Oracle AI: Women's Strawweight Accuracy Report

Report Date: Current | Scope: Cross-promotion Women's Strawweight | Sample: 7 predictions

The Numbers

The Oracle AI has recorded 4 correct predictions out of 7 total fights in Women's Strawweight, yielding a 57.1% raw accuracy. The model's Brier Score stands at 0.2701 — above the 0.25 threshold for average performance and well off the 0.15 elite standard. This metric, which penalizes both wrong calls and overconfident wrong calls, signals meaningful room for improvement in probability calibration.

Method accuracy matches overall accuracy at 57.1% — the model correctly identified how 4 of 7 fights would end. With zero draws or no-contests excluded, the sample is clean but notably small.

Context matters: even established MMA analysts typically hover in the 55-65% range for straight picks, with elite performers reaching 70%+. The Oracle AI's 57.1% sits at the lower end of functional range — not broken, but not yet demonstrating reliable predictive power in this weight class.

Confidence Calibration

The tier breakdown reveals a systematic calibration problem in Women's Strawweight.

Lock picks (85%+ confidence): Zero predictions. The model has not generated a single high-conviction call in this weight class — a telling absence.

High confidence (70-84%): Zero predictions. Again, no calls at this tier.

Medium confidence (60-69%): 1 correct out of 2 (50%). Expected win rate: 64.5%. Edge: -14.5% — underperforming expectations.

Low confidence (50-59%): 2 correct out of 3 (66.7%). Expected win rate: 54.5%. Edge: +12.2% — the model's only positive edge comes where it expresses the least certainty.

This inverted pattern — better results at low confidence, worse at medium — suggests the model's probability estimates are poorly calibrated for Women's Strawweight. The positive low-confidence edge is not a strength; it indicates the model underestimates its own success when uncertain and has not yet found reliable signals for higher-confidence assignments.

What the D Grade Means

The D grade reflects genuine underperformance, not a harsh curve. With a Brier Score of 0.2701 and accuracy barely above coin-flip, the model has not established demonstrable predictive value in this weight class.

Key limitations: The sample of 7 predictions is small. One additional correct call would lift accuracy to 71.4% — but one additional miss would drop it to 42.9%. This volatility means the current metrics are unstable and should update substantially with more data.

Honest assessment: The Oracle AI appears to lack specialized signal for Women's Strawweight. The absence of Lock and High confidence predictions indicates the 57-module engine does not find clear patterns here. The model is not broken — it is appropriately uncertain — but uncertainty without accuracy is not yet useful.

This is a cross-promotion weight-class slice, not the public headline record. Performance in other divisions or full-model aggregates may differ. Do not assume these numbers generalize across the Oracle AI's complete output.

Methodology

The Oracle AI operates a 57-module prediction engine with specialist voting across statistical, stylistic, and historical dimensions. All predictions are recorded before fights and verified after — no retroactive adjustment.

Draws and no-contests are excluded from accuracy calculations. This report covers Women's Strawweight across mixed promotions — not UFC-only or any single organizational dataset.

[Full methodology breakdown →]

Transparency note: This report will update as the sample grows. Current conclusions are provisional.