Oracle AI Featherweight Accuracy Report: 57.9%, Grade D

Cross-promotion weight-class slice — not the public headline record

The Numbers

The Oracle AI has recorded 19 predictions in the Featherweight division across multiple promotions, with 11 correct calls for a raw accuracy of 57.9%. The Brier Score sits at 0.2597, which falls above the 0.25 threshold for average predictive performance and well above the 0.20 mark for "good" calibration. By standard forecasting benchmarks, this indicates meaningful room for improvement in how the model assigns probability to outcomes.

Method accuracy — whether the model correctly predicted how a fight would end — registers at 47.4%, trailing even the modest win-prediction rate. In MMA prediction broadly, public models and betting markets typically operate in the 60-70% accuracy range for straight picks, though this varies heavily by data environment and fight selection. The Featherweight division's speed, volume, and finishing volatility present known analytical challenges that may contribute to this underperformance.

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 Model Succeeds and Fails

The tier breakdown reveals a systematic inversion that demands attention.

Lock picks (85%+ confidence): Zero attempts. The model issued no high-certainty Featherweight predictions, so this tier offers no data.

High confidence (70-84%): 2 correct from 3 attempts (66.7%). This trails the 77% expected rate by -10.3% — a modest underperformance in a tiny sample.

Medium confidence (60-69%): The clearest problem zone. Just 3 of 8 correct (37.5%) versus 64.5% expected, producing a -27% edge. The model is systematically overconfident here, treating uncertain Featherweight matchups as more predictable than they prove.

Low confidence (50-59%): The sole bright spot. 6 of 8 correct (75%) against 54.5% expected yields a +20.5% edge. When the model expresses doubt, it outperforms — suggesting the 57-module engine captures meaningful signal at the lower end of its certainty spectrum, even as it struggles to discriminate among mid-tier confidence fights.

What the D Grade Means

A D grade reflects performance below standard acceptable thresholds for predictive modeling. It does not mean the model is random or useless — the low-confidence positive edge proves otherwise — but it signals that current Featherweight calibration is not reliable enough for actionable decision-making without substantial adjustment.

The sample size of 19 fights is limited. A swing of two or three results would shift accuracy into the mid-60s or below 50%. This uncertainty must be factored into any interpretation. The cross-promotion scope — pulling from multiple organizations with varying roster depths, rule sets, and data quality — adds noise that a promotion-specific model would not face. Do not compare this grade directly to headline records that may use tighter, promotion-specific filters.

Honest assessment: the model's medium-confidence overconfidence is the critical flaw. Until that tier's calibration improves, the D grade is warranted.

Methodology

The Oracle AI operates as a 57-module prediction engine with specialist sub-models voting on outcomes. Predictions are recorded before each fight and verified against official results afterward. Draws and no-contests are excluded from accuracy calculations.

This report covers a cross-promotion Featherweight slice, not a single-promotion dataset. For the complete technical breakdown of module weighting, feature engineering, and validation protocols, see the full methodology documentation.

Report generated from verified prediction log. All metrics reflect actual recorded performance.