Model behaviour drifts quietly. PRISMFORG makes it measurable: versioned evaluation sets, regression gates in CI, and runtime guardrails that log what they blocked and why. Source and issues: github.com/prismforg/prismforg.

Versioned cases with assertions, automated scorers and human review.

Block quality regressions in CI with configurable thresholds and per-run diffs.

Screen prompts and responses for injection, leakage and off-scope topics.