Reviewer Psychology

How a reviewer actually reads a paper, and where they stop to doubt. Use this to prioritize audit findings: an issue that sits where reviewers reliably get suspicious matters more than one they would never reach. Reading path and the suspicion ranking below are general across venues (CS / ML / quantitative science); they are about reviewer behavior, not any single field.

How reviewers read (not front-to-back)

A reviewer handling 3-5 papers in a cycle, ~3-6 hours each, does not read linearly:

Step 1 (3-5 min):   Title -> Abstract -> Figure 1 -> Conclusion
                    => "is this worth reading deeply?"  (>50% of accept/reject leaning forms here)
Step 2 (10-20 min): skim all figures + captions -> skim key Methods parameters
                    => "does the data support the claims / is the method credible?"
Step 3 (1-3 h):     Results + Discussion in detail
                    => hunt for over-claims, logic gaps, missing experiments
Step 4 (~30 min):   Introduction close read + spot-check appendix -> write the review

Optimizing for Step 1 has the highest payoff. The audit should weight problems that a reviewer hits early (title/abstract/figure-1/conclusion) above problems buried deep.

The 3-second scan

In the first seconds a reviewer registers the title (right keywords? claim overreaching?), the author list, and the venue. Within 30 seconds: the first sentence of the abstract (empty throat-clearing or a concrete gap?), the last sentence of the abstract (does the claim overreach?), and Figure 1 (understandable at a glance?).

Where reviewers stop to doubt — ranked by how often it happens

Order findings by this list. Higher = a reviewer is more likely to catch it and let it sink the paper, so it should rank higher in the audit roadmap (within the same severity).

1. Numbers that do not match the claim. "significantly higher" with no number; the abstract says 184, the body says 178; "2× faster" but the figure shows less. → Auditor: cross-check every number across abstract / body / figures / tables; flag every magnitude word with no number. (lanes: notation_and_numeric_consistency, claims_vs_evidence)
2. Missing method parameters. "We ran PCA / fine-tuned" with no version, no hyperparameters, no threshold, no multiple-testing correction. → Auditor: flag undisclosed parameters that block reproduction. (lane: evaluation_fairness_and_reproducibility)
3. Weak citation support. A specific claim backed only by reviews; a citation whose actual conclusion is the opposite; a whole paragraph with no citation. → Auditor: flag claims without a primary-source anchor. (lane: prior_art_and_novelty_grounding)
4. Over-claim. "first to" (when it is not), causal language on correlational evidence, "universally" from one setting, undemonstrated downstream applications. → Auditor: see OVER_CLAIM_GUARD.md; flag wording that outruns the evidence. (lane: claims_vs_evidence)
5. Story does not close. Intro promises A, B, C; Results only deliver A and B. Abstract stresses X; the body barely covers X. → Auditor: check contribution-to-result closure end to end. (lane: claims_vs_evidence, section lanes)
6. Figures disconnected from text. Text cites "Figure 2A" that does not exist; missing axis units / legends / scale bars; color choices hostile to color-blind readers. → Auditor: cross-check figure references and self-containment. (lanes: section lanes, notation_and_numeric_consistency)
7. Language that blocks understanding. Long multiply-negated sentences; tense misuse (Methods/Results not in past tense) that blurs method vs. result vs. inference; undefined abbreviations on first use. → Auditor: flag clarity blockers; tense is covered by the writing skills' de-AI tense check and should be noted when it impedes reading. (lane: section lanes)
8. Results "too clean." Every p-value < 0.001; every effect in the predicted direction; sample size conveniently sufficient in every comparison. → Auditor: note absence of negative / marginal / inconsistent results and limitations. (lane: evaluation_fairness_and_reproducibility)

Reviewer types on a panel

A paper usually gets 2-3 reviewers of these kinds — write (and audit) for all three:

• Domain expert — cares whether your data conflict with published work and whether you cite the core literature. Audit: literature grounding, honest positioning.
• Method expert — cares whether the method choice is justified and parameters are sound. Audit: methods detail, fairness of comparison, ablations.
• Generalist / editor — cares about "so what" and whether a non-specialist can follow the abstract and introduction. Audit: accessibility of abstract/intro, significance framing.

Making the paper "reject-proof"

Reviewers often decide first and then look for evidence to justify it. The defensive moves an audit should check are present:

1. Title + abstract carry no over-claim smell (OVER_CLAIM_GUARD.md).
2. Figure 1 is understandable in seconds.
3. Key parameters are fully disclosed (kills "is this reproducible?").
4. Limitations are stated by the authors (kills "they ignored the obvious weakness").
5. Data + code + supplementary are available (kills "can this be trusted?").

When these are present, the only thing left to attack is the core claim itself — which is exactly the high-level discussion a strong paper wants.

How this skill uses it

• Critical reviewer orders its challenges by the suspicion ranking above (highest first).
• Synthesis breaks ties within a severity tier by suspicion rank, so the revision roadmap surfaces what a real reviewer hits first.
• This file changes ordering and emphasis, not severity definitions — severities still follow REVIEW_CRITERIA.md / editorial_decision_standards.md.