Geopolitical Outlook

The forecast engine

The panel presents the existing watchtower-bayes-outcome-tree-v2 global model at24-hour,7-day and30-day horizons. Each scenario retains its model version, authored prior, posterior estimate, original generation time, horizon end, dependency contributions, confirmation/disconfirmation signposts and exact outcome-resolution rules.

The current engine uses authored priors and log-likelihood contributions, not a fitted historically calibrated geopolitical prediction model. At introduction on2026-10-04, its public registry showed4,632 issued forecasts and zero resolved outcomes. No demonstrated accuracy, calibration, Brier score or log loss is invented. The Track Record view uses the current public registry.

For a prior probability p and retained log-likelihood sum L, the model uses posterior = logistic(logit(p) + L). The existing model retains at most eight dependency groups per scenario/horizon and bounds total L to[-1.5,2]. Related inputs share explicit groups rather than each being treated as independent corroboration. The engine’s published source classes and contribution weights remain visible.

Scenarios are non-exclusive and do not sum to100%. A global escalation estimate is not the probability of war in the country selected on the map. Country and region controls filter context and evidence, never silently localize the global probability. Sensitivity bands are model sensitivity ranges, not frequentist confidence intervals or Bayesian credible intervals.

Evidence and scenario fit

Source-native observations, official notices/advisories, modeled context, publisher reporting, unverified social posts and reference context retain separate labels. A weather warning is evidence that an advisory was issued, not proof of a corridor closure or military action. Some global cross-domain model drivers are hazard notices; the forecast inspector explicitly identifies that limitation.

Existing metric-registry admission rules determine predictive eligibility. Missing data, unregistered fields, connector readiness, inventory counts and reference facilities are not automatically predictive evidence. Official-document publication is not its legal effective date. Generic connector summaries and future-dated notices are excluded from the current evidence timeline. Source support scores are not reused as event probabilities.

Validated cache retention

The platform’s validated last-good read-cache marker is distinguished from an upstream-stale result. Within the original30-minute model or one-hour quote ceiling, a validated cached value can be shown as retained, with its original model/quote timestamp. It is not relabeled as a new source observation. Explicit upstream errors/staleness and values beyond their ceiling remain withheld. Successful scheduled model refreshes synchronize the public read cache so a second cache does not lag the model engine.

Interactive stress test

The Forecast Lab can remove selected retained dependency-group contributions from a published estimate. It recomputes the same log-odds formula using the remaining effective contributions and existing bounds. Removing all groups returns the authored prior.

The result is labeled hypothetical. It does not issue a new forecast, change the source model or registry, prove causality, or re-rank previously excluded candidate signals. Published model values and source timestamps remain intact.

Current country context

The map and country matrix reuse the existing country-risk composite, explicitly labeled as model context—not event probability, verified conflict status or a safety assurance. Components retain missingness and source years. In particular, a numeric CDS value with a baseline/undated/stale source date is withheld from the live-quote column. FX values require an observed flag and a supported recent timestamp. Current/future-year IMF debt series can contain estimates; they are identified as source-year context.

Country references come from mledoze/countries under ODbL1.0. The reused base-map attribution is available here. Markers are country reference points, not incident coordinates. Reporting is associated through explicit country codes or conservative country-name mentions; publisher headquarters/edition country is not used as event geography. Ambiguous place/person names and unlocated records remain unresolved. Association does not establish where an event occurred or that all associated countries are affected.

Market expectations and upcoming catalysts

Prediction markets show source-supplied prices for their exact binary Yes/No contracts, with liquidity, recent volume, snapshot time and contract deadline. Expired contracts are excluded even when the upstream API still calls them active. Stale prices are withheld. These are market-implied expectations, not Watchtower model forecasts; unmatched questions are never blended into a supposed consensus probability. A contract end date is not a prediction that the event will occur on that date.

The calendar retains upcoming source dates for elections, meetings and economic releases within90 days. Its response does not declare a timezone; source clock text is displayed without fabricating UTC instants. Consensus, previous and actual fields remain separate, including missing values. Longer-dated market questions and conditional pathways support a strategic view; the model is not extrapolated into invented90-day or annual probabilities.

Sources, collection and freshness

• Public scenario model, forecast-calibration report and probability-change ledger.
• Country-risk context, normalized data points, cross-domain alerts and domain-source summaries.
• Incident Graph and normalized public OSINT observations.
• Public prediction-market and calendar endpoints.
• Reviewed public news/social working corpus: the existing8,000-news/12,000-social row bounds apply. The collector scans that bounded corpus and selects relevant recent references; it does not access private/custom account feeds or entire social-platform timelines.

A separate five-minute collector reads fixed public routes and the public media cache, with bounded concurrency, time/byte budgets, per-source last-good retention, a singleton lock and atomic publication. Reader filters are read-only and never trigger arbitrary URLs or upstream searches. Source freshness is independent of the newest article date. The global model becomes unavailable for current use after30 minutes; market prices after one hour. Source failures, stale snapshots and missing components are shown, not silently replaced with zeros or new timestamps.

Evidence, source and calendar counts are bounded display inventories, not unique confirmed incidents or exhaustive worldwide coverage. Publication mentions are discovery only and source quantity is not independent corroboration.

Account context, export and replay

Account changes fence outstanding responses and clear the prior selection, comparison, hypothetical groups and open dossiers. Explicit same-account verification can restore an untouched prior selection. No personal portfolio, credential, private report or private source configuration is collected into the public outlook. Exports carry the source-qualified snapshot, selected global horizon and any explicitly hypothetical stress test.

Live requests, model values and map coloring are withheld during replay; the retained classic geopolitical diagnostics remain available for captured context. The public API is /agent/api/geopolitical-outlook, schema geopolitical-outlook-v1.