Oracle AI Light Heavyweight Accuracy Report: 66.7%, B+ Grade

The Numbers

The Oracle AI model has logged 6 predictions in the Light Heavyweight division across mixed promotions, producing a 66.7% raw accuracy rate with 4 correct calls. The Brier Score of 0.1901 falls in the "good" range (below 0.20), indicating the model's probability estimates are reasonably well-calibrated despite the small sample.

Method accuracy matches overall accuracy at 66.7% — when the model predicted a finish or decision, it identified the correct outcome type two-thirds of the time.

Context matters: 66.7% exceeds the ~55-60% benchmark for unaided MMA prediction, but the 6-fight sample is extremely limited. A single additional miss would drop accuracy to 57.1%; a single additional hit would raise it to 71.4%. The Brier Score of 0.1901 suggests the model is assigning probabilities with appropriate uncertainty — neither overconfident nor excessively conservative — but this metric stabilizes only with larger samples. Treat these figures as preliminary indicators, not settled performance.

Confidence Calibration

The tier breakdown reveals a stark pattern driven by sample scarcity:

| Tier | Predictions | Actual Win Rate | Expected | Edge | |------|-------------|-----------------|----------|------| | Lock (85%+) | 0 | — | 92.5% | — | | High (70-84%) | 0 | — | 77% | — | | Medium (60-69%) | 3 | 100% | 64.5% | +35.5% | | Low (50-59%) | 3 | 33.3% | 54.5% | -21.2% |

The model has no Lock or High confidence predictions in this weight class — a notable absence suggesting either genuine competitive parity at 205 lbs or conservative probability assignment. All six predictions fell in Medium or Low tiers.

Medium-tier picks are overperforming dramatically: all three won, producing a +35.5% edge against 64.5% expectation. Conversely, Low-tier picks underperformed by 21.2%, winning just once in three attempts.

The positive edge in Medium tier is the model's only demonstrated advantage in this slice. The absence of higher-confidence predictions prevents assessment of whether the model can identify clear favorites — a critical capability untested here.

What the B+ Grade Means

The B+ reflects solid calibration and above-baseline accuracy constrained by severe sample limitations. The model is not failing; it is unproven at volume.

Honest weaknesses:

• Six fights provide minimal statistical power. Variance dominates signal.
• No high-confidence predictions suggest the model either sees Light Heavyweight as genuinely unpredictable or lacks confidence to commit — neither is ideal for users seeking decisive calls.
• Cross-promotion mixing (per scope note) means these fights span different rule sets, cage sizes, and athletic commissions. The model's performance here does not translate directly to any single promotion's environment.

This is a cross-promotion weight-class slice, not the public headline record. Do not compare these figures directly against promotion-specific reports. The B+ is a conditional grade: promising, pending verification.

Methodology

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

Full methodology breakdown →

Promotion scope: Mixed / not promotion-specific (cross-promotion Light Heavyweight slice).