Unifying Flow and Shocks: The Dual-Representation Siamese TWE Architecture

Why care, dear reader? Everyone wants to do this thing called time-series forecasting. You did it in elementary school, and some of the highest-paid people on earth do it every day at hedge funds and banks, running the casino we call the stock market. We cannot look that far into the future. But what if we could?

What if we had a model that for all purposes was a real reflection of us—our brain mechanics—but faster, better, with less cognitive bias, less emotional reactivity, and fewer of the tribal and amygdalitic responses to the environment? What if we could actually predict the weather a year from now? What if we could tell when someone would have a heart attack based on their current heart rhythms? What if we could predict when the next war would happen, or when the stock market might crash, if these predictions were more than mere guesses?

We would take the "game" out of the system. We would level the informational playing field. That which becomes valuable then exists where the data does not peer: into your mind, dear reader. That final frontier, which is wholly your own. When we all stand on the shoulders of giants together, you may choose: will you be Icarus, or Acelestius, or Odypheus?

Will we be ready for a world where we actually do compete on merit? Where we compete along the pathways that make humans interesting, fascinating, and beautiful? And even if and when we cross that Rubicon, staring at the hollow ghost of ourselves and wondering how hollow it really is—or how hollow we are to have made it—we still feel invigorated. We feel the possibility of finally, for the human species, perhaps since Neanderthals walked this earth, not feeling so existentially, coldly, alone.

To frame this simple problem in front of us: how do we go about predicting the next token of the universe? Macbeth lamented that the sequence creeps in its petty pace from day to day, to the last syllable of recorded time. We build these world engines to trace that creeping pace. Historically, sequence modeling forces a choice between two representations of dynamical systems. Continuous flow or discrete steps. It is a fundamental design fork that requires more than just choosing a model—it calls for a complete rethink of how we represent state.

Continuous models (like CfCs and Neural ODEs) treat state as smooth, uninterrupted drift—think of a car cruising down a highway. They track long-term trajectories perfectly, but they panic when they hit a sudden speed bump or crash into a wall (a discontinuity). In mathematical terms, the differential equations try to compute infinite gradients at the edge of the void, breaking the solver. Discrete models (like Transformers or vector quantization) take snapshots instead, treating the highway as a series of frozen frames. They handle the crash just fine, but they lose the sub-step continuous drift between those frames, resulting in quantization error and drifting off the road entirely.

To bridge this gap, we designed the Dual-Representation Siamese TWE (DR-S-TWE). It processes inputs in representation superposition. The core idea is running parallel continuous and discrete encoders, fused by a surprise-sensitive neuromodulatory gate. The biological mapping (Acetylcholine/Norepinephrine routing) is an analogy for design intuition, not a neuroscience claim. For our purposes, this serves as a stateful representation layer for decision-making under high uncertainty.

Core mathematics

Siamese projection

For an observation $x_t\in {\mathbb{R}}^d$, the continuous stream projects directly to the continuous representation space:

The discrete stream soft-quantizes each dimension over $B$ bin centers $C\in {\mathbb{R}}^B$ using a Radial Basis Function (RBF) softmax. We flatten this distribution and project it to form the discrete embedding. Think of this as bucket-sorting with blurry edges (and a nod to Asimov's psychohistory): instead of saying the input is exactly $72^{\circ }$, we represent it as a distribution across overlapping bins ("Cool", "Warm", "Hot"). If the temperature suddenly jumps, the activation simply shifts smoothly between bins rather than forcing a linear projection to scale infinitely. However, some representations lead to more parameters, and for every new parameter, there are then exponentially more ways to overfit, and so on…

Neuromodulatory gating

A gating network calculates a dynamic routing coefficient $g_t$ based on the prior state and the current observation. Think of this as a surprise alarm in the brain: the model constantly makes a prediction. If the input is close to the prediction (low surprise), the alarm stays quiet ($g_t\to 1.0$), routing the input through the continuous path to track smooth drift. The moment the input deviates sharply (high surprise), the alarm trips ($g_t\to 0.0$), immediately bypasses the smooth continuous stream, and routes focus through the discrete path to absorb the shock of the discontinuity:

