MM3 journeys · 2026-10-16
Five real runs: what we asked, and what the checks said.
Each story is one sentence to a coding agent, a recipe, working code, and a check that reads the code. Scroll it, break it yourself, see the answer flip.
Alignment
Type one sentence. Get the architecture right — in any language.
6languages, same questions
Maturity · security
Keep customers’ data apart. Our first check missed a leak. Here is how it got fixed.
15 / 15planted leaks flagged by the third version
Security
Your agent glued user input into a SQL query. It got caught before merge.
15/15broken copies caught (5 mistakes, 3 languages)
Alignment · Reliability
The payment webhook fired twice. The money moved once.
10 vs 20balance after the same event arrives twice: safe copy vs one-word copy
Over time
Your migration slid backwards. MM3 saw it before you did.
9 of 9yes/no questions still passed on the commit that slid back
Planted tests in tiny apps, not a real project yet. Every number carries a run id on its own page.