Methods, data, limitations and evidence status, published together.
Six open models, ten dilemmas, two framings, three samples per cell. Every raw record is published; the chart is computed from those records.
Forty publicly sourced cases coded on a fixed two axis rubric, every score's rationale published in the data.
Do internal signatures predict how models play? Behavioural fleet measured; latent analysis designed, not yet run.
What evidence would show a model represents evaluation as an internal state? Probing and steering protocol, designed.
Separating persistence caused by a model from persistence caused by the world around it, with falsification criteria.
Per game family: what is measured, the control required, and the claim the evidence never licenses.
How lenders might price credit risk on GPU collateralised debt.
What a functioning forward market for GPU compute would look like.