clear-eyed-deep-reading

Clear-Eyed Reading

Goal: a reader with no background finishes knowing what the work actually does, where its ability comes from, what is really new, and why its result came out the way it did — judged independently of the authors' names, branding, and framing. This can raise or lower the reader's opinion of the work; an undersold structural contribution deserves as much attention as an inflated claim. Critique claims and evidence, never the authors.

Four questions to answer

These drive the research. The writing section below says how to present the answers.

1. What does it actually do? Strip the names: what goes in, what happens at each step, what comes out.
2. Where does the ability come from? Which parts come from this work, and which from data, pretrained models, earlier methods, rules, labels, human choices, instruments, scale, tuning, post-processing, or the evaluation setup.
3. What is really new? For every contribution the authors claim: what is borrowed, what changed compared with the closest earlier work, and what the field would lose without the change. Keep three things apart: how new the technique is, how much it matters, and how far the authors' claim exceeds or undersells the verified change.
4. Why did the result come out this way? What actually carries the result or conclusion, and which other explanation — data, scale, baseline choice, metric, leakage, selection, tuning budget — could also produce it.

How to research

• Identify the exact work and version (preprint version, venue version). Read the primary text; never fill gaps from an abstract, caption, search snippet, or guess. Say plainly what you could not check. Do not bypass CAPTCHAs, logins, or other access controls; report the source as unchecked.
• List every contribution the authors claim — title, abstract, contribution list, named modules, method, figures, experiments, conclusion. Treat each as a claim to check, not as the outline of your explanation.
• For each claim, find the closest earlier work by more than one route: search, the work's references, later work citing it, older names for the same idea, neighboring fields. Compare against the earlier work's primary text, not its abstract. If the primary text is inaccessible, you may rely on a reliable secondary description such as a survey, but say so where the comparison is used. For supporting context that does not decide a judgment, an abstract is acceptable if you say so. Citation counts, downloads, and attention help find candidates; they never prove novelty or significance.
• Before accepting the authors' explanation of the result, ask what else could produce it.
• Look at decisive figures and tables yourself. Keep apart what you observed, what the authors say, external facts, your inference, and unknowns. Visual appeal is not evidence.
• Spend effort in proportion to risk: claims about novelty, significance, causation, reuse readiness, or the absence of a weakness need the strongest evidence. The depth profile sets the overall budget. Stop when more searching is unlikely to change the four answers, not at a fixed paper count.
• If independent subagents are available and allowed, you may split the checks — earlier work on the problem, earlier work on each component, other explanations of the result, material around the work — and reconcile afterwards. Otherwise do the same checks yourself in sequence.
• If outside research is impossible, answer from the work alone and say which judgments (novelty, field position, other explanations, material around the work) are unchecked. Do not score novelty or significance in that case.