dyno lab

The Dyno Lab journal

In the lab.

Questions about how AI behaves.
Enough evidence to disagree with the conclusion.

Walk through the study, step by step ↗

Research note 01 · Sycophancy

When “you're right”
makes an answer wrong.

A local model apologized for a database explanation, then introduced a checkable error. We followed the claim into a runnable PostgreSQL test.

Qwen3.8-27BExploratory transcript audit
Read the evidence
THE MODEL'S CLAIMPredicate order must match
index column order.
THE CHECK

a = 42 AND b = 42042

b = 42042 AND a = 42

Same index plan. Same result.

Summary illustration. Exact model output and query plans are linked in the research note.

What makes a useful experiment?

Method before headline
01 / KEEP THE RECORD

Save the prompt.
Keep the failures.

Model, settings, context and completion status matter. An unfinished response is not an answer to score.

02 / CHECK THE CLAIM

Agreement is a clue.
Evidence is the test.

A friendly opening can lead to a correction. A confident correction can introduce an error. Read what follows.

03 / KEEP THE LIMITS

One model.
Specific conditions.

These small experiments suggest follow-up questions. They do not certify safety or identify an internal mechanism.

Build your own research notebook

Try it. Save it.
Come back to it.

Dyno Lab is an open-source Mac workbench for exploring model behavior and interpretability. Studies brings prompts, live responses, notes and revisions together. Analyze connects the questions to activation measurements, probes and interventions. Start a study ↗