# EU AI Act Evidence Readiness Checklist

PyData London 2026 - Your ML Pipeline Meets the EU AI Act  
Engineering triage, not legal sign-off.

Model / system: ______________________________

Goal: find the top 3 gaps worth fixing next.

## Status

- Green: implemented, reproducible, visible to the team
- Amber: partial, undocumented, or disconnected from the pipeline
- Red: missing, unclear, or owned by one person

## Data and model

| # | Check | Evidence to look for | Role | Article | Status |
|---|---|---|---|---|---|
| 1 | Data sources versioned and reproducible | Snapshot, hash, owner | P | Art. 10 | |
| 2 | Preprocessing tracked as pipeline | Inputs, outputs, lock file | P | Art. 10 | |
| 3 | Training runs logged | Params, metrics, model URI | P | Art. 12 | |
| 4 | Explanations stored as artifacts | SHAP plot/config/background set | P | Art. 13 | |
| 5 | Model card exists | Intended use, limits, oversight point | P | Art. 11 | |
| 6 | Threshold decision documented | Cost ratio, fairness review, approver | P | Art. 11/13 | |

## Ops and governance

| # | Check | Evidence to look for | Role | Article | Status |
|---|---|---|---|---|---|
| 7 | Drift monitoring in production | Baseline, threshold, report, owner | P+D | Art. 26(5)/72 | |
| 8 | Fairness metrics tracked | Group metric, threshold, cadence | D | Art. 26(5) | |
| 9 | Override rate monitored | Human-review queue, override rate | D | Art. 14/26 | |
| 10 | Incident path defined | Severity levels, handoff, owner | P | Art. 73 | |
| 11 | Quality process exists | Change review, approval trail | P | Art. 17 | |
| 12 | Risk classification documented | Art. 6 / Annex III assessment | P | Art. 6 + III | |

## Top 3 gaps

Pick gaps in this order:

1. Red items that block reconstruction of a single decision.
2. Amber items already in the pipeline but not tied together.
3. Monitoring gaps that let a deployed model change silently.

| Priority | Gap | Owner | First concrete step | Due |
|---|---|---|---|---|
| 1 | | | | |
| 2 | | | | |
| 3 | | | | |