Read a paper as a set of claims.

This is the same reference used in Technical mode. Expand it whenever you need the exact numbers, assumptions or further explanation.

Start with the question

A headline promises a robot that understands anything. Its experiment reports a few named tasks on one platform. Which sentence would you put in an evidence table?

Mechanism and worked comparison

A claim is a statement whose support must be identified. A mechanism describes how a method is meant to work; an outcome records what was measured. A prediction model might minimize a representation loss while a controller fails its task. Neither metric can replace the other. Start by drawing three boxes: data enter the mechanism, the mechanism produces an output, and an external evaluator judges that output. Put the paper’s actual comparisons beside those boxes.

The cards link dated primary sources and contain short authored summaries, not copied result tables. For example, an action sequence predicted from demonstrations and a latent dynamics model used for planning solve different intermediate problems. A reported benefit on one benchmark cannot establish a universal claim. The statement “works on all unfamiliar tasks” needs far more evidence than a selected success video.

Annotate what was varied, what stayed fixed, which observations were available at decision time, and which failures were counted. When the source does not answer a question, write “not established by this evidence” instead of filling the gap with a plausible story. Your annotation is a reading exercise, not a reproduction or a human-reviewed literature assessment.

Try and explain

Choose a paper card and record its input, prediction target, action rule and evaluation setting. Classify the universal claim, then explain one test that could disprove it.

Open the reproducible study project or use the implementation studio.