[SCENARIO] Readiness Calibration Loop

[shchukins]

Problem

The product must learn whether deterministic readiness and recommendation outputs match user outcomes. This starts as reproducible analytics, not adaptive ML or hidden model mutation.

User-facing flow

1. System snapshots readiness, recommendation, model version, data quality, and relevant features at decision time.
2. User completes training and provides feedback.
3. Next-day recovery feedback is linked to the previous training day.
4. Analytics join prediction context with actual outcomes.
5. Product surfaces mismatch cases and calibration summaries.
6. Future threshold or rule changes are explicit, reviewed, versioned, and documented.

Definition of done

• Calibration records are reproducible from persisted snapshots and feedback.
• Analytics identify overestimation, underestimation, missing-data uncertainty, and recommendation mismatch cases.
• Dataset export exists for offline analysis without adding online ML behavior.
• Calibration does not mutate historical readiness or silently alter deterministic formulas.
• Model/version metadata is sufficient to compare outputs across changes.

Linked existing issues

• [Feature] Readiness calibration with subjective outcomes #54
• [EPIC] Adaptive Feedback Loop #73
• [Feature] Prediction vs outcome calibration analytics #77
• [Feature] Persist feature snapshots for future ML datasets #78
• [Feature] Dataset export pipeline for calibration and ML #79
• [Feature] Subjective feedback analytics #80

Related future research

• [Research] Automated model experimentation sandbox (autoresearch-like loop) #36 belongs to the research sandbox and should remain outside production calibration until feedback, snapshots, analytics, and export are stable.

Migration note

Created from the accepted end-to-end epic proposal in docs/product/END_TO_END_EPICS_PROPOSAL.md. Existing adaptive-feedback planning is preserved and linked here without mass renames or closures.