--- name: Refutation Banner Over A Confirming Panel slug: refutation-banner-over-a-confirming-panel category: Quality description: Refutation Banner Over A Confirming Panel checks report titles, headings, and figure labels for verdict words and internal labels that can misstate a result. Use it after a run or revision pass when a fidelity finding must stay in the right section and the rest of the report should stay neutral. github: "https://github.com/tangxiangru/AutoR/tree/main/src/skills/a-refutation-banner-over-a-confirming-panel" language: Python stars: 804 forks: 25 install: "npx degit https://github.com/tangxiangru/AutoR/tree/main/src/skills/a-refutation-banner-over-a-confirming-panel ~/.claude/skills/a-refutation-banner-over-a-confirming-panel" installs_to: ~/.claude/skills/a-refutation-banner-over-a-confirming-panel source_path: src/skills/a-refutation-banner-over-a-confirming-panel/SKILL.md collection_size: 25 category_size: 1354 collection_url: "https://dirskills.com/collections/tangxiangru/AutoR" added: 2026-08-22T05:22:22.318Z last_synced: 2026-08-22T05:22:22.318Z canonical_url: "https://dirskills.com/skills/refutation-banner-over-a-confirming-panel" --- # Refutation Banner Over A Confirming Panel Refutation Banner Over A Confirming Panel checks report titles, headings, and figure labels for verdict words and internal labels that can misstate a result. Use it after a run or revision pass when a fidelity finding must stay in the right section and the rest of the report should stay neutral. **Install:** ```bash npx degit https://github.com/tangxiangru/AutoR/tree/main/src/skills/a-refutation-banner-over-a-confirming-panel ~/.claude/skills/a-refutation-banner-over-a-confirming-panel ``` ## README # A refutation banner over a confirming panel ## What goes wrong You investigate the supplied bundle, find something real about it, and the finding is sharper and easier to measure than the analysis you were asked for. Two library skills already cover the consequence: run the requested analysis anyway and add the audit as its own section (`run-the-requested-analysis`), and kill the cheap explanations for a gap before calling it a disagreement (`close-the-gap-to-the-published-number`, whose first two discriminators are units/normalisation and population/split). Read both. This skill is the residue those two do not touch, and it is purely lexical: the word a hypothesis branch is given before any data exists, and the words that end up in the few slots a reader actually reads. Both are invisible from the inside. Both survive an intention to avoid them - a run can carry an explicit written objective to lead with the requested analysis, ship the fidelity verdict in its title anyway, and then spend its one revision pass making that verdict more precise. Intent does not fire here. Enumerate and search does. ## Check 1 - the branch-name test, at hypothesis freeze Before any hypothesis is frozen, write one line for it: > If the source's published claim is exactly true of my data, branch ____ fires, and that > branch is called ____. Read the branch name. If it contains *refuted*, *rejected*, *diverges*, *inconsistent*, *not reproduced*, *fails*, *synthetic*, *surrogate* or *inauthentic*, the hypothesis's subject is the file rather than the phenomenon, and every artifact downstream of it will stamp a failure word across a panel whose data agrees with the source. Prose cannot undo this: the heading and the figure banner are what get read. The fix is to change the subject, not the wording. The hypothesis asks about the thing the task named - does this variable dominate the response, does this ordering hold across conditions, does this arm detect more - phrased so that the source's claim being true fires a branch you would be happy to title a panel with. Provenance hypotheses are legitimate and stay. They are *separate*: their own hypothesis id, their own statistic, their own figure slot, their own section. The specific sentence to refuse is any variant of "computed once, this statistic is both the deliverable and a provenance test". A statistic with two jobs resolves as the audit every time, because the audit is the branch with the interesting name. ## Check 2 - the verdict-word sweep, after the last revision Do not do this by reading. Build the list, then search it. 1. Enumerate, as literal strings: the report title; the first two sentences of the abstract; every section and subsection heading; every figure suptitle and panel title; the first sentence of every caption. 2. Case-insensitively search each string for verdict words - refuted, supported, confirms, diverges, divergence, overshoots, undershoots, inflated, discrepancy, deviation, not reproduced, fails to reproduce, provenance, authentic, synthetic, surrogate, mislabelled - and for anything of the form "X vs published". 3. Search the same strings for internal labels: hypothesis ids, slot numbers, proposition or obligation ids, branch names, stage names. Every hit is a defect unless it sits inside the single section that *is* the fidelity accounting, or the word states the requested analysis's own outcome in plain language: an exemption on that ground has to name the phenomenon and a number, so "H1 refuted" and "the supplied numbers diverge from the published ones" do not qualify. Two fixes, and only these two: - **Internal label in a title.** Delete it. A reader has no copy of your numbering. The label belongs in the caption body or a plan appendix. - **Verdict word in a title.** Replace it with what the panel or section shows, in the source's vocabulary and units. A suptitle states the quantity, the comparison and the conditions. It is not a place to adjudicate anything. Then check the positive space, which is where the graded result usually went missing: for every claim of the source's that your data speaks to, one plain sentence exists saying whether your data reproduces it, with your number and the source's - **including every claim that does reproduce**. A confirmation that appears only as the premise of a refutation ("the effect sits where the source said it does, so the divergence hypothesis fails") is not in the report as a result. Give it its own sentence, in the lead of the section that owns it. ## Checklist - [ ] Every frozen hypothesis has its branch-name line written, and no confirming branch carries a failure word. - [ ] No statistic and no figure slot serves both a requested analysis and a provenance test. - [ ] Title, abstract's first two sentences, all headings, all suptitles and panel titles, all caption openings: enumerated as a list and searched, not skimmed. - [ ] Zero verdict words and zero internal labels among them, outside the one fidelity section. - [ ] One sentence per source claim stating whether your data reproduces it, with both numbers, including the ones that agree. - [ ] The fidelity finding is complete, in one place, and is not the first thing a reader meets.