Seldon Vault_

◌

> About Seldon Vault

What is Seldon Vault?

Seldon Vault is a free, public AI-powered geopolitical forecasting engine. It uses a multi-agent LLM architecture with 11 specialized analysts (including dual-persona pairs with opposing cognitive styles), an adversarial Skeptic with fact-checking capabilities, and a Seldon Arbiter to generate daily probabilistic forecasts across geopolitics, economics, technology, social dynamics, military affairs, cybersecurity, climate, and domestic politics. Inspired by Hari Seldon's psychohistory from Isaac Asimov's Foundation series.

Who is it for?

Anyone interested in understanding global risks and trends. Researchers, analysts, journalists, policy professionals, and curious minds who want data-driven probabilistic assessments instead of pundit opinions. All forecasts are free, public, and require no registration.

What makes it unique?

Unlike prediction markets (Polymarket, Metaculus) that aggregate human bets, Seldon Vault uses an ensemble of specialized AI agents that independently analyze signals and debate through an adversarial review process. Three key domains run dual-persona pairs (Hawk/Dove, Bull/Bear) with opposing cognitive styles — the same news analyzed by an optimist and a pessimist, then merged into enriched forecasts with a spread metric showing the zone of disagreement. Solo-domain analysts use multi-model Council debate (DeepSeek, GPT, Claude) for diversity instead. The final Seldon Arbiter uses a ReACT loop — iteratively reasoning and calling tools (searching analogies, querying economic indicators, fact-checking claims) before producing its synthesis. Every forecast includes transparent reasoning from each agent, a Skeptic's critique with fact-checks, and Bayesian probability updates as new evidence arrives.

How accurate is it?

Every resolved forecast is scored using Brier Score — a rigorous mathematical measure of calibration. A Brier Score of 0.0 is perfect, 0.25 is random chance. Our calibration curve and per-sector accuracy metrics are publicly available on the Metrics page. We track per-agent accuracy to continuously improve each analyst's calibration.

What data sources does it use?

Seldon Vault collects signals from RSS news feeds, GDELT (Global Database of Events), ACLED (Armed Conflict Location & Event Data), Polymarket, Metaculus, FRED (Federal Reserve Economic Data), and other open sources. The Skeptic agent performs independent fact-checking using web search to verify claims before any forecast is published.

How often are forecasts updated?

New forecasts are generated daily at 08:00 UTC. Existing active forecasts receive Bayesian probability updates every 6 hours as new evidence arrives. Maximum daily probability shift is capped at ±15% to prevent overreaction to single events. Forecasts range from short-term (1-7 days) to century-scale (30-100 years).

Is it available in multiple languages?

Yes. All forecasts are generated natively in both English and Russian by the Seldon Arbiter. The interface supports full bilingual switching. We plan to add more languages in the future.

Is there an API?

Yes. Seldon Vault provides a free, public read-only REST API. You can programmatically access daily forecasts, browse the full archive with filters, retrieve accuracy metrics, and get real-time updates via Server-Sent Events. See the API Documentation page for details.

What is the Seldon Plan?

The Seldon Plan is our monthly long-term forecasting report. While daily forecasts cover events on a 1–90 day horizon, the Seldon Plan looks 1–10 years ahead. It runs 6 specialized structural analysts through council debate, a methodological Skeptic review, and final synthesis by Seldon (Claude Opus) to produce master scenarios — interconnected global trajectories with probabilities summing to ~100%. Each report also identifies critical junctures (bifurcation points) and leading indicators to track quarterly.

What is The Mule?

