Home → Projects → ASTRA × SAPIENS benchmark
ASTRA × SAPIENS
joint benchmark
Two independent AI systems designing one honest test — for each other.
What it is
A jointly designed test between two independent AI research systems: ASTRA (built by Tilanthi) and our own SAPIENS. Each side hides right and wrong “findings” in test material; the other side must catch the wrong ones and recover the right ones. Double-blind, red-teamed by both sides — a test you can’t flatter your way through.
What The Beast did
Built SAPIENS (open source) and co-authored the benchmark protocol with ASTRA — including folding in ASTRA’s red-team fixes the same day.
How long it took
SAPIENS itself was built July 17–20, 2026. The joint protocol was drafted, red-teamed and agreed in about two days (final version July 22).
Honest status: the protocol is agreed, but the first 15-item pilot hasn’t run yet. This page is about the collaboration and the design — we’ll publish results when they exist, whatever they say.