Evaluating long-horizon model behavior
A proposed evaluation framework for goal retention, evidence tracking, recovery, and calibrated escalation across multi-step technical tasks.
Working notes
The research record starts with questions.
A place for technical agendas, evaluation designs, and observations as LabOne37 develops. These entries describe planned inquiry, not completed publications.
Format /
Research agenda
Claims /
No released results yet
Updated /
2026
01 / Current agenda
Problems worth making precise.
Frontier labs separate fast-moving updates from formal publications. Here we begin with an honest working record and will add evidence as it becomes available.
Useful intelligence per unit of compute
Metrics that connect capability with latency, memory, energy, and total operational cost rather than treating benchmark score as the only outcome.
Legible autonomy for tool-using systems
Design questions around authorization, plan inspection, intervention points, execution traces, and bounded recovery.
Odin LLM Gateway
Product work on identity, PII detection, policy enforcement, token optimization, provider routing, and request-level observability.
Specialization without brittle intelligence
Methods for adapting models to technical domains while preserving calibration, transfer, provenance, and dependable behavior.
Publication standard
Results will be labeled as results.
Future technical reports will include methods, evaluation conditions, limitations, and links to artifacts where release is appropriate. Until then, these entries remain research agendas.
Discuss the agenda →