The Mule is our weekly experimental contrarian analysis, named after the character from Asimov's Foundation who broke Seldon's rational predictions. While the main pipeline produces evidence-based forecasts, The Mule deliberately looks for hidden connections, alternative explanations, and non-obvious patterns that conventional analysts overlook. It takes the week's mega-chains (major ongoing stories), resolved forecast outcomes, and 14 world leader behavioral profiles, then investigates using 5 tools (web search, historical analogies, economic indicators, leader profiling, entity connections). The result is 2–4 alternative narratives with a self-assessed absurdity index (0.0–1.0), cui bono analysis, and evidence chains. These are NOT predictions — they are food for thought. Previous week's narratives are fed back for theory evolution (new → updated → confirmed → rejected).

> Glossary

Brier ScoreForecast accuracy metric: (probability − outcome)². Lower is better. 0.00 is perfect, 0.25 is random chance, above 0.25 is worse than random. Allows honest comparison of forecast confidence against reality.
Calibration CurveA chart comparing stated probabilities to actual outcomes. A perfectly calibrated system: if the forecast says 70%, the event happens 70% of the time. Points on the diagonal = ideal. Above = underestimating risk, below = overestimating.
NarrativeA causal chain of interconnected forecasts. When one event raises or lowers the probability of another, the system links them into a narrative. Probabilities propagate through the chain with 0.5 dampening per hop, up to a maximum depth of 3 hops.
Seldon CrisisDetected on a narrative (causal chain) when all of the following hold: 4+ sectors covered, 3+ forecasts in the chain, average probability ≥ 75%, average causal link strength ≥ 0.6, and 3+ distinct analysts contributed. Distinct from Critical Alert (a single forecast with P > 80%). A reference to the 'Seldon Crises' in Asimov's Foundation — bifurcation points in history.
Forecast HorizonThe time range of the forecast: short-term (1–7 days), medium-term (8–30 days), long-term (31–90 days), yearly (91–365 days), decade (1–10 years), generational (10–30 years), century (30–100 years). Structural signals are automatically routed to long-term analysis.
Bayesian ProbabilityProbability is not a historical frequency but a degree of belief in an event given current evidence. Updated via Bayes' theorem: each new signal shifts probability through a likelihood ratio (LR). Range: 5%–95%, never 0% or 100%.
SkepticA veto-empowered verification agent. Critiques every analyst proposal: fact-checks via web search (Tavily), detects media biases, scores argument quality (0–100). Proposals scoring below 50 are automatically rejected before reaching the Arbiter.
Likelihood Ratio (LR)A coefficient indicating how much a new signal changes event probability. LR > 1 increases probability, LR < 1 decreases it, LR = 1 is neutral. Range: 0.1–10. The Skeptic can reduce LR but not amplify it. Formula: P(new) = LR × P(old) / (LR × P(old) + (1 − P(old))).
Master ScenarioOne of 2–4 interconnected global development pathways for the next decade, produced by the Seldon structural pipeline. Each has a probability (summing to ~100%) and describes a coherent trajectory spanning economics, geopolitics, technology, society, climate, and military domains.
EpochSequential number of a monthly Seldon Report. Each epoch is a fresh long-term forecast that builds on the previous one — Seldon sees what he predicted last month and updates his scenarios as the world changes.
Critical JunctureA key bifurcation point in the future — a moment when decisions or events could switch the world from one master scenario to another. Each juncture has a timeframe, trigger probability, and affected domains.
World State BriefA structured snapshot of the world at the time of structural analysis: key indicators, structural forces, cycle stages, and recent forecast resolutions for each domain. Compiled from World Bank, IMF, UN, and OWID data plus the system's own forecast track record.
Structural SkepticA specialized critic for long-term forecasts that attacks systematic cognitive failures: linear extrapolation bias, narrative coherence trap, historical determinism, technology hubris, black swan underpricing, and base rate neglect. Scores each domain 0–100 and can reject briefs with multiple deadly traps.
