Built against something real.
Four projects, each answerable to something outside itself: SI constants, the maritime collision rules, database constraints, the filesystem.
- 1191 tests in Lodestar
- 205 physics sanity checks
- 7 layers per topic
- 5 public repositories
Selected work
Each has a case study.
-
Vela Sea
A maritime simulator whose simulation layer runs, and is tested, with no window open.
Playable Case study
-
Veritas
A catalogue of knowledge claims where Postgres enforces the rules, not the application.
Live Case study
-
Anchorfile
A command-line tool that reads a repository and writes the context file an AI agent needs.
Case study
Currently building
VelaScan
Not shipped, and not yet written up. There is nothing to link to, which is the honest state of it; when there is, it will appear above with the others.
About
Everything here has to agree with something outside itself. That is a narrower constraint than it sounds: a simulation can be wrong in a way no test catches unless the test knows what the real answer is, so the projects favour the cases where a constant, a published rule or a database can hold the code to account.
The same constraint decides how they are built. Lodestar recomputes known quantities through the same code paths the simulations use; Vela Sea keeps its simulation layer free of the drawing layer so the whole game can be played by a bot with no window open. Both are the same idea: make the thing checkable, then check it.
The projects are built with an AI coding tool, Claude Code, and the decisions stay with the author. What the AI drafts is held to tests and to outside references, the constants, rules and sources each project answers to, before anything ships.
Contact
Email is the fastest way to get a reply. Everything public is on GitHub.