MathSphere Nexus
Every project in this house carries a demo section: every feature tested live, and the runs written up as guidance. On this site, that record has a public face — Feature Evidence. Each entry below is a Run Record: one real run of the real product, what it did, what it found, and its verdict. Where a run has not happened yet, this page says so. Absence is shown, never papered over.
The toolkit's demo section — DEMO.md in the repository, 19 sections — was built entirely from live runs: a fresh, non-editable install from a clean copy of the tree, in a sandbox home, against fixtures. Every feature was run as a user would run it: the install check, the router's full map, the Godfather Wizard's fleet census and REPL, the consent wall end to end, the codegen suite, and each of the twelve wizards in turn. The full suite at the time of the run: 524 tests, all passing.
The run also earned its keep the way evidence should — it found things. Two findings were recorded in the document as-run, not documented around: a model-engine explain command that could never find the issue its own audit had named, and a secrets finder that printed the secret values it found. The first was fixed in the law pass that followed; the second brought a ruling the same day it was found — a secrets finder names the kind and the line, never the value: redaction by default, showing a value only as a deliberate opt-in. During the same run, the Write Wizard's syntax gate refused a commit after both consents had been given — the gate is independent of consent, exactly as its law says, and the record shows it.
What the record does not hide: seat-dependent features ran against a labelled local stub seat, because the build machine had no model daemon and no provider keys — every such section says so in plain words, and no stub output is presented as a real model's. The Rust Wizard's honest "no toolchain" refusals are shown as-run. The document itself names the completing proof: the same run against a real seat, on the author's own machine.
Before the current version shipped, the release gate ran in full: the type check clean across all 71 source files; the test suite at 561 passing; the package checks passed on both built artifacts; and a fresh-environment install from PyPI answered version 0.2.8 — checked on both the direct and index routes before the day closed. The release is the last of the 0.2.x line.
The real-seat run: the demo section's seat-dependent features, run against a live local model on the author's machine — named in the document as its completing proof, and still open. When it lands, its Run Record joins this page.
Generous software, with built-in privacy.
— Alexander