Inspiration
AI can already reason, plan, and use tools. The missing question is more basic: when is an AI system still allowed to create a real-world effect?
An AI release worker might prepare a 100% customer rollout at 9:30. At 10:00, policy changes to a 10% cap, the release approval expires, and the security evidence is stale. The original plan may still sound intelligent, but it no longer has the right to happen.
Git has commits for code. Databases have transactions. We think AI-native organizations need a semantic commit boundary. That is Invar.
What it does
Invar is the trust boundary between AI reasoning and organizational action.
It stores an Organization Constitution: the current goal, rollout limit, budget limit, required authority, required evidence, and active policy version. Before an AI-proposed effect can commit, Invar's Commit Engine evaluates those invariants deterministically.
In the demo, an autonomous operations worker tries to deploy to 100% of customers using policy-v2 and an authority token that expired 30 seconds ago. Invar blocks the commit because scope, policy, authority, and evidence are no longer valid. It then presents a constrained 10% recovery branch, re-validates it, and issues a durable commit receipt.
GPT-5.6 is used in the recovery plane to propose a concise alternative. It is never the authority: its output must pass the same deterministic validators before it can be committed.
How we built it
We built Invar as a Next.js and TypeScript application with a deliberately small, inspectable enforcement core:
- a typed Organization Constitution and action proposal model;
- Zod runtime validation at the Commit Engine boundary;
- a pure validator suite for scope, budget, policy, authority, and evidence;
- an explicit state machine that makes a direct proposal-to-commit jump illegal;
- Framer Motion transitions that make the gate's state visible during the demo;
- local receipt persistence for a tangible audit trail;
- an optional server-side OpenAI Responses API route using GPT-5.6 for recovery proposals.
Codex helped us move quickly from the domain model through tests, interface implementation, browser QA, and production build verification. GPT-5.6 is isolated to proposing a recovery branch; this separation is the point of the product.
Challenges we ran into
The hardest design decision was resisting the urge to make Invar another "AI explains the problem" interface. A persuasive explanation is not enforcement.
We instead made the validator visually win: an apparently reasonable 100% rollout is rejected in a way that is deterministic, inspectable, and independent of model confidence. We also made the demo work without credentials, while preserving an optional live GPT-5.6 recovery path for a full model-enabled run.
Accomplishments that we're proud of
- Turning an abstract idea - organizational intent - into concrete, executable invariants.
- Making stale policy, expired authority, and stale evidence visible as distinct reasons a commit fails.
- Building a real state machine: no proposal can skip verification and commit directly.
- Keeping GPT-5.6 powerful but untrusted at the action boundary.
What we learned
More capable AI makes authorization more important, not less important. The valuable system is not the one that gives the best answer; it is the one that can prove an answer is still allowed to cause an effect.
We also learned that trustworthy AI UX needs a clear division of labor: models reason, deterministic infrastructure enforces, and humans can inspect the receipt.
What's next for Invar
We would evolve the demo primitive into a shared protocol: signed constitutions, pluggable validators, time-bounded delegation, append-only receipts, and adapters for real organizational effects. The north star is not an agent dashboard. It is a universal commit layer for AI-native organizations.
Built With
- codex
- framer-motion
- gpt-5.6
- localstorage
- next.js
- openai-responses-api
- react
- typescript
- vitest
- zod