You already know: confidence intervals, significance and power (D2); evaluation design, contamination and judges (D10); every system component a paper might claim to improve (D5–D19).
Why this matters for Atlas. Research engineers read far more papers than they implement, so the skill is triage: find the claim, judge the evidence, and decide quickly whether a result is real, large enough, and likely to transfer. Most papers are honest but optimistic: the effect is measured on one benchmark, against an under-tuned baseline, with one seed, at a scale or compute budget that doesn't match yours. This lesson gives a reading procedure, a checklist of the ways evidence goes wrong, and small tools for checking claims numerically.
1. The three-pass method
The question. How do you read a paper in an hour instead of a day?
Start from nothing. Read in passes, deciding after each whether to continue (Keshav):