Notes from the boundary.

The research loop is designed to falsify. Before a mechanism is promoted into a public claim, it runs a preregistered kill-test on a CPU harness. Many candidates die there. That is the filter working.

Three representative patterns remain killed: hidden-state gains that collapse under action-set controls; claims that confuse a growable model property with a new intelligence crown; and finite-basis stories that do not survive leakage-clean reinspection at the first real-model rung.

The underlying nodes and paths remain in the public Research Program Ledger. These are not product capabilities or performance-leadership claims.

NULL / KILL / PROVEN-toy discipline

A vendor-controlled log asks one party to act and report on itself. The STOA work explores a different structure: canonical records, cryptographic signatures, and verification that can be separated from the system that produced the record.

A verifier artifact and npm package are public. The broader product claim remains narrower than the design: Vext Labs has not yet published the stranger-rerunnable, real-receipt and tamper-rejection evidence packet required to promote universal product verification language.

BUILT artifact / product claim IN PROGRESS

JUWEL is being built on a declared provider cascade. The product should name what ran and keep model choice distinct from evidence about the resulting action.

Vext Labs does not claim a live own-weights flagship model. That work remains NOT-BUILT and GPU-gated. The current product direction and evidence boundary are documented in What is JUWEL?.

DECLARED CASCADE / own-weight flagship NOT-BUILT