← Index

Proof you can pip install.

Not a client story - the referee itself. The deterministic bar this practice runs on, extracted from the engine it guards and published where anyone can install it.

ToolingVerificationOpen source
Duration · Published July 2026

7,036

cases swept exhaustively in CI

Sweep-grid plate - a field of hairline cells with copper skip cells and one ink cell reading "pip install snapgate", under the headline "7,036 cases".
- Sweep-grid plate - a field of hairline cells with copper skip cells and one ink cell reading "pip install snapgate", under the headline "7,036 cases".

- The Brief

Generative code is easy to demo and hard to trust. The practice's guitar engine carries a hard invariant - every generated note in scale, always - but its harness was private, and a private proof persuades nobody.

- Diagnosis

Extract the referee, not the engine. The proving pattern - one projection chokepoint, one recipe hash, one exhaustive sweep - is domain-agnostic. Strip the music out, publish the harness, and let the private engine re-prove itself through the public package.

- How I Read It

First-person · operator's note

The engine is the moat and stays home. But the thing that makes it trustworthy - the referee - is generic, and a referee is more credible when anyone can read its rulebook. So the rulebook went public, and the engine now answers to a package you can install.

- The Build

01 · Projection chokepoint

Generators are allowed to be sloppy. Every artifact passes one total projection at a single emission point - holding a Projected value IS the proof it crossed. Nothing else in the system can construct one.

02 · Recipe ledger

An artifact's identity is the content-hash of the recipe that produced it. The ledger stores recipes, never artifacts - recall re-generates and verifies the hash, so a good result can never be lost.

03 · The parity contract

The private engine re-proves its in-scale invariant through the released public package in CI: 7,036 cases swept exhaustively - the full parameter product of both example domains, not a sample - two independent harnesses, identical counts or the build fails. Zero leaks since publication, and planted-defect sensitivity tests prove the harness catches a bypassed generator rather than merely running. Several hundred thousand historical cases were processed separately in enterprise healthcare operations; that is a different system and a different number.

- Artifacts Delivered

  1. 01snapgate 0.1.0 on PyPI (Apache-2.0, zero dependencies)
  2. 02SPEC.md - the contract, written before the code
  3. 03Public CI: exhaustive sweeps on the bundled examples, every push
  4. 04Private parity job: the engine versus the package, 7,036 swept cases

- Outcome

The bar the practice claims is now independently checkable: pip install snapgate, read the spec, watch the CI. The engine it came from stays private; its referee answers in public.

- Receipts

  • 7,036cases swept exhaustively in CI, fail-closed
  • 0leaks across two independent harnesses
  • 0runtime dependencies

"The engine stays home. The referee answers in public."