PUBLISHER · MODEL LAB

Bozzle learns from receipts.

Calibration, prop-type performance and context research from frozen predictions that were graded after the game. NFL, MLB and NBA are never pooled into one calibration score. The lab can recommend what deserves investigation, but it does not silently rewrite model weights.

LEARNING RULE NO TINY-SAMPLE TUNING 30 settled picks before adjustment signals. More is better.

Publisher sign-in required. Model Lab is an owner workflow, not part of the public betting interface.

OFFICIAL MODEL RECORD

Learning status.

Loading Model Ledger…

Sport

Settled0

Hit Rate—

Avg Model Confidence—

Calibration Gap—

Brier Score—

Sample StrengthEARLY

Independent Clusters0

Cluster Hit Rate—

CALIBRATION

Does 70% behave like 70%?

The most important confidence test: predicted probability versus actual hit rate.

Confidence	Settled	Model Avg	Actual	Gap

PROP TYPE PERFORMANCE

Where is Bozzle strongest?

Prop	N	W-L	Actual	Model

CONTEXT COMPLETENESS

Does richer context help?

Context	N	Actual	Model	Gap

ENGINE VERSION PERFORMANCE

Did the new model actually improve?

Version	N	W-L	Actual	Model

SAMPLE INDEPENDENCE

Raw signals versus correlation clusters.

MLB props from the same player/game can be strongly related. The cluster view keeps those receipts while preventing them from masquerading as fully independent evidence; for the independence view, the highest-confidence settled signal represents each player/game correlation cluster.

View	N	W-L	Actual	Model

CONTEXT FACTOR AUDIT

Evidence before coefficients.

This compares outcomes when a structured context factor was available versus unavailable. It is exploratory and can be confounded; it is not automatic causation.

Factor	Available	Hit Rate	Missing	Hit Rate	Difference

BOZZLE'S NEXT QUESTIONS

What deserves investigation?

These are research prompts from the ledger, not automatic betting advice or silent model changes.

Bozzle does not auto-tune itself.

A small sample can make random noise look like a powerful feature. Bozzle records the evidence first, surfaces repeatable patterns, and waits for enough settled predictions before we consider changing coefficients.