Outcome Tree · Global Risk Studio

Source-bound forecasts

The Studio consumes the canonical dependency-capped-authored-llr-v2.4.0 engine. It captures the model and the exact normalized-point, cross-domain-alert and Incident Graph input objects used in that generation. Before publishing the Studio snapshot, it reconstructs the forecast with the same original model time and verifies the posterior and dependency-contribution structure. Model and input-manifest hashes identify the captured analysis.

The capture hook adds no new upstream polling and does not change model mathematics, authored priors, forecast issuance or resolution records. A Studio capture failure cannot fail the existing model. Reader filters never request arbitrary URLs or collect new provider data.

These are uncalibrated authored-model estimates. At introduction on2026-10-05, the public registry contained4,656 issued forecasts and zero resolved outcomes. There is no demonstrated predictive accuracy to claim. Missing Brier/log-loss values remain unscored, not zero.

Horizons and outcome structure

Eight non-exclusive global scenarios are modeled at24-hour,7-day and30-day horizons: banking/liquidity, cross-asset dislocation, strategic supply corridors, grid/digital infrastructure, geopolitical escalation, health/biosecurity, mobility/official controls and food/energy pressure. Their probabilities do not sum to100%.

The explorer shows the selected criterion being met or not met in its precise forecast window. The latter is its mathematical complement, not a claim that the world is safe. Existing three-way conditional pathway allocations are shown numerically only for the source’s published seven-day horizon. The Studio does not invent24h/30d child allocations or unsupported90-day forecasts.

Country, personal and investment probabilities are not inferred from these global outputs. Source confidence/support, connector availability and event counts are not automatically event probabilities.

Which inputs matter

The audit distinguishes potential input rows, registry-eligible candidates, retained contributors, and excluded/context-only inputs. Eligibility refers to the captured model cutoff. Incident inputs preserve the engine’s first120-row model bound, typed severity, freshness and contradiction checks. A record outside that bound can be inspected without acquiring predictive weight.

Retained positive and negative survivors resolve to public input records. Missing/invalid primitives are not displayed as zero measurements; a retained contribution without a valid public input prevents Studio publication. Source links omit credential-bearing URLs and owner-bound evidence. The original engine’s approved dependency grouping, top-eight-group selection and log-likelihood bounds remain authoritative.

Observed classes follow the existing registry. An official weather warning is evidence of an advisory, not proof of military action or corridor closure. Source caveats and source classes remain visible in the input inspector. Shared inputs indicate model dependence, not proven causal connections between global risks.

Scenario experiments

For authored prior p and retained net log-likelihood L, the published engine applies logistic(logit(p) + L), with L bounded to[-1.5,2] and final probability to[0.001,0.995]. The Studio can remove whole retained groups, retain only groups whose members are all observed or all modeled, attenuate retained contributions by a0–1 weight, or substitute a user-assumed prior.

Experiments are hypothetical, do not issue forecasts, do not rewrite input data, and do not re-rank excluded candidate signals. Removing every group or giving retained evidence zero weight returns the selected prior. The published sensitivity band is not reassigned to a hypothetical experiment. Leave-one-group-out bars are model sensitivity, not causal effects.

An active experiment stays on its accepted capture when a newer model arrives during passive refresh; the interface announces the new capture. Explicit refresh resets the experiment. Expired model values cannot drive current experiments.

Compound risks

For modeled marginal probabilities a and b, the Studio reports dependence-agnostic Fréchet bounds:

• Both criteria: max(0,a+b−1) ≤ P(A∩B) ≤ min(a,b).
• At least one criterion: max(a,b) ≤ P(A∪B) ≤ min(1,a+b).

No independence assumption is made, and the marginals are not multiplied into a purported joint prediction. These are mathematical constraints on uncalibrated modeled marginals, not a calibrated compound forecast. Selecting the identical scenario twice is treated as the same event.

History and evaluation

The Studio retains up to288 actual captured generations for a rolling live-model trajectory. Missing earlier generations are not synthesized. Curves filter matching model version and metric-registry hash; they reflect rolling forecast windows and are not observed event rates.

Prospectively issued forecast history and resolved-outcome scores come from the existing public registry. Aggregate scores are labeled as covering all reported versions; current-version groups and selected scenario/horizon issuance history remain separate. Registry model-version scoring does not imply a separately supplied feature-registry cohort. Probability-change entries are an in-app ledger, not claims of external alert delivery.

Freshness, saved work and exports

Successful canonical model generations normally update the Studio on the existing five-minute schedule. Snapshots are atomically published with0600 permissions in a0700 directory. The public read API is /agent/api/outcome-studio, schema outcome-studio-v1; pages are capped at100 inputs. Forecasts expire30minutes after the original model generation. Source errors do not create new timestamps or replace the last accepted capture.

Saved experiment recipes are account-local browser storage, not cloud-synced forecasts. Reopening applies a recipe to the current capture and explicitly reports a changed capture or omitted old groups. Full-input export checks the same capture identity on every page and aborts if the model or account changes.

Account transitions fence late responses and clear prior selections, experiments and saved labels. Live values and requests are withheld during replay. Existing owner-private personal-relevance and forecast-resolution workflows remain available under Advanced Tools with their original access controls.