> How Seldon Vault Works
Seldon Vault uses a multi-stage pipeline that mirrors how an intelligence analysis team works: collect signals, analyze independently, challenge assumptions, synthesize, and continuously update.
Step 1: Signal Collection
Raw news signals are collected from multiple open sources: RSS feeds from major news outlets, GDELT (Global Database of Events, Language, and Tone), ACLED (Armed Conflict Location & Event Data), prediction markets (Polymarket, Metaculus), and economic indicators (FRED). Each signal is timestamped, categorized, and deduplicated.
Step 2: Signal Processing
An AI Signal Processor classifies each signal by sector (geopolitics, economics, technology, social, environment, military, cybersecurity), sentiment, importance score, entities, and temporal scope. Signals are split into 'immediate' (short-term events) and 'structural' (long-term trends that affect decade+ horizons).
Step 3: Knowledge Graph
Processed signals pass through three Knowledge Graph layers. First, Signal Clustering groups semantically similar news from different sources into clusters via embedding cosine similarity — Reuters, BBC, and TASS reporting the same event become one cluster with source_count: 3, reducing noise and token costs. Second, Source Ratings inject per-source reliability scores (computed from historical Brier scores) so analysts know which sources to trust more. Third, Event Chain Linking connects clusters across days into temporal storylines, tracking lifecycle stages from rumor through confirmation, escalation, and resolution. A weekly Mega-Chain Consolidation task uses GPT to group hundreds of narrow chains into ~30-50 broad thematic mega-chains (e.g., 'Russia-Ukraine War', 'Middle East Regional War', 'Global AI Industry'). Mega-chains have broader embedding centroids, enabling better forecast-to-chain matching and giving the Seldon Arbiter rich thematic descriptions instead of hundreds of narrow headlines.
Step 4: Decision-Maker Profiles
Before analysis begins, the system loads behavioral profiles of 14 key decision-makers: Trump, Putin, Xi, Erdogan, Netanyahu, Musk, Altman, Huang, Amodei, Kallas, Merz, Macron, Kim Jong-un, and Powell. Each profile includes a BVI (Behavioral Volatility Index, 1–10) — a measure of unpredictability — along with behavioral patterns, historical precedents, and trigger signals. Profiles are filtered by the regions and sectors of current signals: if the topic involves the Middle East, Netanyahu, Erdogan, and Trump are loaded — but not Powell. Analysts use BVI for calibration: high BVI widens the confidence interval and shortens the forecast horizon.
Step 5: Multi-Agent Analysis with Personas
Eleven specialized AI analysts independently assess the signals in parallel. Three key domains run dual-persona pairs with opposing cognitive styles: Geopolitician-Hawk vs Geopolitician-Dove, Economist-Bull vs Economist-Bear, Political-Hawk vs Political-Dove. Five solo-domain analysts each receive a unique personality profile. Every agent uses the Five Pillars framework: Game Theory, Bayesian Reasoning, Systems Dynamics, Historical Analogy, and Network Analysis. Each persona is defined by four dimensions: risk appetite, contrarian index, temporal focus, and confidence style — shaping how the same evidence is interpreted.
Step 6: Council OR Personas — Cognitive Diversity
The system uses two complementary mechanisms for cognitive diversity, never both at once. Dual-persona domains (geopolitics, economics, politics) achieve diversity through opposing personalities: a Hawk and a Dove see the same signals but reach different conclusions. Solo-domain analysts (Technologist) use LLM Council instead: three different providers (DeepSeek, GPT, Claude) analyze in parallel, debate in structured rounds (up to 3), and merge into consensus. This Council-OR-Personas principle keeps LLM costs manageable: 13 calls instead of 33.
Step 7: Skeptic Review
Every proposed forecast undergoes adversarial review by the Skeptic agent. The Skeptic performs independent fact-checking using web search (Tavily API), identifies logical flaws, biases, and unsupported assumptions. Forecasts with a risk score below 50 are automatically rejected. The Skeptic can veto any forecast it considers poorly supported.
Step 8: Merge Layer — Dual-Persona Fusion
After both Skeptics filter proposals, the Merge Layer matches Hawk/Dove (Bull/Bear) pairs that address the same topic using title similarity (Jaccard threshold ≥ 0.80). Matched pairs are fused into enriched proposals: weighted average probability + persona spread (the gap between optimist and pessimist estimates). A high spread signals genuine uncertainty; a low spread signals convergent evidence. No LLM calls — pure arithmetic. Unmatched proposals pass through unchanged.