Agent PersonaA cognitive profile assigned to an analyst agent, defined by four dimensions: risk appetite (conservative→aggressive), contrarian index (consensus-seeking→contrarian), temporal focus (reactive→strategic), and confidence style (cautious→decisive). Shapes how the agent interprets the same evidence. Not a hard rule — a default tendency that evidence can override.
Dual-Persona PairTwo agents covering the same domain with opposing personality profiles. Hawk/Dove for geopolitics and politics, Bull/Bear for economics. The same signals are analyzed from two angles: one optimistic, one pessimistic. Their proposals are later merged by the Merge Layer into enriched forecasts with a spread metric.
Merge LayerA post-Skeptic processing step that matches dual-persona proposals by title similarity (Jaccard ≥ 0.80) and fuses them into enriched proposals. Produces a weighted average probability and a persona spread — the absolute gap between optimist and pessimist estimates. No LLM calls, pure arithmetic. Unmatched proposals pass through unchanged.
Persona SpreadThe absolute difference between optimist and pessimist probability estimates for the same topic. A spread of 5% means strong convergence — both perspectives agree. A spread of 40% signals a genuine zone of uncertainty where reasonable analysts disagree. Displayed to the Seldon Arbiter as an uncertainty signal.
ReACT LoopReasoning + Acting — an iterative synthesis method used by the Seldon Arbiter. Instead of a single LLM call, Seldon thinks, calls tools (search analogies, query indicators, fact-check, examine event chains, check agent track records), observes results, and iterates. Multiple rounds of investigation produce a more grounded and evidence-based final synthesis.
Quantum CascadeAn interference-aware cascade propagation model. When 2+ shifts target the same forecast, their interaction is modeled as wave superposition: total = classical_sum + interference_term. The interference term can amplify (constructive) or cancel (destructive) depending on coherence between the shift sources. Inspired by quantum mechanical superposition: P = |psi_A + psi_B|^2.
Quantum Persona MergeAn interference-aware merge model for dual-persona pairs. Instead of simple weighted averaging of Hawk/Dove (Bull/Bear) probabilities, models the merge as wave superposition: P = α²·P_hawk + β²·P_dove + 2αβ·cos(φ)·√(P_hawk·P_dove). The phase angle φ comes from coherence between proposals. When personas unexpectedly agree, constructive interference amplifies the merged probability. When they sharply diverge, destructive interference suppresses it. In shadow mode alongside the classical merge.
Coherence ScoreA value from 0 to 1 measuring how 'in phase' two cascade shifts are. Computed from 4 weighted factors: sector match (0.30), direction match (0.35), temporal proximity (0.15), and entity overlap (0.20). High coherence (near 1.0) means shifts reinforce each other; coherence near 0.5 means classical behavior; low coherence with opposing directions means destructive cancellation.
Persona CoherenceA coherence score (0–1) between dual-persona proposals, computed from 4 weighted factors: probability direction — are both on the same side of 50%? (weight 0.35), confidence alignment — same confidence level? (0.25), indicator overlap — looking at the same evidence? (0.25), severity agreement — same severity assessment? (0.15). Determines whether interference is constructive (high coherence), destructive (low coherence), or neutral (~0.5).
Interference TermThe quantum correction added to the classical sum: 2 * sqrt(|s_A| * |s_B|) * cos(theta), where theta = pi * (1 - coherence). Positive = constructive amplification, negative = destructive cancellation, near zero = classical behavior. Safety-capped at max_interference_ratio (default 1.0x the sum of absolute shifts).
Shadow ModeA validation regime where the classical formula always determines the actual probability, but the quantum result is computed in parallel and stored as metadata. The frontend displays both values for comparison. Once accuracy is validated, quantum can be promoted to primary.
The MuleA weekly contrarian analysis agent named after the Mule from Asimov's Foundation — the one who broke Seldon's rational predictions by seeing what the plan could not account for. Deliberately explores non-obvious, conspiratorial interpretations of world events. Uses 5 investigative tools (web search, historical analogies, economic indicators, leader profiles, entity connections) to produce 2–4 narratives with absurdity index, cui bono analysis, and evidence chains. NOT predictions — food for thought.
How It Works →API →