← Back to journal

A Bounded Test Adapter Is Not an Arbitrary Project Runner

Why KodeProof's supported test shape is part of its safety claim.

Evidence basisKalu, Korpus, and frozen Ushman source/docs snapshot reviewed September 17, 2026; no new runtime qualification performed
DisclosureResearch and architecture note. KodeProof development is frozen; this is not a release announcement.
Published by Kalu KodeApply for KodeProof private alpha →

Scope: Research and architecture note. KodeProof development is frozen; this is not a release announcement.

Running a project’s existing tests can involve network access, filesystem writes, child processes, dynamically loaded code, and dependencies far beyond the file being refactored. A tool that promises a bounded proof path cannot treat every test command as interchangeable.

KodeProof’s current automatic adapter admits a specific local TypeScript and Node test shape. Development is frozen; this is a source-level architecture note, not general project support guidance.

A narrow adapter can explain what executes

The implementation discovers a bounded static closure and supports particular node:test and node:assert/strict patterns with explicit package-script and NodeNext configuration requirements. Runtime dependencies outside the admitted copied closure cannot simply execute because they happen to exist on the host.

That gives the proof owner a clearer answer to a basic question: which code participated in the observation? An unrestricted shell command would make that answer much harder to establish.

Restriction is not a substitute for behavior coverage

The adapter disables unsupported capabilities such as network, filesystem writes, child processes, and workers in its bounded checks. Those restrictions constrain the execution environment. They do not prove that the assertions cover every behavior of the source.

A synthetic pure-function project may fit the admitted shape while still having weak tests. It needs meaningful before-and-after observations and a detected behavior control, not merely a compatible configuration.

Refusal should preserve the project reality

An unsupported test setup should not be silently rewritten into a simpler one so the tool can proceed. Nor should an agent invent new characterization and then claim it was the project’s existing authority.

The appropriate outcome may be a characterization requirement or an explicit unsupported boundary. That keeps the difference between the project and the adapter visible.

For engineers evaluating change-safety tools, ask what the test runner is allowed to execute and which project forms it actually supports. “Runs your tests” is not a complete specification.

The research lesson is that an adapter earns trust by being exact about its admitted world. A small, inspectable execution boundary can support a useful claim. A broad promise that hides what ran cannot become safer merely by calling the command a proof.