence frontier retained with build-specific limits
Fresh Final-R6 ARC-Easy, PIQA, and WinoGrande evaluation after training
Scientific translation
Causal research beyond plausible text
Brain's explicit hypotheses, provenance, procedures, and world states are being developed for scientific relationships where the path to a conclusion matters as much as the conclusion itself.
Example relationship path
Compound
Target
Pathway
Cellular effect
Tissue effect
Phenotype
Intervention outcome
Research targets
Drug-target discovery
Pathway analysis
Multi-omics integration
Biomarker discovery
Treatment-response modeling
Rare-disease hypotheses
Protein-function relationships
Mechanistic research
These are research directions, not claimed medical discoveries, validated clinical tools, or treatment recommendations.
Evaluation contract
How a Brain result becomes evidence
A finished run is not automatically a capability gain, and a QSA-backed gain is not automatically quantum advantage. The record follows the exact test that was performed.
01
Freeze
Bind the method, runtime, knowledge identities, parameter state, and complete predictions before target access.
02
Score
Open the target vault once, compare against the inherited Brain answer, and record gains, losses, ties, and coverage.
03
Control
Run information-matched classical, dephased, reset, or fallback controls appropriate to the mechanism being tested.
04
Protect
Verify accepted Brain state, protected checkpoints, donor disconnection, and production authority remain unchanged.
05
Retain
Keep positive, neutral, failed, and revoked paths in the same evidence history so a weak route cannot return as a claim.
Claims & limitations
Research interpretation
Brain is an experimental cognitive architecture under active development. Benchmark results apply to the stated build and protocol. Evidence coverage is not equivalent to answer accuracy. Parameter-count comparisons do not establish protocol-identical equivalence. QSA participation does not by itself establish quantum advantage. Scientific and biomedical applications are research directions, not clinical capabilities.
Experimental Brain access
Talk to Brain / early access in preparation
A limited conversational release of Final-R6 is being prepared for outside testing. Early participants will help evaluate factual recall, reasoning, conversational continuity, memory behavior, explanation quality, and failure modes.
