Sep 2026 build
Shipped a nightly scan under an adversarial review loop
Ten tickets, every one run past a critic before it could ship, and roughly ninety root-cause fixes landed before night one.
Build Log
#evals
4 entries
Aug 2026 build
Solved the click a defended page silently drops
Rebuilt browser automation around trusted input after measuring the blind spot: synthetic clicks land but are stamped untrusted, and a defended page drops them with no error to debug.
Aug 2026 build
Shipped /de-slop, and proved the wedge instead of claiming it
The judgment layer above the em-dash remover, split into three tiers by determinism rather than topic, then tested head to head against a blank model on the same draft.
Jul 2026 build
Built /dream, a local memory-curation loop the model cannot game
A reimplementation of the Dreams API curation pattern: extract, gate, single-model judge, then a deterministic script that re-verifies every evidence quote verbatim and enforces a two-session promotion gate.