@researai/deepscientist 1.5.16 → 1.6.0

package/src/skills/write/references/section_rewrite_checklist.md

@@ -0,0 +1,234 @@
+# Section Rewrite Checklist
+
+Use this checklist when upgrading an existing draft into an oral-quality paper.
+
+## 1. Abstract
+
+Check that the abstract does all of the following in order:
+
+1. problem or deficiency
+2. proposed idea
+3. why the idea is different or principled
+4. the main empirical pattern
+5. one or two strongest concrete outcomes
+6. calibrated takeaway
+
+Rewrite if the abstract:
+
+- dumps too many numbers without hierarchy
+- sounds like a compressed lab note
+- lacks a stable one-sentence memory of the paper
+- tries to answer every reviewer question before the paper has started
+- uses strict-dominance language that the evidence package does not fully support
+
+## 2. Introduction
+
+Check that the introduction:
+
+- states the problem clearly
+- explains why current approaches are insufficient
+- introduces the paper's core idea in plain language
+- gives the reader an intuition anchor early
+- clearly identifies the default setting or operating regime the method is meant to improve
+- places the paper in the broader method family before narrowing to the exact gap
+- makes novelty legible
+- ends with clear contributions or roadmap
+
+Rewrite if the introduction:
+
+- jumps to results too quickly
+- delays positioning until late sections
+- assumes the reader already agrees the method matters
+- has no early figure or verbal mechanism anchor for the main idea
+
+## 3. Related Work
+
+Check that related work:
+
+- appears early enough to establish context
+- distinguishes the paper from the closest competitors
+- separates "similar goal" from "similar mechanism"
+- identifies the nearest neighbors, not just the broad method family
+- clarifies what is genuinely new versus better framed
+
+Rewrite if related work:
+
+- is only a bibliography dump
+- comes too late to help novelty perception
+- never defines the paper's place in the literature
+- leaves the nearest-neighbor novelty boundary ambiguous
+
+## 4. Method
+
+Check that the method section:
+
+- defines the method precisely
+- gives intuition, not only equations
+- explains why the method should work
+- compares the method conceptually to alternatives
+- tells the reader what property to watch for in experiments
+- keeps any implementation-facing statement anchored to actual code, configs, scripts, or verified outputs rather than memory, comments, or stale prose
+- does not consume space that would be more valuable as objection-handling evidence
+
+Rewrite if the method:
+
+- only presents formulas
+- uses notation without interpretive guidance
+- leaves the main conceptual difference implicit
+- never tells the reader how to read the key figure or later results
+- repeats defenses that could be compressed to make room for stronger main-text evidence
+
+## 5. Experiments
+
+Check that the experiments section:
+
+- introduces the evaluation setup briefly but clearly
+- states the main empirical pattern before drowning in numbers
+- breaks the empirical case into explicit reviewer-question subsections when multiple claim clusters are being defended
+- uses displays that match the central claim
+- separates main results from supporting details
+- anticipates obvious objections
+- gives each major display one dominant local claim
+- lets the table or plot carry most of the raw detail
+- gives practical-value or objection-handling evidence main-text space when central to the claim
+
+If the central claim is comparative, also check that:
+
+- one main-text result block still shows named competitors or baseline families
+- the visible metrics are sufficient to justify the comparative wording
+- the comparison surface has not been collapsed into self-only rows plus prose summary
+
+Rewrite if the experiments:
+
+- read like table narration
+- rely too heavily on averages without decomposition
+- bury the key message inside long numerical paragraphs
+- compress setup, headline win, robustness, and ablation into one continuous flow with no clear subsection jobs
+- push interpretive or objection-handling evidence into tiny leftover sentences
+- preserve broad comparative claims after removing the benchmark block that would let a reviewer verify them
+
+## 6. Analysis
+
+Check that analysis:
+
+- explains the mechanism behind the empirical pattern
+- addresses likely reviewer concerns
+- includes at least one targeted interpretive angle
+- connects back to the central claim
+- contains at least one memorable piece of evidence when the paper can support it
+- uses the prose to explain why the pattern appears, not to restate visible numbers
+- names the cross-setting trend and the underlying tradeoff or essence of the result
+
+Possible useful analysis forms:
+
+- robustness analysis
+- diversity or efficiency analysis
+- tuned-baseline analysis
+- qualitative examples
+- failure cases
+- case study
+
+Rewrite if analysis:
+
+- feels like leftover supplementary material
+- repeats results without explaining them
+- mostly paraphrases what the reader can already see in the figure or table
+- cites values correctly but never states what stable pattern the reader should infer
+- never changes the reader's understanding of why the method works
+- could be deleted without changing how memorable the paper feels
+
+## 7. Figures and Tables
+
+For each important figure or table, ask:
+
+- Why is this display here?
