> ## Documentation Index
> Fetch the complete documentation index at: https://docs.supermisson.fun/llms.txt
> Use this file to discover all available pages before exploring further.
# How the AI Works
> The multi-agent debate system behind Supermission's signals. 4 agents, 1 judge, full transparency.
# Multi-agent intelligence
Supermission doesn't use a single model to predict market outcomes. It runs a structured debate between four AI agents with distinct analytical perspectives, then synthesizes the debate into actionable intelligence.
This is the core engine behind [Signals](/ai/signals), [War Room](/ai/war-room), and [Alpha Scanner](/ai/alpha-scanner).
## The agents
**Mandate:** Narrative and source credibility analysis.
Evaluates: source credibility, political dynamics, institutional incentives, social context, and qualitative signals that shape market narratives.
**Mandate:** Microstructure and price-in analysis.
Evaluates: whether new information is already priced in, trading volume patterns, liquidity depth, bid-ask spread dynamics, and market efficiency.
**Mandate:** Ask "what if everyone is wrong?"
Looks for: crowded trades, false consensus, scenarios the market isn't pricing, regime changes, black swan potential.
**Mandate:** Base-rate anchoring with statistical rigor.
Anchors on: historical frequency of similar events, current market prices relative to base rates, time remaining to resolution, and calibration against prior outcomes.
## The pipeline
Each agent receives the same market data: current price, volume, orderbook depth, resolution date, and category. Additionally, each agent pulls from external sources — news feeds, social sentiment (X, Reddit), on-chain data, and historical patterns for similar markets.
Agents analyze independently. They don't see each other's work. Each produces:
* A probability estimate (0-100%)
* Confidence level (how sure they are of their own estimate)
* Key evidence (sourced claims)
* Key doubt (what could make them wrong)
A judge agent reviews all four analyses and produces:
* **Final probability** — weighted synthesis of all four estimates
* **Confidence score** (0-100) — based on agent agreement and evidence quality
* **Key risk** — the single biggest threat to the thesis
* **Agreement ratio** — how many agents lean the same direction
If the synthesized probability diverges meaningfully from market price, a signal fires with the full reasoning chain attached.
## Confidence scoring
| Score | Agent Agreement | Evidence Quality | Meaning |
| ---------- | ------------------ | -------------------- | --------------------------------------- |
| **80-100** | 4/4 or 3/4 aligned | Strong, multi-source | High conviction — agents largely agree |
| **60-79** | 2/4 aligned | Moderate, some gaps | Moderate conviction — meaningful debate |
| **40-59** | Split or 1/4 | Weak or conflicting | Low conviction — proceed with caution |
## Conviction levels
The War Room classifies conviction based on agent agreement:
* **Extreme** — All 4 agents agree at 80%+ confidence. Triggers real-time SSE alerts with sound notification.
* **High** — 3/4 agents agree. Strong directional signal.
* **Moderate** — 2/4 agents agree. Some evidence on both sides.
* **Contested** — Agents split evenly. Market is genuinely uncertain.
## Analysis tiers
AI resources are allocated based on market importance:
| Tier | Criteria | Treatment | Frequency |
| ---------- | --------------------------- | ------------------------------ | ------------ |
| **Tier 1** | Trending OR volume > `$50K` | Full 4-agent debate + judge | Every 15 min |
| **Tier 2** | Volume > `$5K` | 2-agent analysis (Bull + Bear) | Every 30 min |
| **Tier 3** | Everything else | Single-model sentiment | Every 60 min |
Markets also get **event-driven** re-analysis when matched news arrives, regardless of tier.
## Sports-specific intelligence
Traditional news/sentiment analysis fails for sports (28-37% hit rate). Sports outcomes depend on **live performance**, not information edges. Supermission uses a price-movement detection engine for sports markets:
| Algorithm | Signal | Confidence |
| ---------------------- | --------------------------------------- | ---------- |
| **Steam** | Price moves 8%+ in 5-15 min | 60-85 |
| **RLM Proxy** | Tick frequency spike + opposing price | 70 |
| **Spread Compression** | Bid-ask spread \< 50% of 30-min avg | 65 |
| **Late Sharp** | Price move 5%+ within 2h of event start | 75 |
The AI system is calibrated against historical outcomes. After each market resolves, the prediction is scored and fed into the accuracy tracking system. This creates a feedback loop that improves calibration over time.
## LLM infrastructure
| Component | Provider | Purpose |
| ----------------------- | ------------------------------- | ---------------------------------------- |
| **Market Analysis LLM** | MiniMax M2.7 | Multi-agent debate and judge synthesis |
| **Chat Agents LLM** | OpenRouter (GPT-4O-mini) | Oracle, Delphi, and Securo chat agents |
| **Vision** | Gemini 2.0 Flash (Google GenAI) | Image analysis for news context |
| **Streaming** | Vercel AI SDK | Real-time response streaming |
| **Tool calling** | Vercel AI SDK | Structured output and function execution |
AI predictions are not financial advice. The multi-agent system improves signal quality but cannot predict the future. Markets are adversarial — always understand the reasoning chain before acting on a signal.