Known gaps, in the order they would mislead a reader.
A hostile-but-fair peer-review sweep of the report copy (canonical in
CERAI-Paper.md, mirrored on the site in
website/src/content/publications/cerai.ts). Fix in both places. See also
PEER-REVIEW.md.
Three checkable international-law errors an IHL reader would catch, all touching the 75% "Article 49 obligation" threshold that the model is built around:
- Article 49 is a prohibition, not a mandate. GC IV Art 49 prohibits forcible transfer and only permits evacuation as a narrow exception in occupied territory — it does not require it, and it does not apply to armed conflict generally. Reword to "Art 49 permits evacuation of an occupied area as an exception to its prohibition on forcible transfer," and stop describing 75% as an "obligation."
- Art 16 / 24 descriptors are wrong. §08 labels GC IV "Articles 16 and 24" as protecting "children and the elderly," but Art 16 covers the wounded, sick, infirm, and expectant mothers, and no GC IV article specifically protects "the elderly" — the exact confusion the Evacuation Simulator report corrects for its sister project. Cite Art 24 (children) and Art 16 (wounded/sick/infirm/expectant mothers) accurately; ground any "elderly" claim in Customary Rule 138.
- AP I Art 78 is misapplied. It is cited for internal child evacuation, but Art 78 governs only evacuation of children to a foreign country under strict consent conditions. Rely on AP I Art 77 and GC IV Arts 24/50 for internal child protection.
Why it matters. The legal citations are the report's authority; getting them wrong hands a specialist reader an easy discredit.
§01/§05/§10 describe "a database of documented real-world evacuation operations since the year 2000" of "a few dozen documented cases" used to check outputs for face validity, with no source in the works cited. Cite the underlying case sources, or explicitly label the database as the researcher's own unpublished compilation.
Full referee report in PEER-REVIEW.md. Verify each against the current report copy and check off any already fixed.
Major:
- Article 49 mischaracterised (§1.3, §5.2, §8.1). "Requires evacuation" inverts the law: Art 49 is occupation-specific, permissive-exceptional, and paired with a duty of return the tool never models. Re-anchor the endangerment rationale in the precautions rules (AP I Arts 57–58, customary Rules 15–24) and protection-from-attack (AP I Art 51); recast Art 49 as a constraint on evacuation. (Same as credibility item 1 above; this is the single highest-value fix.)
- The tool is described but never evaluated (§5.5, §9.3–9.4, §10.4). No weight vector, no validation table, no worked example anywhere. Publish the weights, aggregation formulas, and a face-validity table scoring the named cases (Aleppo, Mariupol, Mosul, Kabul) against independent characterisations, plus one end-to-end scenario.
- The 75% threshold is false precision with no derivation (§5.2, §5.4). A single cutpoint on a qualitative treaty phrase, with worst-case defaults pushing under-informed scores toward it. Derive it empirically or replace with a banded "consult legal review" zone; disclose the default interaction; suppress time-to-threshold when confidence is low.
- Dual-use / misuse risk treated as a one-line caveat (§10.6, §12.1). A high score tied to a treaty article can be invoked to legitimise unlawful displacement. Add a misuse section: outputs cannot supply legal justification; foreground coercion/consent and non-refoulement; name any guardrails.
Minor: no references section (m1); literature positioning limited to INFORM (m2); aggregation math prose-only and "no double-counting" unproven (m3); Monte Carlo varies inputs not weights (m4); authorship unclear (m5); external data licensing/terms unaddressed (m6); non-refoulement listed flat among IHL provisions (m7); case-count vague/inconsistent (m8).