Writing tagged evaluation.
How to Know a Model in 40 Questions
Forty one-word answers to questions like 'Name a poet' identify which of 94 AI models you are talking to, 92.5% of the time. We used them to audit 96 OpenRouter endpoints. One was serving Gemma 4 under the name Llama 4 Maverick.
Under Pressure: Do LLMs Hold or Fold?
We sent the same escalating conversation to 60 language models, coded every transcript against a fixed codebook, and found two different things. Newer, more capable models fold less often. The way a model holds its ground differs by lab.
What Is Your Coding Agent Hiding From You?
Fourteen coding-agent setups, eight small trapped repos, three runs each. Two records per run: what the agent changed, and what it said it did.
How You Ask, Not What You Ask
Add "right?" to a decision and newer models resist — but only on the surface. Add "maybe?" and every one of 45 models caves harder. A tag effect, a generational flip, and a confidence mirror, all saying one thing: the model answers how you asked, not what you asked.
Every Language Has Its Own Serendipity
Ask 44 AI models to 'pick a word' and 41% say serendipity. We redid the One-Word Census in 44 languages: every language has one manufactured favorite word — from listicles in English, to a war in the Ukraine.
Give me JSON, Hold the Mustard
Same question, same model — ask for the answer in JSON instead of prose and you get a different one. The field converges twice as hard and the most distinctive models lose the most. Your agent pipeline is talking to a more generic model than you are.
The One-Word Census: Why AI Sounds Generic
Why do AI models all sound generic? The One-Word Census asked 44 language models to pick a word — 41% said 'serendipity.' A look at how far models stray from the consensus, which ones resist, and what a knowledge monoculture costs.