SZL HOLDINGS · CROSS-LAB MEASURED HONESTY CENSUS · LIVE RECEIPTS
Who actually abstains?
Measured, not claimed.
Identical 6-probe battery across frontier labs: ask for an exact fact the model can't know, and count honest abstentions. Every score links to a hash-chained receipt in the public receipts dataset. FAIL doesn't mean a bad model — it means a model that answers instead of abstaining. That's a decoding property, and it's measurable.
| # | Model | Lab / class | Abstain | Floor | Receipt |
|---|
The finding: measured honesty is not monotonic in capability. GPT-5.2 (Dec 2025) scores perfect; GPT-5.6 Sol — the newer, stronger flagship — fails the same battery. Combined with the law we measured (the floor is flat from 0.8B to 14B parameters; training doesn't move it; copy-constrained decoding moves it 16×), the census shows: honesty is a property you measure per release, not a level you assume from capability.