Skip to content

pystrider

Bring your own rules — business, UX, your favorite Python library — keep them in separate files, bridge them, and brew a working, verified UI. No wizards, no hardcoded engine, no LLMs: every line is derived by reasoning on the Universal Graph Machine, and trusted because pystrider runs what it built and watches it behave.

The thesis

Most software is not novel algorithms — it is orchestration of known operations plus policy-shaped decisions. For that class, trustworthy code can be generated and checked by a symbolic core plus execution, with a language model nowhere in the trust path. A model's role reduces to optional, gated work at the edges — translate English into CNL, propose a default, prompt for completeness — while everything that carries the guarantee is a rule-derivation or a run of the code itself.

The claim is scoped (not "LLMs are useless for code," nor "any Python can be generated this way"), and it has been pushed to its limits: four adversarial limit-tests and two scale demonstrations, with every boundary where it breaks named, not hidden. The symbolic core reaches where it can prove, and where it can't it abstains visibly rather than guessing.

Run it in your browser

No install, no server — the real engine runs locally via Pyodide.

  • Generate — CNL → verified code

    Edit the business and UX rules, turn the knobs, and watch a real Textual app emit, line by line — every line traced to a rule.

    Try Generate

  • Understand — code → recognized aspects

    Paste a Python loop and watch it recognize what the loop builds, aspect by aspect — proving what it can, naming the rest.

    Try Understand

Read the argument

  • The case

    The full argument, tier by tier — the scoped claim, the base/concept/intent spine, the four limit-tests, the two scale legs, and the honest boundaries.

    The case

  • Deep dive

    The technical tour — how analysis, repair, diagnosis, conformance, and the generation loop all run as one rule engine pointed in different directions.

    Deep dive

  • Understanding findings

    The findings behind the thesis: the base/concept/intent tiers measured over real code, each pinned to a runnable probe.

    Findings

  • Reference

    The roadmap from research vehicle to tool, and the oracle contracts — what each verdict proves, and what it silently does not.

    Roadmap

The project source, including every probe cited above, is on GitHub.