Honestly don't remember inserting previous README.md, uploading prefer md.

Research notes/GPT5.4pro_research_framing_metis_spikenaut.md

@@ -0,0 +1,70 @@
+# Research Framing: Metis, SAAQ 2.0, and Spikenaut
+
+## Core thesis
+
+Metis is best framed not just as a hybrid model, but as a **scientific instrument for discovering sparse routing laws** in dense Mixture-of-Experts systems. It uses MoE latent telemetry, hardware-aware fatigue dynamics, and SNN routing to expose structured attractor-like basins; those discovered laws are then intended to transfer into **Spikenaut**, a pure native SNN built from scratch. The current foundation of SAAQ 2.0 is an L2-normalized voltage bound,
+
+\[
+V_{\text{norm}} = \frac{V}{\sqrt{\sum V_i^2 + \epsilon}}
+\]
+
+which acts as a global pressure limiter on latent voltage and prevents runaway domination by a single route. 
+
+## AI / neuromorphic framing
+
+### Short abstract
+
+We present a hardware-grounded neuromorphic distillation framework in which a spiking routing engine is used to probe the latent structure of dense Mixture-of-Experts language models. Rather than distilling weights into a smaller ANN, the system records route-level telemetry under bounded voltage and fatigue-like constraints, then uses those traces to discover reusable equations for sparse computation. In early experiments with OLMoE, L2 normalization breaks routing collapse and reveals prompt-dependent routing neighborhoods: abstract English concentrates in a high-index basin, while Rust syntax and mathematical statements cluster into a lower shared logic band. The resulting telemetry is intended to drive symbolic regression and transfer the discovered laws into Spikenaut, a pure SNN implementation. 
+
+### One-sentence contribution
+
+This work treats **quantization as a discovery instrument**: a way to extract equation-level sparse routing laws from dense MoE systems and port them into a native SNN. 
+
+### Why this matters to AI
+
+The research matters because it shifts the goal from “compress a model” to “discover the dynamical laws that make sparse computation work under real hardware pressure.” Your docs already show a chronological progression from routing collapse, to stabilized attractor formation under L2 normalization, to prompt-family-specific routing separation. That means the quantized SNN layer is functioning as an analytic probe, not just an inference wrapper. 
+
+### Why this matters to neuromorphic engineering
+
+The framework is naturally relevant to neuromorphic research because it emphasizes sparse event-like routing, bounded activity, fatigue-like adaptation, and hardware-in-the-loop constraints. The RE4 path-tracing telemetry is especially unusual and interesting here: your own notes say it supplied the dynamic thermal and power fluctuations that informed the fatigue limits used in the `tick_gpu_temporal` loops of the Rust/CUDA engine. 
+
+## Neuroscience-respectful framing
+
+### Short abstract
+
+This project is best described as a **bio-inspired dynamical systems testbed**, not as a literal model of the brain. A global normalization law and fatigue-like adaptation are imposed on a spike-routed latent system derived from dense language-model embeddings. Under those constraints, different task families appear to settle into different routing neighborhoods, suggesting attractor-like partitioning of state space. The scientific value is not that the model “is biological,” but that it provides a controllable system for studying how normalization, activity bounds, and repeated-use fatigue shape structured routing dynamics. 
+### Language to use with neuroscientists
+
+Use phrases like **attractor-like basin**, **fatigue-like adaptation**, **homeostatic normalization**, and **task-conditioned state-space partitioning**. That language is strong but still disciplined. It invites comparison to attractor networks and adaptive routing without claiming that code and math literally live in fixed brain regions.
+
+### Language to avoid
+
+Avoid saying that the system has discovered “how the brain routes semantics,” or that the routing neighborhoods are direct analogs of cortical areas. The clean claim is that this is a hardware-grounded, bio-inspired model in which structured inputs generate reproducible basin-level routing patterns under normalization and fatigue constraints. 
+
+## How original the work is
+
+Methodologically, the work is highly original because it combines three ingredients that are rarely unified in one pipeline: MoE latent telemetry, hardware-derived fatigue priors, and symbolic regression intended to recover closed-form SNN equations for transfer into a native spiking model. Empirically, it is still early-stage: the major next step is to show that the same basin grammar and fatigue laws survive across multiple MoE families and remain useful once implanted into Spikenaut. 
+
+## Why the Metis -> Spikenaut pipeline matters
+
+The strongest opportunity is not ordinary weight distillation. It is **dynamical distillation**. Metis can serve as the teacher/probe layer that reveals routing basins, transition rules, and fatigue cadence; Spikenaut can then test whether those laws can be implemented directly in a native SNN without carrying over the dense teacher. Your dataset README already frames this relationship clearly: Metis proves the math, while Spikenaut implements it natively. 
+
+## What would fascinate scientists
+
+A computational neuroscientist would likely find the project interesting if you present it as a sparse dynamical system with attractor-like partitioning and fatigue-like adaptation. A neuromorphic engineer would likely care about the hardware-in-the-loop aspect, especially the use of telemetry-derived heartbeat constraints. A biologist may also be interested, but usually only if the bio claims remain modest and testable. The strongest hook is that the system may reveal a general principle: structured tasks can settle into adjacent routing neighborhoods when activity is globally bounded and repeated use incurs fatigue. 
+
+## Copy-paste positioning statements
+
+### AI / neuromorphic version
+
+Metis is a hardware-grounded spiking Mixture-of-Experts discovery framework that uses SNN routing telemetry to uncover sparse laws in dense MoE models. Rather than distilling weights into another black-box network, it captures route-level dynamics under SAAQ 1.0 constraints and uses those traces to recover reusable equations for native spiking systems. Those equations are then transferred into Spikenaut, a pure SNN built from scratch. 
+
+### Neuroscience-respectful version
+
+This project studies how global normalization and fatigue-like adaptation shape task-dependent routing in a spike-routed latent system. The goal is not to claim a biological replica of cortex, but to build a controllable model system in which structured inputs produce reproducible attractor-like routing neighborhoods that may inform sparse computation and neuromorphic design. 
+## Suggested next-step claims to validate
+
+1. Basin structure replicates across different MoE families.
+2. Basin transitions and dwell times can be captured by symbolic regression.
+3. The discovered equations transfer into Spikenaut and reproduce the same routing grammar.
+4. The hardware-derived fatigue cadence improves stability or efficiency relative to naive quantization.