Automate Capture Research
Independent research into how AI systems remember, combine capabilities, and reason. Papers, experiments, and software from a small applied research studio.
Using known structure to guide vector retrieval.
A study of retrieval through a session–document–chunk hierarchy, with reported results on synthetic data.
The research record
Poisoned models passed a routine capability gate: no accuracy change above 1.9 pp. Counting the individual answers that changed separated them at no extra cost. A study of detection, with its negative results and harness failures reported alongside.
Hidden state injection improved reported accuracy on a direct medical reasoning task for one Phi-to-MediPhi model pair. Transfer failed in the other three tested pairs.
Forced-routing tests compare expert scores with the learned router in Phi-mini-MoE. The reported correlations are small, but isolating one expert does not measure the full routing objective.
Activation profiles suggest that some feed-forward neurons could be skipped for particular domains. The reported compute estimates and output similarity do not establish measured inference speed or task accuracy.
Claude Opus 4.6 generated candidate solutions to ten research-level mathematics problems. This report records prompting behavior and output completeness; mathematical correctness remains unverified in this publication.
ARMS explores storing and retrieving model states through spatial organization. Its projection-based ARM prototype is an early experiment; the proposed native-state architecture has separate validation requirements.
Questions we investigate
How can stored structure and attention states help a model use information beyond its immediate context?
HAT · ARMS 02What transfers between model representations, and how can we distinguish transfer from task-specific fitting?
Blades 03How do routing, sparsity, and evaluation choices affect the behavior we observe?
Routing studies · Sparse pathwaysFrom research to practice
Tools and interactive experiments make the ideas available for closer inspection.