Investigating artificial intelligence beyond Assumptions.
We isolate how intelligent systems reason, learn, and break under genuine stress—retaining negative outcomes, releasing live testbenches, and publishing reproducible experimental trails.
Investigating systems where outcomes remain unvarnished.
Each program starts with a falsifiable hypothesis. We evaluate baselines against noisy real-world conditions and retain every negative result.
Robust Reasoning RL
When verifiers are corrupted, biased, or noisy, what does a reasoning model actually learn? We isolate reward hacking by measuring uncertainty-aware reward shaping against GSM8K baselines.
Quantum Advantage Simulator
Testing neural quantum states, VQE, and QAOA against exact classical references without presuming an asymptotic advantage.
ATLAS Manifold Engine
Evaluating persistent topology for multi-hop retrieval. Evaluated on HotpotQA; graph manifolds failed to outperform direct cosine similarity.
NeuroSim Dynamics Platform
Multi-compartment Hodgkin–Huxley dynamics connected directly with continuous-time synaptic plasticity and scientific visualization engines.
AgentWatch Instrumentation
Full-fidelity provenance, execution replay, and state divergence tracing for autonomous AI-native development workflows.
Interactive Empirical Testbench.
Manipulate independent variables in real time. Watch how reward hacking emerges under corrupted verifiers, where quantum simulators meet classical tensor networks, and how neurons fire under current injection.
Percentage of training batch verifier labels replaced with adversarial noise.
We dismantleintelligence to its computational foundations,leaving no assumption unchallenged.
We investigate how computational systems reason, adapt, and decompose under genuine stress. Rather than accepting optimistic benchmarks or polished vendor summaries, our research isolates fundamental algorithmic limits, measures verifiable failure rates, and open-sources every artifact so anyone can verify our conclusions directly.
Evidence over intuition. Every hypothesis tested to failure.
Standard literature often hides negative outcomes, selectively tunes seeds, and overclaims generalization. Shinraya Tech Labs archives complete runs, unadulterated failure modes, and open benchmarks.
Isolating Reward Drift Under Corrupted Ground Truth
Reinforcement learning with generative reward verifiers degrades quickly when reward models are queried out of distribution. We systematically inject adversarial labels, measuring policy entropy and reward divergence across verified math benchmarks.
Demarcating Genuine Advantage from Optimization Artifacts
Quantum algorithms frequently benchmark against unoptimized classical baselines to demonstrate artificial speedups. We pair exact tensor network contractions and state vector references against VQE and QAOA formulations to record the real computational boundary.
Preserving Hypotheses That Failed in Empirical Tests
In the ATLAS evaluation, high-dimensional topological invariants and persistent homology failed to surpass direct dense vector embeddings on multi-hop QA tasks. We published the negative findings alongside complete HotpotQA validation runs.
Infrastructure built to render theory executable.
Research cannot advance without rigorous instrumentation. Hover or select any system vector below to inspect its architecture.
RepoPilot
Autonomous repository modification and loop evaluation.
A code-modifying agent that iterates against test feedback, verifies execution boundaries, and checks invariants autonomously.
VEX
PulseStack
FluxCore
Good questions deserve independent reproduction.
Clone our experimental repositories, challenge our evaluations, or join our Discord community to discuss AI research and systems engineering directly.