# Claims Every substantive claim this project makes, what would falsify it, and how long checking takes. This file exists because a dashboard invites belief and a claims list invites attack. If Closure is worth anything, it is worth someone trying to break — and that requires the claims be stated in a form that can lose. **Nothing here has been checked by a domain expert.** Every figure was produced by measurement against held-out data, controls, or independent labellers. None of it has been checked by a person who knows camera-trap density estimation. That is the single largest caveat on the page and it is not closable by more engineering. --- ## A. Claims about the field **A1. Of works in this corpus that use a named density estimator, six method families have no automated implementation: REM, REST, TTE, SCR, spatial-count, N-mixture. Only CT-DS has any (5).** - *Falsified by:* one paper implementing automated detection or distance measurement for any of those six. A reader who knows the field would name one or fail to in under a minute. - *Check:* `artifacts/field_grid.json`, `automated` per row. - *Weakest link:* `automated` is set from title text only. A paper that automates without saying so in its title is invisible to this claim. **This claim is probably too strong and is the one I would attack first.** **A2. The corpus holds 1,003 works; 369 use a named estimator; SCR is the largest family at 160.** - *Falsified by:* naming works in the field absent from `artifacts/census.json`. This will succeed — see B1. - *Check:* `artifacts/census.json`, `artifacts/field_grid.json`. **A3. The seed manuscript cites 14 of 298 field works using a named estimator (5%), including 1 of 159 SCR works.** - *Falsified by:* the seed's reference list disagreeing. Directly checkable against the paper. - *Not a criticism.* A distance-sampling methods paper has no obligation to cite the SCR literature. The number exists so that "you have not engaged with SCR" becomes checkable rather than an impression. ## B. Claims about coverage **B1. The corpus is 59% complete against an independent probe. It is NOT a census.** - *Method:* the corpus was built by definition queries plus citation closure; the probe was built by citation traversal from one paper. Of 140 probe works classified as this field, 83 are in the corpus. - *Falsified by:* a better-constructed probe returning materially different recall. - *Known bias:* both mechanisms are DOI-bearing and English-only. **Neither can see non-indexed or non-English work at all, so 59% is an overestimate of true coverage.** **B2. Across 10 seed papers and 140 held-out real citations, 79% of works the authors actually cited were recoverable in the pool.** - *Check:* `artifacts/holdout_eval.json`. - *Weakest link:* "in the pool" can mean position 5,000 of 13,000. Pool membership is necessary, not sufficient. The usable figure is recall@10 = 24%. ## C. Claims about the instrument **C1. Classifier reliability is population-dependent: Cohen's κ 0.95 on a CT-DS-heavy sample, 0.71 on an SCR-heavy sample of the same field.** - *Falsified by:* a third sample showing no such dependence. - *If it holds,* it means a single accuracy number for a classifier over a heterogeneous literature is not a property of the classifier. This is the most transferable finding here. **C2. Ordering the pool by citation frequency or by semantic relevance carries no demonstrated signal.** Frequency r = +0.10 against expert relevance; semantic reranking p = 0.66 against real author citations, with gains on pre-2022 papers and losses on recent ones. - *Falsified by:* a ranking that beats graph order on held-out real citations at p < 0.05. - *Note:* an earlier r = +0.83 result was withdrawn — model graders judging a model reranker. **C3. Several per-class precision figures rest on n < 6 and should not be read as measurements.** - SCR n=3, spatial-count n=2, index-RAI n=3 in the roster measurement. ## D. What is NOT claimed - That it reads papers. There is no PDF parsing or full text anywhere in the pipeline. - That it covers gray literature. Zero gray-literature documents have been ingested. - That the ordering of any list means anything. - That any of this has changed a decision, a manuscript, or an outcome. **It has no users.** --- ## The check that would matter most Someone who knows this field reads A1 and either names a counterexample or does not. Ninety seconds. If they name one, the automation claim is wrong and the `automated` flag needs full text rather than titles. If they cannot, then a first-in-family claim for six method families is a real observation about a real literature — and that is a finding, not a dashboard. Everything else here is instrumentation for producing more claims of that shape.