Verist

Most AI frameworks help you build faster. Verist helps you change safely. We’ve all been there: You tweak a prompt to handle an edge case, and it silently breaks 5% of your common cases. You don't find out until a user complains. Verist is a deterministic workflow kernel that lets you treat AI decisions like code commits. It captures production baselines so you can run impact analysis before you deploy. The Workflow (Git for AI) 1. Capture: Verist stores every input, output, and trace as an immutable artifact. 2. Tweak: Update your prompt, model, or schema. 3. Verify: Run verist test to recompute against history and see the diff. Why Verist? - 🚫 No Silent Regressions: If an output schema changes or a value drifts, you see it in the diff. - ⏪ Deterministic Replay: Rehydrate any past run exactly as it happened—no "magic" retries hiding failures. - 👩‍💻 Human Overrides Persist: Unlike other frameworks, if a human corrects an AI decision (an "Overlay"), Verist preserves that truth even if you re-run the model. - audit-first: Every step is a ledger entry. "Why did the AI decide this?" is essentially a database query, not a guess. Who is this for? Engineers building production workflows (claims processing, contract review, support triage) where correctness matters more than autonomy. If you are building a "chat with PDF" toy, Verist is overkill. If you are handling money or legal data, Verist is your safety net. — Open Source | TypeScript | Postgres | LLM-Agnostic