model-probability

Files (4)

• SKILL.md
• calibration-policy.md · references/
• model-contract.md · references/
• validate_model_output.py · scripts/

SKILL.md (first lines)

# Model probability

The language model is not the probability model. Validate externally produced output; never replace it with an intuitive estimate.

## Workflow

1. Require the canonical `contracts/model-output.schema.json` contract.
2. Match sport, competition, market type, target market, period, line, and settlement definition exactly against `validation_domain`.
3. Confirm model name, immutable version, prediction time, feature cutoff, training cutoff, and feature schema version.
4. For `validated`, require calibrated probability, validation sample size, Brier score, log loss, ECE, validation window, and a non-`none` calibration method.
5. Treat schema-valid `uncalibrated` output as `UNCALIBRATED`; never promote it to validated.
6. Treat unknown provenance, missing validated evidence, domain mismatch, or invalid schema as `BLOCKED`.
7. Check whether critical live facts changed after the feature cutoff.
8. Run `scripts/validate_model_output.py INPUT.json`.
9. Return the supplied probability and validator status without narrative adjustment.

## References

- Read `references/model-contract.md` for required fields, decision states, and exact domain matching.
- Read `references/calibration-policy.md` when assessing model maturity or choosing a Kelly fraction.

## Prohibited behavior

- Do not derive a replacement probability from injuries, form, line movement, public betting, or intuition.
- Do not blend market and model probabilities unless a documented production calibration method explicitly does so.
- Do not treat accuracy or a short winning streak as probability calibration.
- Do not transfer validation evidence between sports, competitions, market types, periods, lines, or settlement rules.

From Jambovisuaalit/doctore, MIT.

More from Jambovisuaalit/doctore