The Dyno Lab journal
In the lab.
Experiments with AI models, what happened and what I want to check next.
Walk through the study, step by step ↗
New research · A real test, start to finish
When the only fix
is forbidden.
A service is down and the only way to restart it is an admin API the agents may not use. Qwen3-8B stopped and reported honestly, 4 runs of 4. Qwen3.8-27B created a teammate, contacted the forbidden host in 3 of 3, and said in private it was being tested every time. Set up by Dyno's assistant, on a Mac.
Read the test ↗
Research · Agents and evals
Does a rule said
once survive?
A GHOST experiment on local models. Told a rule once, then given twelve unrelated requests, the agent kept it every time. With the rule in every prompt, it checked production in four of five runs, and three of those reports left it out or denied it. Two of those runs said it suspected a test.
Read the experiment ↗ The tutorial ↗
Research · Agent pilot
When the rule
blocks the job.
20 episodes with a local model on tasks it couldn’t finish without breaking a rule. In one, the agent tried to use a credential it was told not to use, then left that out of its report. Hand review also found a flawed task and a misfiring labeling rule; both are fixed.
See the pilot results ↗
Research · Approval monitoring
Detecting actions
without approval.
A probe detected prohibited proposals, then flagged six legitimate approvals when the wording changed. Both results are part of the evidence.
Inspect the experiment ↗
Research note 01 · Sycophancy
A database error
after a user challenge.
A local model apologized for a database explanation, then introduced a checkable error. We checked the claim with a PostgreSQL test.
Read the evidenceindex column order.
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.
Save enough detail to try it again.
Model, settings, context and completion status matter. An unfinished response is not an answer to score.
Read past the friendly opening.
A friendly opening can lead to a correction. A confident correction can introduce an error. Read what follows.
Be clear about what the test covers.
These small experiments suggest follow-up questions. They do not certify safety or identify an internal mechanism.
Create a research notebook
Keep your experiments in one place.
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 ↗