Empirical Research Programs

Follow the question. Keep the evidence.

Five independent lines of inquiry into machine reasoning, agent memory bounds, quantum boundaries, geometric representations, and biophysical networks. Each project makes its scope and limitations explicit.

Jump to Program:Robust Reasoning RL ↓ACI Architecture Framework ↓Quantum Advantage Simulator ↓ATLAS ↓NeuroSim ↓
Reasoning & reinforcement learning

Robust Reasoning RL

“When the reward is wrong, what does a reasoning model learn?”

Framework and pilot complete · final GPU study blocked
Source CodeResearch Report
Scope & Methods

Matched baseline and robust GRPO training, uncertainty-aware reward shaping, controlled verifier corruption, and independent mathematical evaluation.

Observed Evidence

The repository documents pretrained GRPO training and save/reload runs, independent GSM8K evaluation, model qualification, and a frozen execution package.

Limitations & Boundaries

The primary hypothesis is unestablished. Pilot runs and software fixtures do not establish improved reasoning or reduced reward overfitting. The final GPU study remains externally blocked.

Agent memory & attention bounds

ACI Architecture Framework

“How can long-horizon agents bound context so tool output stops crowding out the goal?”

v0.1.0 reference package · live probe validated
Source CodeResearch Report
Scope & Methods

Softmax Inundation Bound, Generational Context (Gen-0/1/2), Attention-Weighted Eviction (AWE), deterministic CCB-001 simulation, and live transformer attention probing on Qwen2.5-0.5B.

Observed Evidence

The repository documents formal bounds, 11/11 passing unit tests, -97.5% token compression in CCB-001, 8/8 CIC governance scenarios, and empirical Qwen2.5-0.5B eager forward-pass attention retention.

Gen-0 EphemeralTTL & Distillation (-97.5%)
Gen-1 EvaporativeAWE Attention Cache
Gen-2 TenuredPinned Invariants
Live Qwen2.5-0.5B ProbeStable ~0.406 Share
Limitations & Boundaries

Softmax inundation bounds assume uniform projection energy bounds; live transformer measurements were performed on 0.5B parameters; SWE-bench verified evaluations on frontier neural models remain in progress.

Quantum & classical computation

Quantum Advantage Simulator

“Where do exact, learned, and variational methods actually help?”

v0.1.0 simulator studies released
Source CodeResearch Report
Scope & Methods

Exact simulation, neural quantum states, VQE, QAOA, finite-shot noise mitigation, circuit search, and minimal-basis molecular chemistry.

Observed Evidence

Released CPU studies preserve configurations, per-seed results, checkpoints, and checksum manifests. Small-system comparisons include failed optimizations and harmful mitigation outcomes.

TORUS KNOT LATTICE [3D]
DRAG TO INSPECT
Limitations & Boundaries

No quantum advantage, GPU speedup, or hardware performance advantage is demonstrated. CUDA and IBM QPU runs were not executed for v0.1.0; the measurements are small-system CPU simulations.

Retrieval & geometric learning

ATLAS

“Does structure add useful signal beyond semantic similarity?”

Evaluated · central hypothesis not supported
Source CodeResearch Report
Scope & Methods

Controlled semantic, graph, and topology retrieval comparisons, alongside experimental local matching and compute-to-data prototypes.

Observed Evidence

Held-out HotpotQA and MuSiQue evaluations did not establish an improvement over semantic retrieval. Validation selected cosine-only retrieval as the primary V2 path.

Limitations & Boundaries

Evaluation uses internal held-out subsets of public development data, not official hidden-test scores. The result concerns the tested formulations, not all topology methods. Local prototypes have no established privacy guarantee or independent security audit.

Computational neuroscience

NeuroSim

“How can neuron and network dynamics be explored across scales?”

Public implementation · scientific validation unverified
Source Code
Scope & Methods

Repository-described Hodgkin–Huxley dynamics, multi-compartment models, synaptic plasticity, network simulation, and Python/C++ visualization and compute paths.

Observed Evidence

The public README describes the architecture, models, interfaces, and simulation commands. These document intended and implemented scope, rather than a measured scientific result.

Limitations & Boundaries

The reviewed README does not establish biological validation, human-brain-scale performance, or verified MPI/CUDA scaling. No such results are claimed here.

Editorial summaries reviewed against public repository documentation on 11 October 2026. Independent reproductions are welcome through our issue trackers.
Open Verification & Community

Good questions deserve independent reproduction.

Reproduce an experiment, challenge an assumption, or join the discussion with fellow researchers on Discord and GitHub.

Join Discord (UT9uaeY46e) Contribute to the WorkGitHub Organization