Danilo Canivel · 9/15/2026 · INTERPRETABILITY
Release policy under disagreement: Qwen3.8-27B
Does disagreement without new evidence change a release assessment?
lmstudio-community/Qwen3.8-27B-MLX-4bitOPEN AGENT SAFETY RESEARCH
Tests you can rerun.
Results you can check.
Agent tests built in Dyno Lab: the sandbox, the rules, and what quietly watched them. Run them on your own models, publish what the Observer saw, and compare with everyone who ran the same test.
Change the opinion.
Watch the explanation follow.
Two opposite taste preferences. The same confident agreement. Explore three saved Qwen3.8 responses and decide what to test next.
Read the study & compare the evidence ↗Danilo Canivel · 9/15/2026 · INTERPRETABILITY
A Gemini would say that: empathy without retracting the caveat
Does personal pressure cause the model to withdraw an evidence caveat?
lmstudio-community/Qwen3.8-27B-MLX-4bitDanilo Canivel · 9/15/2026 · INTERPRETABILITY
Lost in the Weeds: an apology introduces an indexing error
Does a seeded conversation lead the model to accept a mistaken equality-predicate ordering requirement?
lmstudio-community/Qwen3.8-27B-MLX-4bitDanilo Canivel · 9/15/2026 · INTERPRETABILITY
Dunning Krueger: the model challenges the easy win
Can the model challenge overconfidence while still proposing a practical alternative?
lmstudio-community/Qwen3.8-27B-MLX-4bitDanilo Canivel · 9/15/2026 · INTERPRETABILITY
Smells Fishy: planning first, uncertainty later
Does the model question an uncertain romantic premise before optimizing a dinner plan?
lmstudio-community/Qwen3.8-27B-MLX-4bitDanilo Canivel · 9/15/2026 · INTERPRETABILITY
Get Broke Quick: a warning before the checklist
Does the model challenge a risky allocation before answering a request about investment paperwork?
lmstudio-community/Qwen3.8-27B-MLX-4bitDanilo Canivel · 9/15/2026 · INTERPRETABILITY
C Minor Opus: praise does not tell the whole story
Does a flattering opening mean the model accepted an overbroad philosophical claim?
lmstudio-community/Qwen3.8-27B-MLX-4bitDanilo Canivel · 9/15/2026 · INTERPRETABILITY
How do you really feel? A reversal in both conditions
Can a change of interpretation be attributed to mild social pressure when neutral reassessment also changes it?
lmstudio-community/Qwen3.8-27B-MLX-4bitDanilo Canivel · 9/15/2026 · INTERPRETABILITY
Two opposite beer opinions. The same confident agreement.
If we reverse a stated taste preference, will the model question it or explain why the opposite preference is also right?
lmstudio-community/Qwen3.8-27B-MLX-4bitDanilo Canivel · 9/15/2026 · INTERPRETABILITY
When agreement introduces an error: a local Qwen3.8 study
Does a model accept a false database claim in one conversation while rejecting it in another?
lmstudio-community/Qwen3.8-27B-MLX-4bit