RESEARCH

Research

From computation in hidden states to reliability in the real world—and the research process itself.

All 10 research projects

Original concept artwork for this project

Embodied AI safety·v0.3 · Open-source execution verification workbench

Sentinel-EVC

Final-action validation and verifiable execution

Final-action checks, incremental certificate reuse, pre-motion integration checks and signed execution records.

Explore research
Verifier-Bit Ledger concept artwork

Information theory & reasoning · ICLR under review

The Verifier-Bit Ledger

Verifier-in-the-loop reasoning

How much improvement can verifier feedback buy? An information ledger connects uplift ceilings, ghost baselines and executable audits.

Explore research
Original concept artwork for this project

Representation & Reasoning·Ongoing research

Attractor-JEPA

Compositional computation in latent representations

Explores compositional computation in latent representations, focusing on reasoning depth, state stability and learning.

Explore research
Original concept artwork for this project

Representation & Reasoning·Training mechanisms under investigation

CIDS-GA

Causal credit and capability consolidation

Using active causal experiments from the same state to identify the contribution of intermediate operations, then studying how repairs can become model capabilities.

Explore research
Original concept artwork for this project

Representation & Reasoning·Theory and manuscript work in progress

Model collapse and feature detectability

Model collapse and feature detection

Investigating the limits of recursive training, rare information, and feature detection, including information that remains hidden within models.

Explore research
Original AutoResearch concept illustration: a question, evidence cards and research actions.

Research Automation·Local workbench and controlled experiments in progress

AutoResearch

Automated research and recursive improvement

Connecting questions, literature, search, refutation, and frozen experiments so that sustained exploration accumulates evidence future work can build on.

Explore research
Original concept artwork for this project

Optimization & search · Paper submission to IEEE TEVC forthcoming·Working manuscript and research implementation

Diagnosis-First Optimization · DFO

Pricing exploitable structure before black-box search

Before black-box search, ask what structure an instance offers, what finding it costs, and when the evidence calls for an uncertain answer.

Explore research
Original initialization and signal-propagation concept illustration: filaments passing through layered channels.

Representation & Reasoning·Two theoretical manuscript and numerical-study branches

Initialization and Signal Propagation

Initialization and Signal Propagation

Study what individual units and backward signals experience beyond average activation scale, and how depth changes input distinguishability.

Explore research
Original label–trajectory channel concept illustration: an observational lens and mirrored paths.

Embodied AI safety·Information-theory working manuscript and self-contained computational witnesses

Reading Tasks from Behaviour

Reading Tasks from Behaviour

When a task label is observed only through a trajectory, what information comes from geometric differences and what comes from time direction?

Explore research

Research manuscripts

Research manuscripts currently in development. These are presented as working papers.

Conceptual illustration of policy-interface information
VLA runtime detection · Working paper

Interface-Attached Information Bounds for Runtime Detection in VLA Policies

Studying how information available at policy interfaces constrains runtime failure detection.

Related research
Conceptual illustration of latent features and detection conditions
Representation & feature detection · Working paper

Two Regimes of Feature Detection in Superposition: Quartic Dark Matter and the SAE Error Floor

Exploring feature-detection conditions and error bounds in superposition and sparse autoencoders.

Related research