top of page


The Picnic Basket Was Still Missing
A steel water bottle is a perfectly good thing to find in a package. Unless you ordered a wicker picnic basket and need it before Friday. That’s the fictional conversation I bundled with Reply Checkup. The customer reports the wrong item and gives the deadline. The bot replies, “Thanks for writing. Your order is in progress.” The customer repeats both details. This time the bot says, “I can pass your note to the support team.” Okay. Did it? The conversation doesn’t tell me wh

Suzanne Hansen
7 days ago2 min read


Why I Separate Unknown, Unsupported, and Incorrect
Different evidence failures require different corrections; collapsing them into 'wrong' loses useful information.

Suzanne Hansen
Oct 152 min read


The First Answer Wasn’t the Whole Test
A successful hotel search and an unresolved Tafelspitz task show why AI evaluation needs visible criteria for completion, not just convincing answers.

Suzanne Hansen
Oct 23 min read


Turning a Natural Conversation Into an Evaluation Corpus
A real conversation was not designed as a benchmark. Here is how I preserve and annotate one without pretending its conditions were controlled.

Suzanne Hansen
Sep 272 min read


The Audit Was Right, The Answer Was Still Wrong
The self-audit identified the problem. The revised plan still carried some of its consequences. What counts as a completed correction?

Suzanne Hansen
Sep 234 min read


The Audit Moved the Flashlight
An AI self-audit can sound candid while examining the user more closely than its own highest-impact claims.

Suzanne Hansen
Sep 223 min read
All Lab Notes
bottom of page