Oracle AI Middleweight Accuracy Report: 73.3%, B Grade

Cross-promotion weight-class analysis | 15-fight sample | Recorded pre-event, verified post-event

---

The Numbers

The Oracle AI model has recorded 73.3% accuracy (11-4) across 15 middleweight predictions in a mixed-promotion environment. This falls within the competitive range for MMA prediction systems, where 65-75% represents solid performance against closing betting lines.

The Brier Score of 0.2076 sits in the "good" range (threshold: <0.20 for elite, <0.25 for average). This metric—measuring both accuracy and calibration—indicates the model assigns probabilities with reasonable precision, though not yet at elite calibration levels. A Brier Score above 0.20 reveals room for improvement in probability assignment, even when directional picks succeed.

Method accuracy registers 35.7%, reflecting the inherent difficulty of predicting how middleweight fights end. Weight classes with higher knockout rates typically show lower method precision due to finish volatility.

No draws or no-contests were excluded from this sample. All predictions were logged before fight night and verified against official results.

---

Confidence Calibration: Where the Edge Lives

The tier breakdown reveals a calibration inversion that demands attention:

| Tier | Record | Actual Win Rate | Expected | Edge | |------|--------|-----------------|----------|------| | Lock (85%+) | 0-0 | — | 92.5% | — | | High (70-84%) | 2-1 | 66.7% | 77.0% | -10.3% | | Medium (60-69%) | 5-1 | 83.3% | 64.5% | +18.8% | | Low (50-59%) | 4-2 | 66.7% | 54.5% | +12.2% |

The model's positive edge concentrates in Medium and Low confidence tiers—the exact opposite of ideal calibration. High-confidence selections are underperforming expectations by 10.3 percentage points, while Medium picks exceed expected rates by 18.8%.

No Lock-tier predictions have been issued in this middleweight sample, preventing assessment at the highest confidence threshold. The absence of Lock activity itself signals uncertainty: the model has not identified any middleweight bout as sufficiently predictable to warrant maximum conviction.

This pattern—strong performance in moderate-confidence ranges with High-tier drag—suggests the model's probability scaling may be miscalibrated at upper bounds. The directional signal remains valuable; the probability assignment requires refinement.

---

What the B Grade Means

A B grade communicates honest performance: above average, below elite. The 73.3% headline accuracy is respectable. The 0.2076 Brier Score exposes the gap between picking winners and assigning correct probabilities.

Weaknesses are real and documented:

• Sample size limitation: 15 predictions provide directional signal but insufficient volume for statistical confidence in tier-level calibration. One or two result swings materially alter percentages.
• Cross-promotion scope: This report aggregates middleweight data across multiple organizations. Performance varies by promotion due to roster depth disparities, judging consistency, and competition level. These results do not claim comparability to any single-promotion headline record.
• Method prediction lag: 35.7% method accuracy trails directional accuracy significantly, indicating the model reads fight outcomes more reliably than fight dynamics.

The grade reflects this honesty: the model delivers value, but not with the consistency or calibration precision that would merit an A.

---

Methodology

The Oracle AI operates a 57-module prediction engine with specialist voting architecture. Individual modules assess distinct fight dimensions—striking efficiency, grappling transitions, cardio decay curves, stylistic matchups—before ensemble aggregation produces final probabilities.

All predictions are time-stamped before events and verified against official commission results. Draws and no-contests are excluded from accuracy calculations.

This report covers a cross-promotion middleweight sample in a mixed data environment. It does not represent the model's full promotional record or headline public performance.

Full methodology breakdown →

---

Report generated from verified prediction log. Next update upon sample expansion.