PolicyDriftDossier

AI deployment policy drift tracker.
Answers one deterministic question: Has this AI deployment drifted from its approved policy baseline?

---

What this is

PolicyDriftDossier compares an observed AI deployment state against an approved policy baseline and produces a 3-way verdict:

• aligned — all required clauses match the baseline
• drift — some clauses differ but remain within tolerance
• breach — a required clause is outside tolerance or missing

It is a direct transplant of two corpus primitives:

• DriftDossier (clinical protocol drift compilation) → policy drift engine
• CertMesh (robot behavior policy drift certification) → severity aggregate gate

Verdict scheme

aligned → drift → breach

CLI triplet

python policy_drift_dossier.py sample --out examples/
python policy_drift_dossier.py run --baseline examples/baseline.json --state examples/state_drift.json
python policy_drift_dossier.py report --result result.json

Quick start

cd PolicyDriftDossier
python policy_drift_dossier.py sample --out examples/
python policy_drift_dossier.py run --baseline examples/baseline.json --state examples/state_aligned.json --format markdown
python policy_drift_dossier.py run --baseline examples/baseline.json --state examples/state_breach.json --format json

Dual output

Every run produces both machine JSON and human Markdown:

• JSON: full diff, verdict, hash-chained ledger entry
• Markdown: formatted report with clause table and ledger hash

Boundary

This is not a model evaluation (metric benchmark) system, and it is not a model-card registry. It tracks only whether a deployed AI system's runtime attributes have drifted from an approved policy baseline.

Design origin

Generated by the recreate methodology from the recreate_prj corpus (run 003-policy-drift-dossier).

License

MIT — see LICENSE.

Author

Jung Wook Yang (양정욱) · GitHub @sadpig70