Governance & audit
An AI prediction can drive a real decision — a loan, a repair, an alert. That only works if you can trust the system that produced it. Pradra Studio builds that trust into the platform rather than leaving it to discipline.
Everything is versioned
Datasets, feature sets, models, prompts, knowledge sources, workflows, and plugins all carry versions. A change creates a new version instead of overwriting the old one. Because nothing is silently mutated, you can always answer "what exactly produced this result?" — and you can always go back.
Nothing skips the gate
Every model faces its capability's acceptance criteria before it can reach production — no exceptions. This holds whether a person clicks "promote", an automated retraining job proposes a candidate, a prompt version is activated, a fine-tune finishes, or an agent is published. There is no side door to production.
A candidate that fails is rejected and production is left untouched — the system working as designed, not a failure.
Full lineage
Because everything links, any production behavior traces cleanly back:
- a classic model → its experiment → the dataset and feature set versions it learned from → the evaluation that let it pass;
- a generative answer → the pinned model, the prompt version, and (with retrieval) the exact document chunks it used;
- an agent's action → the full step-by-step trace of how it decided.
This lineage is what makes an audit — or a post-incident review — possible.
Everything is audited
Sensitive actions are recorded in an audit trail: who did what, and when — role changes, promotions, approvals, consent grants, secret rotations. Credentials and secret values are never written into that trail, so the record is safe to keep and review. Prediction history plays the same role for model behavior, with a configurable retention window.
Safe by default in operation
- One production version at a time per capability — never ambiguity about what is serving.
- Rollback is a pointer switch to a known-good version — instant, never a retrain.
- Honest monitoring that says "not enough evidence" instead of a reassuring zero.
- Human control over automation: retraining approval by default, a cap on automatic promotions, and consequential agent actions always waiting for a signature.
- Hard isolation: data is scoped to your organization and workspaces — see Organizations & workspaces.
The short version
Versioned, gated, traceable, audited. If a prediction reached production in Pradra Studio, you can show how — and undo it if you must.