Build software that does not depend on one implementation

Most applications give every user identical code. One hidden bug can therefore affect every installation. Polyform uses coding agents to create several implementations from one complete behavioral specification, then gives installations different combinations of those implementations.

The workflow

1. Specify the complete behavior. Describe the application, every public function in the design, its interface, rules, errors, and concrete examples.
2. Generate the code. Run polyform generate --count 5. Codex creates shared conformance tests and five implementation candidates for every function.
3. Ship one program. The program contains all accepted implementations. Each installation receives a different combination to use.
4. Compare real outcomes. Applications report success or failure and the implementation IDs involved, without sending user content.
5. Remove risky code. The dashboard shows implementations that appear unusually often in failures. Quarantine one to stop distributing it and send replacements.

A concrete example

Imagine a ZIP extractor with five filename decoders and five decompression implementations. One installation might use decoder 2 with decompressor 4. Another might use decoder 5 with decompressor 1. All must follow the same public specification and pass the same tests.

If decoder 2 contains a rare bug, failures should occur more often where decoder 2 is running. The dashboard links that result to its public source. You can inspect it, quarantine it, and have those installations use another already-shipped decoder.

Why multiple implementations help

AI makes it practical to create and maintain more code than a human team could write by hand. That extra code creates diversity: a flaw in one approach may not exist in another. As AI-assisted vulnerability discovery makes attacks faster, an independently written alternative may already be available before you have time to design and deploy a traditional patch.