rmems/Spikenaut-SNN

Other • Updated 26 days ago • 18

Note The experimental SNN itself — LIF neurons, e-prop/STDP training experiments, Q8.8 FPGA export targets.

Note v3 action-proposal trajectory dataset for the supervisor stack: state_telemetry + outcomes with chronological embargoed splits, deterministic teacher policy, axon-encoder encoding params. The learned policy only proposes; the deterministic safety filter is final authority.

Note Canonical bare-metal hardware corpus: RTX 5080 telemetry under max-settings gameplay (RE4 path-tracing seed; multi-title). Real timestamps + NVML throttle bitmask — the state signals Spikenaut's teacher needs.

Note SAAQ / SEMM research data: Semantic Attractor Clustering — an SNN routing LLM embeddings into distinct physical pathways — plus the symbolic-regression benchmark behind the SAAQ 1.5 adaptation law.