Closed-form continuous-time integration

The fused embedding integrates into the state $h_t$ using a closed-form analytical decay update. It is a stateful transition without the numerical instability of traditional solvers. (Think of it as a cup of hot coffee cooling down over time: instead of a computer running complex calculations at every microsecond to estimate its heat, we use a simple formula to solve for the temperature at any arbitrary future point. For our researchers, this means continuous conformal flow—like the gravitational warping in Interstellar, time becomes a coordinate we navigate seamlessly, rather than a step counter):

Implementation

The forward pass executes these steps sequentially: RBF binning, gate computation, superposition, CfC decay, and a gated blend of continuous and discrete decoders. A standalone PyTorch reference module implements this sequence as forward_step in our research branch. (Think of it as a standalone model file representing an earlier and simpler stage of development).

Full training and evaluation harnesses live in the time-series research repository. This includes the pre-training scripts, per-domain tuning, and the zero-shot bakeoff code: https://github.com/connerkup/stock_trading_agent/tree/research/thalamocortical-geometric-2026-07-04

Empirical benchmark performance

So why care? Dear reader, sequence calibration shapes the execution systems of our entire trading engine. It is not an academic exercise. First, a disclaimer: results below are from task-specific training of DR-S-TWE on each benchmark, compared to Amazon Chronos 8M in zero-shot mode. They do not describe the larger pre-trained checkpoint evaluated without fine-tuning. We want to evaluate the trajectory with analytical rigor and a willingness to challenge assumptions.

We evaluated tuned DR-S-TWE against Chronos 8M zero-shot on two non-stationary domains. The Lorenz chaotic attractor represents continuous dynamics with sensitive dependence. Melbourne daily temperature represents seasonal drift interrupted by sudden weather-front shocks—weather fronts moving in like silent observers.

On Lorenz, tuned DR-S-TWE edges Chronos zero-shot ($+0.9890$ vs $+0.9847$). Both models sit near the persistence baseline. The material gap appears on Melbourne. We see a $+0.196$ correlation improvement and a 52% reduction in MAE ($0.2299$ vs $0.4820$). This is consistent with routing sudden, discontinuous shocks through the discrete path while tracking seasonal drift continuously.

Research note: zero-shot pre-trained checkpoint

Separately from the tuned benchmarks, we pre-trained a $10\text{M}$-parameter TWE-Large checkpoint. This model was trained for five epochs on a synthetic multi-domain mixture. We evaluated it zero-shot against persistence and echo-state baselines. As a "Speaker for the Dead" (a nod to Card), this section reports the unvarnished truth of the model’s current limitations, not just its triumphs. This track represents early research, not the headline claim of this post. I’m not holding my breath either way (spoiler: zero-shot transfer is hard, and the void of out-of-distribution data is vast).

Zero-shot Melbourne and prognostics beat the echo-state network on correlation. Lorenz and long-horizon closed-loop rollouts still lag simpler baselines. Resolving these bottlenecks is the next step for deploying stateful sequence engines in complex decision environments. Full nine-domain numbers and configurations are in the repository report. Scaling pre-training and competition entry remain ongoing work. The future is arriving fast. We can either stand by as passive observers of non-stationary systems or step up as active architects of the models that navigate them. The journey ahead is uncharted for sure, and I’m not holding my breath either way. One thing is clear though: the dialogue between continuous flow and discrete shocks will be a defining narrative of our modeling efforts. Navigating that ocean will be the task of us, and the necessity of our next research sprint.

Collaboration and authorship note

This architecture was designed collaboratively. The high-level cognitive constraints—the subcortical thalamocortical pacemaker, Siamese TRN contrastive gating, and hyperbolic concept pathfinding—were mapped out by myself. The mathematical translations, PyTorch implementations (CfC, RBF softmax, and Poincaré operations), and pre-training pipeline code were expounded, compiled, and iterated across various research phases by agentic AI coding assistants (including Claude Fable, Claude Opus, Claude Sonnet, GPT 5.5, and Gemini/Antigravity).

Disclosure: The core SNeWE/TWE dual-representation continuous-discrete architecture and neuromodulatory surprise routing mechanisms described in this article are subject to U.S. Provisional Patent Application No. 64/108,066, filed July 9, 2026.