Version Control for AI Decisions

Every AI system produces decisions. None of them treat those decisions as versioned artifacts you can own, branch, replay, and merge. The context that shaped the decision evaporates the moment the request ends. CVCS fixes that.

Core Principle

Identity = Log × Interpreter

Your AI agent's identity doesn't live in model weights. Models are rented from OpenAI, from Anthropic, from whoever. What you own is the accumulated history: every decision, correction, and context change that defines how this system behaves in your domain.

If your agent breaks when you swap GPT-4 for Claude, you built it wrong. All identity work that cannot live in the weights must live in the log. The log is the only artifact you fully own. Everything else is rented.

Same log + new model = upgrade, not replacement
Same log + new branch = safe experiment
Same log + replay = full audit, years later

Observability

"What did it do?"

Logs tell you what happened after it happened.

CVCS

"What did it know?"

Versions the exact cognitive state which prompt, which policy, which knowledge was active so you can reproduce any decision, years later, on any model.

Stop treating decisions as ephemeral

See it in action. Every decision is versioned. Every context change is audited.

1240
commits
8
branches
3891
decisions
98.2%
accuracy

Decision Details

Model
claude-opus
Branch
production
Hash
a3f91b2
Status
success

The log is the system

Most AI deployments have no durable memory of themselves. Which prompt produced this decision? Did behavior change because you updated the policy, or because the model changed? If you left your current vendor tomorrow, would you lose three years of behavioral history?

CVCS is the infrastructure layer that answers these questions. Open source. Postgres-native. No lock-in.

Built with v0