SHIELD Geopolitical AI

01 · Forecast Output

Probabilistic Scenario Bands With Calibrated Confidence

For any user-defined territory, SHIELD returns a ranked set of forward-looking risk scenarios across multiple time horizons (30, 90, and 180+ days). Each scenario specifies a calibrated probability, a structured narrative, the supporting retrieved evidence, and the adversarial critique notes produced during multi-agent review.

The output is engineered for analytical comparability rather than narrative consumption: every probability is paired with a watch list, an explicit confidence indicator, and the source identifiers used to construct the assessment, preserving traceability between the model's claims and the underlying corpus.

03 · User Workflow

From Territory Definition To Calibrated Report

The interaction model proceeds in five sequential stages. The user delineates a region of interest, either through manual country selection or through geometric primitives (rectangle, circle, polygon) on the interactive map. The system resolves the selection to a canonical set of ISO-3166 alpha-3 country codes and queries the indexed corpus for evidence relevant to that scope.

Retrieved evidence is then forwarded to the multi-agent reasoning module, where an iterative debate produces a candidate assessment. The orchestrator synthesises the debate into the final structured scenario report. Throughout the workflow, all intermediate artefacts are persisted, enabling post-hoc audit of the analytical chain.

04 · Three-Layer Architecture

Forecasting, Retrieval, and Multi-Agent Reasoning

The platform is structured as three orthogonal analytical layers, each addressing a distinct concern. The Forecasting Engine produces base-rate probabilities through gradient-boosted ensembles (LightGBM and XGBoost) trained on a country-month feature matrix that integrates event counts, conflict fatalities, governance indicators, geographic proximity, and graph-derived signals.

The Text Intelligence layer performs retrieval-augmented evidence assembly through dense embeddings and country-boosted semantic search, with a DeBERTa-v3 escalation classifier providing supplementary structural signal. The Multi-Agent LLM layer then orchestrates the analyst-critic-orchestrator debate over the combined evidence and forecast features. The separation of concerns is deliberate: each layer can be evaluated, retrained, or replaced independently, preserving the explainability and auditability of the analytical pipeline.

05 · Multi-Agent Debate

Three Coordinated Agents: Analyst, Critic, Orchestrator

Reasoning is delegated to three coordinated language-model agents. The Analyst agent drafts an initial assessment over the retrieved evidence and the quantitative forecast features. The Critic agent performs an adversarial review of that assessment, surfacing methodological gaps, counter-evidence, single-source biases, and unaddressed alternatives. The Orchestrator agent synthesises the debate into a final report with calibrated probabilities and explicit uncertainty.

By construction, every claim emitted by the system is anchored to a retrieved evidence chunk; free generation is structurally precluded. This grounding requirement, combined with the adversarial critique pass, materially mitigates the hallucination risk inherent to language-model reasoning in high-stakes domains.

06 · Retrieval-Augmented Generation

Indexed Corpus, Country-Aware Semantic Search

The retrieval pipeline ingests evidence from fifteen open data sources, normalises records into a unified relational schema (DuckDB), chunks the resulting text, and produces dense vector representations using BGE-M3 embeddings indexed in a FAISS inner-product store. A country-aware filter restricts candidate documents to the resolved ISO-3 set, ensuring that the language model is never exposed to evidence outside the user's chosen scope.

The corpus contains approximately 130,000 country-month observations spanning daily event-level granularity and monthly forecast aggregations across six feature tables. The trained gradient-boosted forecaster reports a macro ROC-AUC of 0.989 and a precision-recall AUC of 0.805 on a held-out test set.

07 · Data Sources

Fifteen Open Feeds Across Five Thematic Categories

All evidence consumed by the platform is sourced from open, keyless feeds, ensuring full reproducibility of the underlying base. The aggregated corpus spans five thematic categories, providing complementary signals for the analytical pipeline.

Conflict events · ACLED, GDELT, UCDP, COW
News and humanitarian reporting · RSS feeds, ReliefWeb
Macroeconomic and governance indicators · World Bank, V-Dem
Geographic and demographic baselines · GeoBoundaries, WorldPop
Environmental signals · FIRMS (active fires), ERA5 (climate reanalysis)