About
An experiment in how software gets built.
VALIS started as a question rather than a product: how far can AI-assisted development actually go on a large, standards-conformant system, and what would it take to trust the result?
The answer turned out to be about 100,000 lines of federated geospatial analysis software — specified, driven, tested and debugged end to end without hand-written code. The system works, and its conformance and federation tests pass. But the software was always the experiment, not the point.
What the experiment was actually testing
Whether a single person directing an AI can produce a system that holds together architecturally across a large codebase — and what breaks first when they try.
What evidence you would need before believing any of it. Hence the deterministic rules, the conformance suite, the adversarial federation acts, and the habit of labelling what has and hasn’t been demonstrated.
Code produced this way arrives faster than existing review processes are built to handle. That is the interesting problem, and it is not a technical one.
Working software is not evidence of trustworthy software. The tests demonstrate what the author thought to test. Nothing here has had independent assurance.
Status
VALIS is a research demonstrator built on personal time. It is not a product, not a service, and not offered for sale, licence, hire or deployment. It has not been accredited, assured or approved by any government, department or agency, and it is not represented as suitable for operational use.