A FIELD GUIDE TO INCONVENIENT EVIDENCE
A little wonder.A lot of proof.
We don’t ask you to trust the number. We hand you the way to check it.
THE JOURNAL / CONTENTS
Pick a chapter.
ASSAY-001: Jev, checked
An independent, pre-registered test of TypeSafe’s launch claims. Calibrated on one corpus of two; zero type errors.
A stealth model, checked
A claim from X, tested the same day. The model passed. Our ruler had eight defects, four found by us and four by a second reviewer. All published, with the words.
Shuffle the options
A probe: reorder Jev’s options and 5.3% of answers change. Same order, 0%. Every flip below 0.9.
The register
Fourteen independent measurements of Jev’s calibration, collected from their own files. Not re-scored. Not ranked.
ASSAY-002: Snap, checked
The same pre-registered test on SubQ’s Snap. Inside the 0.05 line on Banking77, outside it on CLINC150, underconfident. Zero type errors.
OPEN SOURCE
Six of our own tools
MAP THE SOUL, COSMIC COMMONS, the BIFROST extension, CLAW, WING and QUILL. Each one lives in a public repository under an open license.
The shelf ↗ALSO FROM THE WORKSHOP
LEVEL
A free learning commons for credential study, money skills, responsible AI, and Texas career paths. Free for every learner. No paywall, ever.
jllevel.com ↗