Thompson write-up: 'resolving a 109-year paradox' and similar headings outrun what has been externally checked (line-by-line list and replacements) #327

The Thompson fish write-up describes a good result in language that overstates what has been externally checked. The core numbers are fine (600-point input held fixed: curve TPS folds on 6/6 pairs, 8p Thompson on 0/6, better held-out interior-landmark prediction on 6/6; 791 species LOFO: Thompson 50.9% vs 55.3% data-driven ceiling). What has not happened: no one outside the team has agreed that the comparison criteria are fair (Wainwright set the test and has not signed off; Zelditch and Goswami/Felice have not been asked yet, see #326). Until that happens, "resolving a 109-year paradox" is a framing, not a finding.

Each line below is a place the framing runs ahead of the evidence, with a proposed replacement. None of these change a number.

docs/KEY_DISCOVERIES.md

• L27, L54: "Discovery I — Resolving the 109-Year Paradox in D'Arcy Thompson's 1917 Fish Transformations" → "Discovery I — Why 17-landmark fits and 600-point outline fits of Thompson's 1917 fish pairs disagree, and which one his grids pass". The "paradox" is our name for a discrepancy between two landmark protocols; nothing in the 1917–2024 literature calls it a paradox, and §1.9 (the historiographical audit) does not show anyone posing it as one.
• L75: "(or even an exact 34-parameter Thin-Plate Spline!)" → drop the exclamation marks throughout (also L204 "Box fails!", L207 "Fan Audit: morphology animations don't match real forms (planarian severe, Raup T ignored) #1!", L208 "Radial Audit: morphology animations don't match real forms (planarian severe, Raup T ignored) #1!"). A table cell is not a place to cheer.
• L232: "Resolving Harmon et al.'s (2010) 'Rare Early Burst' Paradox" → "A possible contribution to Harmon et al.'s (2010) rare-early-burst problem". This is one dataset and one decomposition; Harmon's question is field-wide and is not resolved by it.
• L339, L343: "Empirical Resolution of Turing Open Problems TOP-03 and TOP-07" → "Empirical verdicts on TOP-03 and TOP-07 under our stated reductions". A 1D scalar steady-state reduction (L340) decides the reduced problem, not the open problem.
• L347: "Discovery II — Genuine Mathematical & Biological Discoveries" → "Discovery II — Results on Rule 30 and Conus textile". "Genuine" is a word a reader supplies, not an author. §2.7 already has the right three-tier audit (proved / conditional / open); the heading should match it.

docs/THOMPSON_METHODS.md

• L1: title "Resolving D'Arcy Thompson's 1917 Fish Transformations Paradox" → "Methods & Provenance: the Thompson 1917 fish-transformation tests".
• L264: "How Querying the KG's pk:evidenceAgainst Triples Sparked the Breakthrough" → "…led to the 600-point outline study". A remediation item in an audit file is not a breakthrough.
• L41 (table): Wainwright's call is listed under "Human Principal Investigator & Domain-Expert Feedback" as a methods pillar. He gave two critiques, which were remediated; he has not reviewed the remediation. Add "(critique received 2026-10-21; remediated 2026-10-22; remediation not yet reviewed by him)".

docs/thompson_1917_fish_transformations_manuscript.tex

• L69: section title "A 109-Year Methodological Gap" is defensible (a gap, not a paradox); keep.
• L90 (Fig. 1 caption): "Resolving the 109-Year Discrepancy" → "The 17-landmark vs. 600-point outline discrepancy".
• L408, L428: "Resolving Harmon et al.'s (2010) 'Rare Early Burst' Paradox" → as above.
• L526, L535, L554: three subsection items headed "Why It Failed & How We Resolved It" about Huxley, Medawar and prior TPS tests. Huxley and Medawar did not fail; they lacked a 600-point outline dataset and a computer. → "What was missing and what we added".

docs/thompson_methods_and_provenance.tex

• L196 (banner): "How We Resolved D'Arcy Thompson's 1917 Fish Paradox" → "How the Thompson 1917 fish tests were run".

docs/thompson_1917_outreach_emails.md (Bookstein, Evans/Larouche, Sherratt, Joly drafts; Zelditch and Goswami are redrafted in #326)

• L83 "struck us deeply", L87 and L124 "a striking … discrepancy", L173 "deeply honored", L238 "Remarkably", L264 "a remarkable confirmation of your 2018 findings", L102/L173/L273 co-authorship invitations in a first email. Every one of these is doing the reader's evaluating for them. The ask in each email should be one checkable question (see the Sharpen Zelditch and Goswami/Felice outreach to one validation question each #326 drafts for the pattern) and the co-authorship offer should wait for a reply.

What I am not flagging

The numbers, the pre-registration, the blind FishShapes test (§1.12 reports its own null: "most of that skill is the nearest relative", L277) and the §2.7 three-tier audit are the honest parts of the document and should be the model for the rest.

Proposed rule for the repo, to go in AGENTS.md: a heading may say "resolved", "proved" or "discovery" only when the file names the outside check that earned the word (a referee, a dataset author's written agreement, a Lean module with no sorry, or a pre-registered test whose criteria were met). Otherwise say what was measured.