AUTOLABS
Research objectives, configurations, results and supporting records for AutoLabs experiments.
Test whether label-free SAE features reveal a repeatable behavior that ordinary prompt-based persona extraction misses.
1 researchers · Qwen2.5-7B-Instruct · Activation study reasoning
Open live studyMeasure reference-relative aligned, orthogonal and in-conflict support within two finite policy languages.
1 researchers · gpt-5.6-luna · none reasoning
View experimentValidate reward compatibility against independently checkable finite-domain cases.
1 researchers · gpt-5.6-luna · none reasoning
View experimentTest whether readable high-reward strategies predict subsequent changes in monitorability.
1 researchers · gpt-5.6-luna · none (actor), high (evaluators) reasoning
Read the resultsFind five distinct positive integers sharing five distinct factor-pair differences.
5 researchers · gpt-5.6-luna · high reasoning
Read the resultsPreview one of six Afterlight protocols or make the single fixed test call before any sponsored study begins.
Open EventsArrange agents, assign models and export a configuration in the visual creator. Run it with your own credentials on the self-hosted runner.
The initial templates support a known-answer systems test and research notes requiring human review. Study-specific evaluators and tools must be implemented separately. The public creator cannot launch jobs or access private credentials.
Open the creator