The question
A reasoning system can be asked to solve a problem of its own, or to solve someone else's. The task can be identical in every formal respect and still be approached differently. This programme studies that difference in language models: when the framing of a task changes from self-directed to externally given, what changes in the behaviour — and how much of the observed change is real rather than an artefact of measurement.
The second half of that sentence is where most of the current work goes. A difference that does not survive a rerun is not a finding, and a great deal of published behavioural work on language models does not report whether it would survive one.
How the work is organised
Paired conditions
Every observation comes from a matched pair: the same underlying item presented under a self-directed and an externally framed condition. Comparisons are made within the pair, not across runs.
Separated sources of variation
Construction of the material, sampling by the model, and the execution environment are varied independently, so that a difference can be attributed rather than merely observed.
Pre-registered analysis
Decision rules are written before the data is seen, and reviewed adversarially. Analyses that were chosen after looking at results are reported as exploratory, never as confirmation.
Current status
The programme is in an instrumentation phase. Current work is about making the measurements trustworthy: separating sources of variance, establishing what reproduces and what does not, and retiring earlier interpretations that later runs failed to confirm.
What is published, and what is not
Methods and limitations are documented in full. Results are published only after they have been reproduced and reviewed; an interpretation that a later run failed to confirm is withdrawn rather than softened. Where a claim has been withdrawn, the documentation says so and says why.
What is being measured · Methods and limitations · About the programme