Climate stressors precede armed-conflict events by 6-18 months

Five independent claude-sonnet-4-6 calls, one per persona — Skeptic, Fact-Checker, Researcher, Compliance-Guard, Falsification-Auditor. Each writes its hardest objection from its own seat, paired with the methodological resolution it would accept. Run on the static catalogue spec Jun 3, 2026; a live council for any topic is at /try.

1. Skeptic #01
   Raised

   The observed 6-18 month cross-correlation between drought indices and conflict-news signal volume could be generated entirely by global food commodity price transmission rather than a direct climate-to-conflict causal pathway. When regional drought reduces Sahelian or MENA grain harvests, it contributes to spikes in the FAO Food Price Index and World Bank commodity price series, which then drive conflict onset through urban purchasing-power shocks—a mechanism that would produce the same apparent temporal precedence in the data without any local climate causation. Compounding this, using discourse/Reddit news volume as the conflict proxy (rather than verified ACLED or UCDP-GED event records) introduces a media-attention confounder: international outlets systematically cover drought conditions months before they report on resulting violence, artificially inflating cross-correlation at exactly the 6-18 month lag range independent of any real causal structure.

   Resolved

   Incorporate the FAO Food Price Index monthly series and World Bank pink-sheet commodity prices (wheat, sorghum, maize) as covariates—both retrievable via /api/worldbank—inside a panel Vector Autoregression estimated per region, and replace the news-volume conflict proxy with georeferenced ACLED event counts or UCDP-GED onset dates as the dependent variable. Compute partial cross-correlations between the detrended drought index and verified conflict onsets after conditioning on the commodity price vector; the decision rule is that the hypothesis is supported only if the residual partial cross-correlation coefficient at 6-18 months remains statistically significant (|r| exceeding the Bartlett band, p < 0.05) in ≥3 geographic clusters, and the same lag structure replicates when ACLED counts replace discourse volume—falsifying the media-attention artifact as the operative mechanism.

2. Fact-Checker #02
   Raised

   The primary uncertainty stems from two compounding sources. SPI/SPEI drought-onset detection carries a temporal ambiguity of approximately ±1.5–2 months (1-sigma) depending on accumulation-window choice (SPI-3 vs. SPI-6), and NDVI-based crop-failure proxies embed a phenological lag of 1–3 months relative to actual canopy stress onset. The named "ACLED-equivalent" news/discourse signals from /api/discourse and /api/reddit introduce an additional 4–8 week systematic reporting delay and a recall rate of roughly 60–80% for sub-national events in the Sahel and Horn of Africa, producing a combined 1-sigma temporal uncertainty budget of approximately ±2–3 months. This means the 6-month lower bound of the SUPPORTS threshold is only ~2 sigma above zero lag, making it nearly indistinguishable from measurement artifact — especially given that both drought and conflict peak seasonally in these regions, which can inject a spurious 12-month cross-correlation lobe.

   Resolved

   Replace /api/discourse and /api/reddit conflict proxies with version-stamped ACLED v3 data filtered to quality flag = "confirmed," and define drought onset using CHIRPS v2.0 precipitation anomalies combined with MODIS MOD13A3 NDVI restricted to pixel quality flag ≤ 1 ("good"/"marginal"). Apply STL seasonal decomposition with a 12-month periodic window to both series before cross-correlating, eliminating the annual co-cycle that otherwise generates spurious lags. Construct 95% block-bootstrap confidence intervals (≥1,000 resamples, block length = 3 months to preserve autocorrelation) on the peak-lag estimate per region; the SUPPORTS criterion is met only if all four CIs exclude zero lag and fall entirely within 6–18 months in ≥3 of 4 regions, and the hypothesis is FALSIFIED if the CI overlaps zero or the bulk of CI mass falls outside 6–18 months in ≥2 regions.

3. Researcher #03
   Raised

   The dominant uncontrolled confounder is global food commodity price volatility, which operates through the following channel: drought events in major grain-exporting regions (e.g., Russia/Ukraine 2010, U.S. Midwest 2012) drive spikes in global cereal prices within 2–4 months, after which food-import-dependent states across MENA, the Sahel, and Central Asia experience downstream fiscal stress, subsidy failures, and food-riot mobilization approximately 4–14 months later—producing an apparent 6–18 month lag between locally-measured drought indices and locally-measured conflict signals that is in fact mediated by global commodity market transmission rather than local agro-climatic scarcity. Because the cross-correlation design does not condition on global food price dynamics, the estimated lag statistic conflates the intended sub-national climate-to-conflict pathway with the global price-shock-to-unrest pathway, and this inflation is particularly severe for the MENA cluster, where domestic food production is structurally decoupled from conflict risk.

   Resolved

   The FAO Food Price Index (FPFI) monthly series, available from FAO GIEWS and updated in near-real-time, should be included as a global time-varying covariate in each regional cross-correlation model; country-level food import dependency ratios from FAOSTAT Food Balance Sheets (Import Dependency Ratio series, accessible via World Bank WDI indicator AG.IMP.TOTL.Q) should be interacted with the drought signal to allow price-transmission sensitivity to vary by food self-sufficiency. A structurally stronger fix is to impose country × calendar-year fixed effects in the panel regression underlying the cross-correlation, which absorbs all contemporaneous global and national macroeconomic shocks—including commodity price cycles—leaving only within-country, within-year sub-national drought variation to identify the lag, and thereby cleanly separating the local climate pathway from the global market pathway.

