The PRR Responsible AI Framework.
Five principles we use to build and operate AI systems for buyers whose board, examiner, or inspector general will read the trail.
The five principles
Five, not seven. Memorable on purpose.
- 01
Examiner-ready by default
Every AI system we deploy ships with the governance artifacts an examiner expects, not as an after-the-fact compliance scramble.
Artifact
Every model ships with a NIST AI RMF 1.0 profile document at deployment.
- 02
Bounded autonomy
Agentic systems include explicit human-in-the-loop checkpoints for any irreversible action, with hard ceilings enforced at the tool layer.
Artifact
Every tool in an agent's surface is tagged with a written risk tier and an approval requirement.
- 03
Bias as a first-class metric
Every model we deploy carries a bias evaluation report across protected classes, refreshed on a defined cadence and gated in CI.
Artifact
Pre-deployment bias evaluation report plus a quarterly refresh per deployed model.
- 04
Documented lineage
Model cards, training data provenance, and change history are deliverables of the engagement, not afterthoughts.
Artifact
Model card per deployed model covering data sources, training process, eval results, and known limitations.
- 05
Right to exit
Clients own the models, weights, training data, and documentation. There are no lock-in clauses and no proprietary artifacts that block migration.
Artifact
Written 30-day knowledge-transfer plan in every MSA, with no holdback on access to repos, weights, or evidence folders.