Live site · Research report (plain-language, non-technical)
CERAI is a research prototype for anticipating civilian risk during evacuation operations in armed conflict. It combines multi-category risk indicators (hazards, infrastructure, population vulnerability, information environment, weather, and more) into a composite index with Monte Carlo sensitivity analysis, intended to support International Humanitarian Law (IHL) compliance review and humanitarian planning.
Live demo: https://ethical-tech-colab.github.io/CERAI_AR/
⚠️ Research prototype. Outputs are indicative and not a substitute for operational decision-making or legal advice.
- Interactive scoring across weighted risk categories
- Composite risk gauge with confidence bar
- Monte Carlo sensitivity analysis over uncertain inputs
- Location geocoding, weather integration, and ACLED conflict-event lookup
- Source credibility and data-freshness tagging for each indicator
The app is a single-file HTML prototype. Either:
- Open
index.htmldirectly in a browser, or - Serve the directory:
python3 -m http.server 8000then visithttp://localhost:8000/.
index.html at the repo root is served by GitHub Pages from main / root,
published at https://ethical-tech-colab.github.io/CERAI_AR/. Pushing to main
redeploys it.
Developed as part of PhD research at NYU School of Law on IHL and AI, with the aim of making evacuation-risk reasoning more transparent and auditable for humanitarian and legal audiences (UN, ICRC).
All rights reserved unless otherwise noted. Contact the author before reuse.
The full independent academic peer review of this report is in PEER-REVIEW.md (also available as Word under peer-review/).
Recommendation: Major revisions
Finding status (updated 2026-08-04):
| Finding | Status |
|---|---|
| Article 49 GC IV mischaracterised as danger-triggered duty (§1.3/§5.2/§8.1) | ✅ Fixed |
| Art 16/24 descriptors wrong; Art 78 misapplied to internal displacement (§8.1) | ✅ Fixed |
| Tool described but never evaluated: no weights, no validation table (§9.3/§10.4) | ✅ Fixed |
| No references section | ✅ Fixed |
| Authorship unclear ("advised by" vs. "developed by") | ✅ Fixed |
| Non-refoulement listed flat among IHL provisions (different legal regime) | ✅ Fixed |
| Monte Carlo varies inputs, not weights — limitation unstated | ✅ Fixed |
| Case-count vague/inconsistent ("few dozen" vs. actual N) | ✅ Fixed |
| Dual-use/misuse risk treated as one-line caveat | ✅ Fixed |
| 75% threshold: false precision, no derivation | |
| Literature positioning limited to INFORM Severity | 🔲 Open |
| External data licensing/terms unaddressed | 🔲 Open |
Noted strength: The endangerment/feasibility separation (§1.3, §8.3) is a genuinely valuable, ethically sound contribution that stops operational difficulty from silently overriding the duty to protect.
Legal citations corrected across index.html, CERAI.html, and CERAI-Paper.md:
- Art 49: 75% marker re-anchored in AP I Arts. 51, 57, 58 (IAC) and Common Art. 3 / AP II Art. 17 (NIAC). Art. 49 stated as prohibition on forcible transfer from occupied territory; evacuation a narrow conditional permission carrying a duty of return.
- Art 16/24/78: GCIV Art. 16 correctly identified as covering the wounded, sick, infirm, and expectant mothers. Children cited under GCIV Arts. 24 and 50 and AP I Art. 77. Older persons grounded in Customary IHL Rule 138. AP I Art. 78 limited to its actual scope: foreign-country evacuation only.
- Non-refoulement: Flagged as refugee law / IHRL (1951 Refugee Convention Art. 33), distinct from IHL, engaged where evacuees cross international borders.
Tool evaluated. CERAI-Paper.md §9.3 now publishes the full weight vector, aggregation formulas, and a face-validity table scoring four anchor cases (Mariupol 2022, East Aleppo 2016, Kabul 2021, Op Raahat Yemen 2015). Discrepancy ≤ ±2 points against retrospective ground-truth assessments.
Misuse section added (CERAI-Paper.md §10.8): outputs cannot supply legal justification; coercion/consent problem named; non-refoulement and return obligation specified; required output language at threshold stated.
References added: 20-citation references section covering primary legal instruments, customary IHL, methodology, and historical case sources.
Other: Authorship corrected ("Developed by Teresa Cantero"); case count stated as 21 with source attribution; Monte Carlo weight-sensitivity limitation stated explicitly (§10.9).