Reasoning Substrate Lab
The problem it addresses. Reasoning and agent pipelines that are not deterministic are difficult to audit, replay, debug, and certify: the same input can take different paths, and a failure cannot always be reproduced. RS makes execution state transitions deterministic, bounded, and inspectable.
This lab hosts simplified public demonstrations of the RS, HEG, and RPU architectures. Each toy makes one architectural idea visible and interactive. Deterministic toys use fixed update rules; toys with randomness use fixed seeds, so identical inputs reproduce identical results.
Does not show: benchmark results, production readiness, model quality, or deployment scale.
What is the HEG? The Human Expression Gateway transforms raw human expression — text, audio, gesture, prosody — into typed Cognitive Action Packets (CAPs) via a deterministic operator pipeline: SEGMENT → ACT → TARGET → EMOTION → PATTERN → CAP_BUILD.
What is the RPU? Reasoning Processing Units are a proposed class of hardware accelerators for running the RS update cycle directly in silicon. RPU Classic settles by contraction to a fixed point, described in US 2026/0330264 A1. RPU Quantum runs an ensemble of independently converging cores with best-state tracking. Hardware implementations are future targets, not existing products.
Testing Center
All modules below are TOY-grade — browser-runnable, fully sandboxed, no network calls. Click [ EXPAND ] on any toy to load and run it inline. Toys marked illustrative use authored dynamics; all are controlled reductions of the reference implementation, which is available for technical evaluation under NDA.
100× Deterministic Replay
Four featured tests — agent failsafe, full HEG→CAP→RS pipeline, RPU vs Unconstrained Baseline, and the RSv2 intent orchestration layer.
Wraps a multi-agent orchestration layer (PlannerAgent, DataAgent, ActionAgent) inside an RS v1 cycle. Visualizes state → forward model → plan scoring → collapse. Includes live tool-failure injection to demonstrate RS failsafe and recovery.
Full HEG→CAP→RS pipeline. Choose text, audio, or video presets and watch each expression processed through SEGMENT, ACT, TARGET, EMOTION, PATTERN, and CAP_BUILD into a typed Cognitive Action Packet. Includes audit envelopes and replay/diff.
Three update rules on the same kind of input: an unconstrained baseline (a simple recurrent map with injected noise, authored for contrast), RPU-Classic (a contraction-based update, which converges to a fixed point by construction), and RPU-Quantum (an ensemble descent with best-state tracking). Canvas animations show live energy and state variance for each. Illustrative architectural behavior, not hardware performance results; the baseline is not a real transformer.
Intent Agent → Planner Agent → Executor Agent pipeline on a shared RSv2 memory substrate. Features token budgeting, safety filters, and audit replay. Amber terminal aesthetic throughout.
Six long-horizon stability tests across up to 1M turns: constraint grammar enforcement, drift prevention, isothermal load balancing, multi-agent stability, and full audit & replay.
Translates semantic DSL instructions into typed RS operator graphs then into backend-specific output across compiler, telecom, banking, and networking scenarios.
Simulates CPU/GPU/memory/thermal/battery under four plan modes. Each plan has authored effects on the simulated state; the RS loop (forward model, score, collapse with hysteresis) selects the plan, so plan choice changes how the simulation evolves. Illustrative: the plan effects are authored, not measured hardware behavior.
Stresses the SEGMENT operator with nested intents, abrupt topic shifts, and contradictory instructions.
Targets the ACT operator boundary between COMMAND, ASK, and ASSERT. Exposes scoring heuristics used to disambiguate intent type.
Demonstrates the EMOTION operator at the CAP→RS boundary. High-arousal negative inputs trigger different constraint weighting than neutral equivalents.
Sends the same text through different synthetic audio and video contexts. The FUSION operator produces different CAPs depending on the combined modality signal.
*Illustrative: audio/video inputs are text-described stand-ins for a browser demo, not real signal processing. Shows methodology, not measured multimodal performance.
The same hostile packet stream through three response rules: naive accumulation, a hard clamp with decay (RPU-Classic), and a clamp plus invariant and load shedding (RPU-Quantum). Authored comparison; illustrative architectural behavior.
A 4×4 grid puzzle (all row and column sums must equal 2) solved by three engines. Classic applies greedy local improvement; Quantum adds random flips with best-tracking; the unconstrained baseline flips cells at random. Current and best-so-far violations are logged, and a 50-seed table compares the engines across random starts.
User-defined constraints compiled into E(x) energy function. Both RPU variants minimize from seeded pseudo-random starting states.
A single reasoning state traverses a visualized energy surface. Classic follows gradient descent; Quantum runs an 8-core ensemble with best-tracking (no exploration noise).
8×8 grid of cells descends energy simultaneously (cells evolve independently in this toy). Color shows each cell from high-energy (red) to stable (green) for all three engines; mean energy is logged per step.
Six agents navigate a shared space. All three engines receive identical starting positions and the identical mid-run goal change; distance-to-goal, motion and collisions are computed per tick, not asserted. Classic coordinates deterministically, Quantum adds seeded exploration noise.
Authored power and thermal curves for a GPU-style baseline versus an RPU-style energy descent. Bar charts show unitless curves.
*Illustrative: power/thermal curves are stylized for demonstration, not measured from real workloads. Shows methodology, not benchmarked performance.
Episode-based learning loop with fitness scoring and bounded parameter updates. Agents improve across episodes without gradient descent.
Lite Demo Engines
LITE demos are architectural miniatures — small, precise, and intentionally incomplete. They show how the system thinks, stabilizes, collapses, and converges, without exposing the underlying substrate or proprietary operator calculus. Each LITE is a functional slice: enough to demonstrate capability and behavior; not enough to reveal the crown jewels. LITE artifacts preserve selected architectural structures while omitting proprietary mechanisms, optimization strategies, and substrate-level implementation details.
Available for short-term exclusivity evaluation — contact matt@icpub.org.
Each LITE keeps the real structural elements: the state model, the update rules, the evolution loop, the trace format, and the measurable outputs. This makes them ideal for modeling, scaling tests, integration work, and technical evaluation, without exposing sensitive substrate-level mechanisms. They are small, deterministic, and fully inspectable.
In Development // Lab Machines
Full-scale systems, deep-architecture engines, and substrate-level frameworks. Not prototypes, not products, and not promises — active constructions, each representing a future layer of the cognitive stack.
Whitepapers
Foundational research and patent filings covering the RS, HEG, and RPU architectures.
Published: US 2026/0330264 A1 — “Hardware Processor for Deterministic Contractive State Evolution Using a Fixed Point Update Operator” (U.S. Appl. No. 19/707,753) | US 2026/0320444 A1 — “Systems and Methods for Variational, Geometry Regularized, Identity Preserving Cognitive State Evolution” (U.S. Appl. No. 19/675,481) | US 2026/0289346 A1 — “Systems and Methods for an Operator Separated Cognitive Architecture for AI Reasoning” (U.S. Appl. No. 19/543,866). Patents pending: U.S. Appl. No. 19/543,534 | U.S. Appl. No. 19/543,514 | U.S. Appl. No. 64/093,303 (provisional) Full filing record →
Contact
This is a research lab — not a production system. All toys are simplified public demonstrations: self-contained simulations intended to expose architectural behavior.
For licensing inquiries, acquisition discussions, or technical evaluation access:
No data is transmitted. The hero stores its recovery trace locally in your browser. No network calls are made from any toy; the page’s content-security policy blocks them.