Calibration-weighted truth.

A self-improving prediction oracle where every AI agent's influence is weighted by its historical Brier score. Python engine, Solidity mirror on Base Sepolia, every probability and every weight independently verifiable.

Watch the 90s demo View on GitHub Try the interactive demo
100%
Accuracy
50-case benchmark, seed=42
0.0724
Brier score
Lower than any single agent
742
Python tests
90 adversarial · 55 Foundry · CI green
3.3s
Consensus latency
3 parallel agents, local LLM

How it works

One question, three parallel agents, one calibrated answer. The weight formula is two lines of math; you can read it in swarm_oracle/weights.py and find the same math, bit-for-bit, in contracts/src/CalibrationRegistry.sol.

$ python swarm_verify.py "Did BTC close above 100K on May 5, 2026?" 
========================================================================
  SWARM ORACLE  |  Calibration-Weighted Consensus
========================================================================
Question : Did BTC close above 100K on May 5, 2026?
Agents   : 3
Elapsed  : 3.30s

Individual votes:
  agent-oracle     | strategy=api         | P(YES)=0.030 | conf=0.90 | weight= 10.00 ( 60.1%) ████████████········
  agent-reliable   | strategy=web_search  | P(YES)=0.050 | conf=0.80 | weight=  5.56 ( 33.5%) ███████·············
  agent-novice     | strategy=knowledge   | P(YES)=0.500 | conf=0.00 | weight=  1.07 (  6.4%) █···················

Consensus:
  Weighted P(YES) = 0.0303
  Variance        = 0.0143
  Decision        = NO
========================================================================

1. Question in

Any binary (YES/NO) prediction question. CLI, FastAPI endpoint, or direct Python.

2. Parallel research

Each agent runs a different research strategy — API lookup, web search, knowledge-only — and reasons independently.

3. Calibration weighting

weight = 1 / (brier + ε) scaled by a confidence ramp for new agents. Well-calibrated agents get more vote.

4. Weighted consensus

Linear opinion pool produces a single P(YES). If weighted variance crosses a threshold, the result is flagged DISPUTE rather than forced.

5. On-chain verification

CalibrationRegistry.sol and SwarmConsensus.sol mirror the math in WAD (18-decimal) fixed-point. Anyone can recompute weights from public Brier scores.

6. Self-improving

Every resolution becomes training data. Brier scores update, weights re-derive, future predictions sharpen. The protocol gets smarter without a re-deploy.

The benchmark

A 50-case deterministic benchmark (seed=42), balanced YES/NO, with agents designed to fail on different subsets. DISPUTE = correct abstention when agents genuinely disagree — that is accuracy, not failure. Reproduce with make benchmark.

Method Accuracy Brier ↓ Disputes Notes
swarm 100% 0.0724 18/50 Best Brier of all methods
majority vote 92.0% 0.0785 0
average 98.0% 0.0935 0
agent-oracle 84.0% 0.1029 0 Best single agent
agent-reliable 80.0% 0.1332 0
agent-novice 68.0% 0.2009 0

The swarm protocol beats every single agent on Brier score — including the oracle's 0.1029. The variance gate converts genuine disagreement into honest DISPUTE outcomes rather than forcing a wrong answer. Reproduce locally with make benchmark.

The architecture

Six Python modules feed three Solidity contracts. The Python ↔ Solidity boundary is policed by a 14-test parity suite that compares bit-for-bit on a frozen corpus.

Swarm Oracle Architecture Calibration-Weighted Multi-Agent Prediction Consensus Question Input (CLI / API) verifier.py — ThreadPool Orchestrator agent-oracle strategy: API lookup Brier: 0.08 · weight: 60% agent-reliable strategy: web search Brier: 0.15 · weight: 34% agent-novice strategy: knowledge Brier: n/a · weight: 6% consensus.py + weights.py Weighted linear opinion pool · Dispute detection YES NO DISPUTE BASE SEPOLIA L2 — On-Chain Verification Layer CalibrationRegistry Brier storage + WAD weight computation SwarmConsensus Vote aggregation + resolution events RewardDistribution ETH pools · 70/30 split Pull-payment pattern AgentIdentity Soulbound ERC-721 Reputation NFTs Self-Improving Feedback Loop Resolution → Brier update → Weight recalculation → DPO training data 742 Python tests · 55 Foundry tests · 14 parity tests · 3.3s consensus · zero external deps

The on-chain layer

Four contracts, written in Solidity 0.8.24, optimised for Base Sepolia. Weight math is pure WAD (18-decimal fixed-point) — no external libraries, no on-chain sqrt, no approximations.

CalibrationRegistry.sol

Per-agent Brier-score storage. computeWeight(agent) reproduces the Python formula bit-for-bit on a 14-case parity corpus.

SwarmConsensus.sol

Vote aggregation. Reads weights from CalibrationRegistry, computes weighted P(YES) and squared-variance dispute threshold, emits Resolution(YES|NO|DISPUTE).

RewardDistribution.sol

Per-question reward pools. 70/30 split between correctness payouts and calibration improvement. Pull-payment pattern, no re-entrancy surface.

AgentIdentity.sol

Soulbound ERC-721 per agent node. Transfers blocked. Stores cumulative Brier and resolution-count on-chain for transparent reputation.

Try it locally

Two minutes from git clone to a calibration-weighted answer. No API keys. No paid services. Local LLM optional — demo mode runs with zero network calls.

Quickstart 
git clone https://github.com/solmonger/swarm-oracle.git
cd swarm-oracle

# Demo mode — no LLM required, deterministic, 3 seconds
python swarm_verify.py --demo "Did BTC close above 100K on May 5, 2026?"

# Or: full pipeline with a local llama.cpp / Ollama server
export LLM_API_URL="http://localhost:8080/v1/chat/completions"
python swarm_verify.py "Will ETH close above $3,000 on June 1, 2026?"

# Or: one-shot Docker
docker compose up                         # API at http://localhost:8000/docs
docker compose run oracle demo            # CLI demo, no LLM needed

Verify the math

Test & verify 
make test               # 742 Python tests
make test-solidity      # 55 Foundry tests
make test-integration   # End-to-end pipeline
make benchmark          # Reproduce the comparison table above (100% accuracy, 0.0724 Brier)
make adversarial-compare  # Sybil vs bribery attack cost comparison
make economic-model-mvp   # Minimum viable pool by market size