Four frontier AI models are the input. The engine is the output. This is the proprietary layer that turns disagreement into calibrated probability — and makes EDJE a trained prediction assistant, not a chatbot wrapper.
Every market gets scored through a twelve-module pipeline. Forty-seven heterogeneous signal sources — macro press, polling aggregators, sports data, crypto markets, state media inversions, whale flow, Reddit sentiment, X-curated sharps, pollster blends, commodity + geopolitical primary reporting — feed the reconciliation engine. The ensemble runs Algorithm 4.0, M1 through M12, with an AI secondary opinion as the final check. The user sees one clean answer, not forty-seven competing opinions.
The weights weren't chosen — they were solved for. We backtested every model's standalone accuracy across 148 resolved markets spanning politics, sports, macro, crypto, and culture. Then we ran a constrained Brier-score minimization: find the weight distribution (summing to 1.0) that maximized ensemble calibration against ground truth.
The output: Claude pulls more weight because its reasoning is tighter on structural claims. Perplexity carries real citations and verifies numbers against sources. ChatGPT is the balanced generalist. Grok adds live-event signal. Each weight reflects where that model outperforms the group — not where it votes most often.
Weights are re-calibrated quarterly as new markets resolve. The engine learns from the markets it predicts.
Four stages. Each one is doing real work. The whole pipeline runs in ~5 seconds end-to-end because the models fire in parallel, not sequence.
Your question hits Claude, ChatGPT, Grok, and Perplexity simultaneously. Total latency = slowest model, not sum of models (~5s).
Each model returns a probability + factor reasoning + sources. The engine collects all four into a structured vote block.
If any model is >15 points from the group median, it's flagged as "outlier signal" and named explicitly in the final output.
Claude Opus runs reconciliation prompt engineering against the trained weights, producing one calibrated answer with confidence interval and risk flags.
A naive weighted average fails exactly where it matters: when models disagree sharply. Picture this — three models say 70%, one says 20%. The 20% isn't noise. It's a signal: one model sees something the others don't.
EDJE's engine treats disagreement as information. The magnitude of disagreement tightens or widens the confidence interval. Outliers get explicit callouts: "Grok flags base-rate concern — track this." The user doesn't lose the signal in the averaging.
A model is calibrated if its 70% confidence actually resolves 70% of the time. Not better, not worse. Reality tracks the prediction. This is the hard metric. Public ledger at getedje.com/transparency.
Every resolved market updates the training set. Every quarter the weights get re-calibrated. The ensemble has gained +16 points of accuracy since launch, while the best single model stays flat.
The frontier models are commodities. Anyone can call Claude, ChatGPT, Grok, Perplexity with an API key. The moat is everything that sits on top. That's EDJE.
Four peer-reviewed research artifacts published to the EDJE library. Each one documents a specific layer of the engine. Open-source where appropriate, proprietary where the math is the moat.
Brier-score minimization against 148 resolved markets. Deriving the 0.32 / 0.26 / 0.19 / 0.23 distribution.
READ → ◆ METHODOLOGY · VOL IIHalf-Kelly implementation when probability comes from a weighted ensemble, not a point estimate.
READ → ◆ METHODOLOGY · VOL IIIWhy using Claude Opus to synthesize four model outputs beats naive weighted averaging on disagreement scenarios.
READ → ◆ METHODOLOGY · VOL IVThe >15-point threshold, false-positive rates, and why we surface outliers to users instead of absorbing them.
READ →