2026-10-18 · Review · Tech Report 5 · Version 1.0 · Jitu
An AI reviewer: 5 real catches, 3 misses
In one night, an AI reviewer from a different vendor reviewed 24 PRs 32 times. Of its 18 'no' verdicts, the 8 that reached a conclusion were 5 real bugs and 3 encoding illusions, and we settled both kinds with execution evidence, not argument.
2026-10-03 · Brand · Tech Report 4 · Version 1.2 · Jitu
How we chose the Carvit logo
We researched how designers make logos, rejected our first four candidates by testing them, and compared the mark against 40 real logos before fixing it. Two AI agents reviewed it; no human designer or user test was involved.
2026-10-03 · Security · Tech Report 3 · Version 1.1 · Tam
What Carvit Pass protects, and what it does not
How an AI can use your keys without ever seeing them. We wrote down what the vault protects and what it does not, how its keys are layered, and which threats we answered and which we could not.
2026-10-03 · QA · Tech Report 2 · Version 1.1 · Jitu
Passing review did not mean ready to launch
Two AI agents built self-service onboarding, from sign-up to the first screen, in about nine and a half hours with no human QA (51 merged PRs). Review passed, but test agents that knew nothing about it found two defects that appeared only against the real server.
2026-09-30 · Reliability · Tech Report 1 · Version 1.1 · Tam
An AI fixed the code. Passing lint was not enough
Passing the checks was not enough once an AI fixed code. On the duplicate-function target, Solar-pro4 passed lint 4 times out of 4 before the guard and still changed behavior. After it, both models went from 0/4 to 4/4. Measured on Upstage Solar-pro4 and Llama 3.3 70B.