4. Compliance-Guard #04
   Raised

   The PREDICTS clause explicitly positions this lag-correlation finding as a "preventive-finance trigger" and a driver of stabilization-aid allocation — channels governed by IFRS S2 (climate-related risk disclosure materiality), the EU CSRD/ESRS E1 double-materiality standard, and World Bank Operational Policy fiduciary frameworks for climate-linked development lending. If an investment manager, sovereign wealth fund, or multilateral development bank cites a sub-threshold correlation as validated causal precedence to justify a reallocation of capital or humanitarian pre-positioning, that reliance could constitute a material misstatement under SEC Rule 10b-5 or an SFDR Article 9 mis-classification, and could trigger improper pre-commitment of CERF or EU Humanitarian Aid anticipatory-action reserves against an evidence standard those protocols explicitly require to be independently verified.

   Resolved

   The finding must not be cited as actionable until (a) the cross-correlation achieves statistical significance at the pre-specified lag in at least three geographically distinct clusters and the result survives out-of-sample holdout validation against the official ACLED database as independent ground truth — not solely discourse proxies — and (b) the methodology is peer-reviewed and formally integrated into a recognized early-warning standard such as FEWS NET or the UN OCHA Anticipatory Action trigger framework; until both gates are cleared, all outputs must carry an explicit disclaimer stating that the observed lag-correlation is a hypothesis-stage finding, has not established causality, and must not serve as a standalone trigger for financial allocation, insurance parametric payout, or policy commitment under any climate-risk disclosure or humanitarian-finance regime.

5. Falsification-Auditor #05
   Raised

   The FALSIFIES condition "No coherent temporal precedence" is not operationalized with a numerical boundary, making it practically unreachable. Both drought indices and conflict-news volume are strongly autocorrelated time series; naive cross-correlation significance tests treat each monthly observation as independent, but autocorrelation corrections typically reduce effective sample size by 60–80%, so any |r| > 0.15–0.20 at arbitrary lag will exceed nominal α = 0.05 thresholds even under the null. The 6–18 month search window spans roughly 30% of a typical 48-month lag range, and across four geographic clusters the experiment will almost always find at least one spurious peak inside that band — the Monte Carlo false-positive rate for this design under phase-randomized nulls is empirically around 40–60%, meaning the FALSIFIES region is nearly unreachable regardless of true causal structure.

   Resolved

   Generate 1,000 phase-randomized surrogate pairs per region (preserving autocorrelation spectrum of both input series) and compute the full cross-correlation envelope; redefine the FALSIFIES threshold operationally as: the 6–18 month lag correlation fails to exceed the 95th percentile of the surrogate null distribution in ≥3 of 4 regions, with a minimum effect size of |r| ≥ 0.20 after autocorrelation-corrected degrees-of-freedom adjustment. Supplement this with a sensitivity sweep across at least two independent drought reanalyses (ERA5 vs. MERRA-2) and two discourse proxies (news volume vs. Reddit sentiment) to bound instrument-noise variance; if the inter-product spread in lag-peak location exceeds ±4 months, the claimed 6–18 month window must be widened accordingly or the hypothesis restated as falsified by irreducible measurement ambiguity.

Unlike the static stress tests above (synthesised against the frozen catalogue spec), this is what a 3-voice council found in the most recent biweekly review. Refreshed on the 1st and 15th of each month at 09:00 UTC. Each voice runs one bounded web search via Anthropic's web_search_20260209 tool, cites what it finds, and recommends a verdict. The verdict diverges from the curated catalogue status (forming) — the synthesis below explains why.

Agent draft incorporating the 9 cited findings from the live council above. Not auto-merged — surfaces here for human review. To accept, open a PR editing site/src/_data/hypotheses.json with the revised fields below. To reject, ignore and the proposal will refresh on the next council run.

1. forming

   May 30, 2026 · added to catalogue at status "forming"

• The UN Peacebuilding Fund ($100M annual envelope) redirects 25–35% of discretionary project disbursements to climate-proximate conflict-risk geographies across the Sahel and Horn of Africa by the 2028 replenishment cycle, adopting sub-national drought-index exceedance and crop-failure proxy breach as tier-1 eligibility triggers.
• World Bank IDA's Crisis Response Window incorporates CHIRPS-derived drought-week thresholds and NDVI crop-failure proxies as parametric pre-commitment triggers for up to $800M in stabilization tranches under the IDA-21 allocation cycle, compressing average disbursement lag from approximately 11 months post-onset to under 4 months by end-2027.
• African Risk Capacity expands its drought-parametric sovereign risk-transfer programme from 35 to 43–47 member states, adding a conflict-prevention tranche of $400–600M in annual coverage capacity targeted at high-climate-conflict-risk Sahel and Central Asian corridors within three underwriting cycles (by 2028).
• IFC and the InsuResilience Global Partnership co-structure a parametric climate-displacement bond (notional $200–500M) using dual triggers — CHIRPS drought-index breach plus ACLED 90-day conflict-onset exceedance probability — bringing the first issuance to market within 18 months of IPCC AR7 Working Group II endorsement of climate-conflict lead-time evidence.
• USAID Bureau for Humanitarian Assistance shifts 20–30% of its annual prepositioning budget (approximately $120–180M) upstream by 9–12 months relative to current conflict-onset thresholds, following formal adoption of ACLED-climate composite early-warning indicators as mandatory allocation criteria no later than the FY2027 Congressional Appropriations cycle.