# Assertive Scientific Writing Internal verification can be strict without making the paper timid. Write the reader-facing manuscript as a confident scholarly argument. ## Direct by Default - State observed results directly: “CARBON improved Pearson correlation across the evaluated tasks,” not “The supplied evidence appears to suggest that CARBON may potentially improve…”. - Use active, specific verbs: improves, reduces, identifies, captures, supports, outperforms, reveals. Choose the verb whose strength matches the evidence. - State the contribution and why it matters without self-deprecating workflow language. - Write the user's study as a research paper, with its scientific question, methods, findings and significance. Do not make the assistant's material inventory, file checks or lack of a new replication the subject of the paper. Preserve scientifically important design and measurement information; put engineering execution details in task notes or the local work package, not in the article's scientific supplementary material. Research reproducibility information and required disclosures remain part of the scholarly record. ## Hedge Only Where the Inference Changes Use qualification for causal claims from observational evidence, extrapolation beyond evaluated data, uncertain provenance, incomplete statistics, or genuine alternative explanations. Put one clear boundary at the relevant paragraph, section, caption, or limitations passage; do not repeat it sentence by sentence. Avoid stacked hedges such as “may potentially suggest,” repetitive “within the supplied materials,” and audit phrases such as “user-authoritative evidence.” The evidence ledger holds provenance; the manuscript communicates the science. Apply [publication-surface.md](publication-surface.md) to mixed scientific and production-status paragraphs. Figure choices govern which image is used; they are not a finding to announce. Prefer one consequential limitation over repeated workflow disclaimers, without removing uncertainties that change interpretation. ## Limitations Limitations should identify the few constraints that change interpretation or generalizability. Do not add ritual disclaimers for every ordinary design choice. A concise, specific limitation is stronger than pervasive defensive prose. ## Organize the argument around the earned result Use the closest actually read work to state what the study adds under the tested conditions. After analysis or a substantive revision, check the title, abstract, introduction's contribution and conclusion against the same result. A useful negative or uncertain result may change the question the paper can answer; it need not be hidden or converted into a positive claim. Distinguish improvement of a complete method from evidence that one component caused it. A component-removal comparison whose uncertainty includes no change does not establish that component's independent benefit, and does not prove it has no effect. Consider other changed controls, evaluation windows and resource budgets before attributing the result. Resolve this through a narrower claim or an authorized targeted experiment, not stronger adjectives or significance seeking. For Results, connect the decisive quantitative comparison and its uncertainty to the scientific question. For Discussion, explain the contrast with prior work, plausible alternatives and the remaining discriminating question. Group repeated limitations where they change interpretation; keep locally necessary qualifications and required disclosures. Reader-facing labels such as "supplied" or "confirmation" are appropriate only when they identify a meaningful study design, not merely how the assistant received or checked a file. This is contextual editing, not a word blacklist or automatic deletion rule. ## Apply the saved author-expression preference Read `configuration.author_voice_restoration` from the same task. `off` uses ordinary scientific editing. `standard` preserves the author's characteristic terminology, emphasis and useful phrasing while improving concrete weaknesses. `strict` makes the minimum necessary changes to the author's expression and argument structure; explain any substantial restructuring in the revision notes. Neither option permits changing results, claim strength or scientific meaning. Use the supplied draft and the author's own or explicitly authorized writing as the voice reference. Do not treat a target-journal exemplar as that author's personal voice. If no author text exists, retain the user's explicit expression preferences and state that a personal voice could not be learned; continue useful writing without inventing one or requiring a new writing corpus. Preserve the prior draft. For each affected passage, identify its scientific function, retain the evidence and terminology, and fix only the actual reader problem. Clear text may remain unchanged. Compare the revised passage with the original for numbers, negation, qualifications, attribution and conclusion strength. Give the independent reviewer both versions and the actual saved preference. Do not use a detector score, word blacklist or sentence-rhythm quota as authorship evidence. Historical author-voice receipts, hash-bound confirmation and fixed approval sequences are not prerequisites for this current host method.