Notes from the field.
Ideas on better data, useful benchmarks, and the questions worth asking before you trust a number.
Introducing EmbodiedTrace: an inspectable path from observation to action
Explore a 100-episode sample connecting visual observations, generated subgoals and structured robot actions — and the questions to bring to a pilot discussion.
Introducing Datumwell
A place to find your reference point. Why we’re starting with sources, context, and a deliberately small collection.
What makes a dataset useful?
Quality starts with the question. A practical guide to provenance, coverage, definitions, and access.
Before you compare, check the conditions
A benchmark result has a scope. Making a useful comparison means understanding what stayed the same.