Corridor

Corridor is a local, privacy-first proof of concept for turning minimized synthetic observations into a reviewed automation hypothesis.

What it does

Corridor gates observations behind a scoped authorization receipt, minimizes fields, groups only exact compatible repeats, applies deterministic policy, binds human review to the evidence and candidate version, and runs eight no-I/O adverse cases before producing a typed lifecycle timeline. The current PoC cannot execute external actions.

How it was built

The provider-independent core is written in Rust. Axum exposes the local API, a Rust CLI supports explicit ingestion and observation, and a Next.js control plane presents evidence, candidates, policy, approval, replay, and timeline. The optional OpenAI adapter uses the Responses API with Structured Outputs and targets GPT-5.6. Codex accelerated the cross-stack implementation, tests, fixture, launch paths, and documentation while the privacy and autonomy boundaries stayed explicit.

What makes it different

Corridor is designed around auditable automation review, not employee surveillance. Raw keystrokes, screenshots, messages, document bodies, health data, and secrets are outside the contract. Model output stays advisory until deterministic checks pass, and the demo is synthetic, in-memory, loopback-only, and incapable of external action.

Judge path

The repository includes one-command setup, sample synthetic data, an exact demo script, and a full verification target. make check passes 42 Rust tests, 19 web tests, strict linting and type checking, and two Next.js production builds.

The meaningful Build Week extension is bounded by commits b6df258 and bba2f12, authored July 19 after submissions opened. The latter records the verified judge pack, unified Bun workspace, hardened launcher, and explicit runtime boundary.

Built With

• axum
• bun
• codex
• gpt-5.6
• next.js
• openai-responses-api
• rust
• structured-outputs
• typescript