How to read a frontier paper in 30 minutes

10 atomic recall cards. Export to Anki and let spaced repetition do its slow work.

In Anki: File → Import, choose this TSV, set field separator to Tab, deck = Tokens to Embodied Minds · Ch 12, note type = Basic.

Front	Back
What are the six steps of the 30-minute paper reading protocol?	(1) Abstract: claim + number + baseline. (2) Figure 1: paper genre. (3) Main results table: numbers + fairness check. (4) Ablations: load-bearing components. (5) Limitations + appendix: disclosed and hidden failures. (6) Three prior works: what to read next.
Why is the output a 1-page diff rather than a summary?	A diff says what changed relative to what you already knew. A summary says what the paper contains. A diff forces you to identify novelty — the delta over prior work — rather than re-narrate the paper's own framing.
What three categories of information are consistently under-disclosed in limitations sections?	(1) Full compute cost including failed runs. (2) Data quality and curation details. (3) Evaluation coverage — the benchmarks where the method fails are usually absent from the main table.
What does an ablation row with a very small delta tell you about that component?	The component is a default choice — it barely matters for the primary metric. It was included for completeness or because it is standard practice, not because it was the key innovation.
What does a paper-graph encode that a paper-pile does not?	Dependency relationships between papers — which technique builds on which prior work. This lets you read in dependency order, identify foundational papers (high incoming edges), and locate new papers relative to what you already know.
Why should you read the three most-cited prior works in a paper's reference list?	These are the papers the new paper is building on. If you have not read them, you are missing the context that makes the new paper's contributions legible. They are the prerequisite nodes in the paper-graph.
What does an ablation entry of 'without component X: N/A' in a table tell you?	Component X was either (a) impossible to remove without breaking the system, (b) tried and failed in a way the authors chose not to report, or (c) the main baseline is effectively 'without X'. None of these are the authors' preferred narrative.
How does the paper reading protocol apply to DealLens deal memo analysis?	The same structured extraction — claim (investment thesis), number (valuation, round size), comparison (sector comps), limitations (risks) — produces a standardized deal summary. Three prior works becomes three comparable precedent transactions. The protocol is domain-agnostic.
What is the JHU humanoid build for Chapter 12?	Read GR00T N1 and GR00T N1.5 using the protocol in 60 minutes total. Write a 1-page diff naming exactly what changed (frozen VLM, simplified adapter, FLARE loss, DreamGen data) and why. Post to reports-lab.
After applying the protocol to 50 papers, what structural property of the paper-graph becomes actionable?	The nodes with the highest number of incoming edges — the papers most frequently depended on by others — are the foundational works in your area. These are the papers worth re-reading deeply rather than just diffing. The graph makes this priority list automatic.