Skip to content

Home

Jump to bottom
OfficialSubmind edited this page Apr 6, 2026 · 1 revision

SubMind — Intelligence Research Engine

A lawyer presenting evidence in court, not a chatbot guessing at answers.

What SubMind Is

SubMind is a forensic intelligence engine that produces auditable, probabilistic predictions. Every claim is traceable to specific sources, every number has a source chain, and the full reasoning is auditable.

Core Design Principles

  1. Courtroom Analogy: The prediction is the verdict. Sub-claims are arguments. Sources are exhibits. The Brier score is the lawyer's win/loss record.
  2. Counter-evidence is required, not optional: Adversarial dual-search means the system architecturally must search for evidence AGAINST every claim.
  3. Source independence at fetch time: Correlated sources are penalized before aggregation, not after.
  4. Confidence intervals, not point estimates: Uncertainty is expressed as ranges derived from source quality variance.
  5. Public accountability compounds: Every resolved prediction improves the track record, which is the primary moat.

Architecture

lib/       → Immutable foundation (LLM routing, search, grounding, scoring)
pipeline/  → Product logic (grows by stage addition only)
api/       → Thin glue (Vercel serverless routes)
components/→ Pure UI (no business logic)

Core Pipeline Flow

  1. Parse Query → Classify domain, extract entities
  2. Claim Decomposition → 3-5 testable sub-claims (LLM + Zod)
  3. Adversarial Dual-Search → Supporting + counter-evidence per sub-claim
  4. Grounding Validation → 4-status: GROUNDED, UNGROUNDED, UNVERIFIABLE, CONTRADICTED
  5. Probability Scoring → Confidence intervals from source quality variance
  6. Aggregation → Bayesian with explicit uncertainty
  7. Store → Supabase (predictions, sub_claims, accuracy_scores)

The Moat

  • Accountability layer: Public Brier scores, auto-resolution, calibration curves
  • Source independence scoring: Penalizes correlated sources structurally
  • Adversarial architecture: Counter-evidence is mandatory, not optional
  • Entity extraction: Cross-query pattern detection over time

Tech Stack

Layer Technology
Frontend Next.js 15, React 19, Tailwind CSS 4
LLM (primary) Gemini 2.0 Flash
LLM (failover) Groq, Cerebras, OpenRouter, Anthropic
Search Tavily → Brave → Serper (cascade)
Database Supabase (PostgreSQL)
Hosting Vercel (Hobby tier)
Validation Zod

Prediction Domains

  • Sports game outcomes
  • Geopolitical conflict and war
  • Viral online trends
  • Technological invention
  • Financial markets (equities, macro, fixed income, commodities, crypto, FX)

Pages

Last updated: April 6, 2026

Clone this wiki locally