+- What exact question does it answer?
+- Is the caption enough to read it independently?
+- Does the surrounding prose tell the reader what takeaway matters?
+- Does the surrounding prose interpret rather than duplicate the display?
+- Should this live in the main text or appendix?
+- Is it the right display type for this job: intuition, result, tradeoff, or case study?
+
+Move material to appendix if it is:
+
+- complete but not central
+- exhaustive rather than explanatory
+- useful only for deep verification
+- too dense to remain legible in the main text
+
+Make sure the main text has a balanced display program:
+
+- one early intuition or mechanism figure
+- one main result display
+- one analysis or tradeoff display
+- one memorable qualitative or case-study display when available
+
+If the paper is being written section by section, lock this display program in the plan first and then preserve those roles through integration. Do not let later sections silently demote a planned main-text display into a passing prose mention.
+
+## 8. Conclusion
+
+Check that the conclusion:
+
+- restates the main contribution cleanly
+- summarizes the evidence at the right level
+- makes the implication clear
+- avoids overclaiming
+- stays inside the strongest evidence zone
+
+Rewrite if the conclusion:
+
+- merely repeats the abstract
+- adds new unsupported claims
+- lacks limitation or scope awareness when needed
+
+## 9. Limitations and Reproducibility
+
+Check that the paper:
+
+- acknowledges real limitations
+- signals what is reproducible now
+- points to appendix or code artifacts where appropriate
+- describes artifact availability consistently across abstract, main text, appendix, and reproducibility statements
+- does not present planned, commented, or stale code paths as if they were verified current artifacts
+- includes prescriptive future-facing guidance for follow-up quests on the same problem: which routes are now closed by falsification (and the regime in which they are closed), which knobs were not exhausted, which cross-distribution or scaling axes remain unprobed, and which alternative families would be the highest-value next pass. State each as a directed claim ("a future pass should…", "do not retry X without first…"), not a passive enumeration of open questions, so a sibling quest reading this section during cross-quest recall can lift it directly into idea-stage selection.
+
+Do not use limitation language performatively. The limitation should calibrate trust, not just check a box.
+
+## 10. Appendix Package
+
+Check that the appendix has explicit jobs such as:
+
+- protocol and evaluation detail
+- full result tables
+- human evaluation protocol details
+- hyperparameter sweeps or ablations
+- tuned baselines or sensitivity checks
+- extra examples or failure cases
+- proofs or formal derivations
+- reproducibility detail
+
+If the main text was compressed for pacing, check that the appendix still visibly carries the overflow jobs needed to defend the main claim.
+
+Rewrite if the appendix:
+
+- looks like unstructured leftovers
+- contains critical evidence that the main text never signals
+- fails to answer the reviewer questions most likely to arise
+- has no visible progression from protocol to results to stress tests to qualitative evidence
+- behaves like a short method tailnote instead of a reviewer-support package
+
+## 11. Final oral pass
+
+Before considering the rewrite complete, ask:
+
+- Can someone summarize the paper after one read?
+- Is the central idea visually and verbally anchored early?
+- Does each main-text page or section have a clear dominant job?
+- Are the strongest figures and analyses in the main text?
+- Does the evidence budget look intentional rather than frugal or overloaded?
+- Is there breathing room on major result pages?
+- Does the paper feel guided rather than compressed?
+- Does it read like a human-curated argument rather than an LLM summary?
+- Has the final integration pass removed planning language, self-narrating rhetoric, and other drafting scaffolding?
+
+If not, the draft is not oral-ready yet.