Determinism
Identical validated inputs, policy versions, and code versions must produce identical score artifacts.
Methodology
Learning without hindsight contamination.
ACI-IPS is designed to improve through structured prediction-to-outcome comparison while preventing data leakage, backfilling, and single-case overfitting.
Master Spec v1.4Schema lockedFixture validated
Methodological controls
The model must be inspectable before it is persuasive.
Explainability
Every score exposes signal contribution, confidence, limitations, and relevant policy context.
Forward calibration
Prediction snapshots are frozen before outcomes and later compared through a model-delta report.
Bounded improvement
Permitted changes include controlled weight tuning, new pre-cutoff signals, and approved rubric refinement.
Calibration cycle
Prediction snapshot → outcome comparison → model delta.
- 01Freeze the evidence cutoff, inputs, policy bundle, and score artifact.
- 02Wait for the defined forward outcome horizon without altering the snapshot.
- 03Compare predicted rank, confidence, and scenario range with the realized outcome.
- 04Classify misses by data, signal, policy, calibration, or decision-translation cause.
- 05Propose bounded changes and validate against fixtures and unrelated cases.
Prohibited calibration behavior
Do not inject post-cutoff facts into a historical prediction snapshot.
Do not alter the outcome definition after reviewing performance.
Do not tune the model solely to make one famous case appear correct.
Do not publish a score without the policy, data, and code versions required to reproduce it.