Step 8b: Quantum Persona Merge (Shadow)
After the classical Merge Layer fuses Hawk/Dove pairs, the system computes a quantum interference shadow. Instead of just averaging probabilities, it models the merge as wave superposition: Merged P = α²·P_hawk + β²·P_dove + 2αβ·cos(φ)·√(P_hawk·P_dove). The phase angle φ is derived from coherence between proposals — computed from 4 factors: probability direction (both on same side of 50%?), confidence alignment, indicator overlap, and severity agreement. When personas unexpectedly agree (high coherence), interference is constructive and the quantum merge exceeds the classical average. When they sharply diverge, interference is destructive and suppresses it. Currently in shadow mode: the classical weighted average always determines the actual probability. The quantum result is stored as metadata for comparison. After 30 days, Brier scores of both approaches will be compared to decide on promotion.
Step 9: Seldon Synthesis (ReACT)
The Seldon Arbiter uses a ReACT loop (Reasoning + Acting) to synthesize approved forecasts. Instead of a single LLM call, Seldon iteratively reasons and calls tools: searching historical analogies, querying economic indicators (FRED), fact-checking claims via web search, examining event chains, and reviewing agent track records. After multiple rounds of investigation, it selects the top 3-5 forecasts, assigns calibrated probabilities (5%-95%), generates English descriptions, and detects Seldon Crises. Translation to Russian is handled separately.
Step 10: Bayesian Updates
Active forecasts are updated every 6 hours using Bayesian inference. When new evidence arrives, the system recalculates probabilities based on evidence strength and prior confidence. Maximum daily shift is capped at ±15% to prevent overreaction. The full probability history is recorded and displayed as a chart for each forecast.
Step 10b: Quantum Cascade
When multiple cascade shifts converge on the same forecast (a 'fork-target'), the system models their interaction as wave interference rather than simple addition. Each pair of shifts is scored for coherence — how aligned they are in sector, direction, timing, and shared entities. High coherence amplifies constructive shifts (both pushing the same way); low coherence with opposing directions causes destructive cancellation. Currently in shadow mode: classical formula always applies, quantum result is computed alongside for comparison. The purple dashed line on probability charts shows what the quantum formula would produce.
Step 11: Accuracy Tracking
When a forecast resolves (event occurs or doesn't within the horizon), it is scored using Brier Score: (predicted probability - actual outcome)². Lower is better. 0.0 is perfect prediction, 0.25 is random chance. We track accuracy per forecast, per agent, per sector, and overall. This data feeds back into agent calibration — underperforming agents receive their accuracy stats in prompts to self-correct.
Step 12: Institutional Learning
Resolved forecasts receive automated post-mortem analysis: which analyst was most accurate, what error pattern occurred (anchoring bias, overconfidence, etc.), and key lessons learned. When similar topics arise in the future, each analyst sees their own past accuracy on comparable situations — creating a feedback loop where agents learn from their mistakes over time.
Step 13: Auto-Resolution
A Resolution Agent automatically verifies whether forecasts have come true. For quantitative forecasts (interest rates, asset prices), it queries APIs directly (FRED, Yahoo Finance). For qualitative forecasts (geopolitical events, political decisions), it uses web search to find evidence and an LLM to assess whether the forecasted event occurred. High-confidence verdicts auto-resolve the forecast; lower confidence cases are flagged for manual review. Crisis-level forecasts are never auto-resolved.
Step 14: Seldon Plan — Long-Term Structural Forecasting
Monthly, a separate structural pipeline runs 6 futurist analysts (Economist, Geopolitician, Technologist, Sociologist, Climatologist, Military) through multi-model council debate. They receive structural data (World Bank, IMF, UN Population, OWID), historical analogies via RAG, and a world state brief synthesized from 90 days of daily forecast history. A specialized Structural Skeptic attacks their methodology using 6 'deadly traps' of long-term forecasting (extrapolation bias, narrative coherence, historical determinism, etc.). The Seldon Arbiter (Claude Opus) then synthesizes everything into 2–4 master scenarios with probabilities, critical junctures, and leading indicators. Each month, Seldon sees his previous report to provide continuity.
Step 15: The Mule — Weekly Contrarian Analysis
Every Sunday, a separate pipeline runs The Mule — a conspiratorial meta-analyst named after the character who broke Seldon's rational plan in Asimov's Foundation. It takes the world snapshot (active mega-chains), the week's resolved forecast outcomes (what happened and what didn't), and previous narratives for theory continuity. The Mule — a persona inspired by Pelevin, Zhirinovsky, and Alex Jones — investigates hidden connections using 5 tools: web search for fact-checking, historical analogies via RAG (including 16 real conspiracy cases — CIA coups, LIBOR scandal, NSA surveillance), economic indicators from FRED, decision-maker behavioral profiles (14 world leaders), and entity connection mapping. The output is 2–4 deep narratives with a self-assessed absurdity index, cui bono analysis, evidence chains, and indicators to watch. Theory lifecycle tracks evolution across weeks: new → updated → confirmed → rejected. These are explicitly NOT predictions — they are alternative interpretations designed to challenge conventional thinking.