Brain / QELM research system
Brain coordinates hybrid model execution, persistent memory, continual learning, and QELM research on locally controlled infrastructure. The working system is in active development and is not open as a public chat or node download.
Research discipline
Phase 37Z fixes the architecture, corpus, optimizer, token budget, and seeds before training. QELM is compared with a conventional decoder-only control at nearly the same parameter scale, and public scores cannot be used to select or repair either model.
System atlas
Brain separates routing, quantum execution, memory, learning, and governance so each path can be inspected and tested independently.
Control plane
Brain binds tokenization, hybrid transformer blocks, sampling, model routing, memory, and the active quantum backend. Runtime modes keep local execution primary while optional teacher systems remain explicit and traceable.
Current experiment
The sealed Phase 37X route retained 10,173,600 associative-memory parameters that its memory-free benchmark path did not read. Phase 37Z replaces that inactive table with a three-layer causal GRU and a residual MLP using the exact same parameter count.
No parameters are added by the replacement.
Memory topology
Each store has a distinct role. Durable writes move through governance instead of treating every generated response as memory.
Canonical writes carry provenance, confidence, validation state, and an identified writer.
memory/ + core/memory_governance.pySQLite FTS5 provides durable, searchable records for local semantic recall.
core/memory_manager.pyA separate associative path experiments with Grover-oriented retrieval.
core/quantum_memory.pyQELM sub-bit state is retained through a dedicated memory and controller pair.
core/subbit_subspace_memory.pyExtracted entities and relationships persist as structured graph data.
core/knowledge_graph.pyPhase 37Z protocol
The comparison contract is designed to fail closed when corpus, contamination, retention, or execution requirements are not met.
Frozen before training
Phase 37Z requires the checksum-pinned raw WikiText-103 source. Its smaller prepared cache is explicitly forbidden for this run.
Required gate
The exact training sources are audited against 91,008 examples across 74 expanded benchmark tasks before either model can train.
One QELM + one control
Both routes share the corpus, optimizer contract, token budget, and context window. Seed searches and post-benchmark retraining are disabled.
Reject on regression
The QELM candidate must retain canonical behavior and structured-safety performance while the reclaimed parameters receive real gradients.
No selection authority yet
Public benchmarks run only after internal acceptance. The contract does not authorize a universal architecture-superiority claim.
Public access
Brain's current quantum implementation is not distributed. The public record, related open repositories, lab tools, and later evidence remain accessible through the paths below.