Oracle AI Heavyweight Report: 50% Accuracy, D Grade

The Numbers

The Oracle AI Heavyweight model sits at 50% accuracy across 10 predictions — a coin-flip result that demands honest examination. The Brier Score of 0.2767 falls above the 0.25 "average" threshold, indicating forecasts were both inaccurate and poorly calibrated. Method accuracy trails at 40%, suggesting struggle beyond simple winner-loser binary calls.

Context matters: Heavyweight MMA features the sport's highest knockout rate and shortest average fight duration. Variance is structurally elevated. A 50% mark in this division differs from identical performance in flyweight, where outcomes skew more predictable. Still, this underperforms general MMA prediction baselines, where informed models typically clear 55-60% against closing odds.

The sample size of 10 predictions carries real uncertainty. A 50% rate with n=10 yields wide confidence intervals — true underlying accuracy plausibly ranges 20-80%. This is not a settled estimate. It is a warning that more data is required before drawing firm conclusions about model capability in this weight class.

Confidence Calibration

The tier breakdown reveals a stark pattern: the model fails when confident and succeeds when uncertain.

Lock picks (85%+): Zero attempts. The model never reached peak confidence in any Heavyweight fight during this period.

High confidence (70-84%): 0-for-2 (0% actual vs. 77% expected). A catastrophic -77% edge. Both "High" calls lost. This is the report's most concerning finding — when the model felt strongest, it performed worst.

Medium confidence (60-69%): 3-for-3 (100% actual vs. 64.5% expected). A +35.5% edge. The model's only profitable tier. These fights were read correctly, though the small sample (3) limits confidence in this pattern.

Low confidence (50-59%): 2-for-5 (40% actual vs. 54.5% expected). A -14.5% edge. Slight underperformance relative to expectation, but within noise given sample size.

The inverse relationship between stated confidence and actual results is a calibration failure. In well-calibrated systems, "High" tiers should outperform "Medium" tiers. Here, the opposite holds. This suggests the Heavyweight module's probability estimates are systematically miscalibrated — overvaluing factors that do not translate to victory in this division.

What the D Grade Means

A D grade signals performance below acceptable professional standards. It does not mean the model is useless; it means current outputs should not be relied upon without substantial human judgment overlay.

The grade reflects two simultaneous failures: directional accuracy (50%) and probabilistic calibration (0.2767 Brier). Either alone might warrant a C-range mark. Together, they produce a clear warning.

Weaknesses are specific and addressable. The model appears to misweight variables in Heavyweight — possibly overvaluing striking metrics against the division's volatile finishing dynamics, or failing to account for the accelerated aging curve in larger athletes. The zero Lock attempts suggest the module recognizes its own uncertainty but has not learned to convert that recognition into accurate probability assignment.

Scope limitation: This report covers a cross-promotion Heavyweight slice, not the Oracle AI's full public record. Performance in other weight classes or promotion-specific environments may differ. Do not assume this grade generalizes.

Sample size remains the largest caveat. Ten fights is insufficient for stable measurement. The next ten predictions could shift accuracy dramatically. Treat this as an early diagnostic, not a final verdict.

Methodology

The Oracle AI operates as a 57-module engine with specialist sub-models voting on outcomes. Predictions are recorded before fight night and verified against official results. Draws and No Contests are excluded from accuracy calculation — none occurred in this sample.

This report covers Heavyweight fights across mixed promotions, per the public scope note. It is not promotion-specific.

For full technical methodology, see the complete breakdown.

Last updated: Current reporting period. Next update following additional Heavyweight predictions.