OMENZResearch
SCRIBE is OMENZ's public record: what we tested, what we expected, what actually happened, and what we changed because of it. Failures included.
Changed? failed its first test, then shipped
SCRIBE 002 · Nov 1, 2026 · finding + change
Changed? tells an agent whether a page really changed since its last visit, or just got a new timestamp. The first version failed our own frozen test: it called a JavaScript-only page "unchanged". That was 1 miss, 3.4%, over our 2% bar.
We didn't lower the bar. We fixed it: pages that need a browser now come back UNKNOWN, never UNCHANGED. Version 2 then took a fresh test on 157 pages it had never seen, rechecked 4 days later: 0 of 43 real changes missed, 93.4% accuracy. In the wild it caught two small businesses changing their opening hours. Known false alarms (live counters, rolling dates, captchas, social feeds) are listed on the tool page.
Agents decide to verify before they read
SCRIBE 001 · Oct 27–30, 2026 · finding + change (Confío)
We tried to make fact-checking cheaper for AI agents by handing them verified evidence: exact quotes, sources, timestamps. They mostly ignored it and checked the original website anyway, often before even reading what we gave them.
So we measured when they decide to verify. Answer: at planning time, from how the tool describes itself. We moved the claim there. Compared with agents given no evidence, re-checks on simple lookups went from 100% to 0%, and high-stakes tasks used 54–67% fewer tokens, with every answer correct. A follow-up showed the flip side: when a trustworthy-sounding source was wrong, agents that didn't spot-check repeated the error, and every agent that spot-checked caught it. Claims have to be true and checkable.
Small, simulated tests (8 businesses, 2 AI models).
Open question: do lenders check a token's state?
OKENZ · open research question · Nov 1, 2026
In Ankr aBNBc → Helio (2022) and the eBTC mint → Curvance (2026), a token's state changed and a lending market kept accepting it. Before you accept a token as collateral or payment, do you check the token itself (supply jumps, contract upgrades, admin changes, pauses), or rely on oracles and caps? If a token-side change ever cost you money or a pause, we'd like to hear about it.