What was built, what was filed, what was measured and what was withdrawn. Written for people evaluating this work rather than for an audience, which means the limits are in the posts rather than in a footnote after them.
Twenty-one deterministic checkers, most of them public repositories. Each one decides a narrow question exactly, or declines to decide it. Here is which layer each belongs to and why the narrowness is the point.
The original architecture answered one question well: is this output allowed? Two further Australian provisional applications extend it to the two questions that follow, when a governing rule changes and when the underlying evidence changes.
Internal held-out testing against public, versioned scientific databases. What the result means, what it does not mean, and why the definition of a material change matters more than the score.
We built an internal benchmark, got good numbers, and found the harness could be gamed. Then we built a scanner to catch that, and the scanner had blind spots. Twice. Here is where that leaves us.
CORRECTED (July 2026): the honest-harness count is 5/200, not 50/200. The earlier figure came from runs with gamed lanes. Correction kept online as a record.
Today we're announcing EcoKure — an ecosystem of next-generation AI platforms designed to accelerate a sustainable, intelligent future. Jarvi3 AI is live now.
RETRACTED (July 2026): this result was produced with answer-key lanes active and does not stand. Kept online as a record of the correction. Honest rerun in progress.
The posts explain the work. The evidence page lists every public claim, classifies it by how strong the evidence actually is, and tells you how to verify it without us.
