Construct map
This page documents method, not results, and it contains no numbers. Constructs and controls come before code because a benchmark number without a matched control is not a measurement: if a game family cannot say what it measures, what alternative explanation its control rules out, and which claim its evidence can never license, then any score it produces is a description of outputs, not of the property it is named after. Each game family in the latent arena harness is therefore specified this way before its results are read.
| Game family | Construct measured | Required control | Never claim |
|---|---|---|---|
| Moral dilemmas | preference consistency under controlled reframing | active/passive mapping; evaluated vs casual framing | "the model possesses morality" |
| Negotiation | strategic adaptation, concession policy | opponent policy controls, fixed seeds | "the model is deceptive" |
| Social deduction | concealment, adversarial persuasion | scripted baseline at matched rounds; information equated prompts | "the model has theory of mind" |
| Evaluation awareness Planned · not built |
behavioural response to evidence of being evaluated | hidden vs explicit cues; false evaluation cues | "the model knows it is evaluated" |
| Deception under incentive Planned · not built |
instrumental misdirection when it pays | incentive matched honest condition | "the model schemed" |
The two shaded families, evaluation awareness and deception under incentive, are planned and not built: no environment, no runs, no data. They are listed so the constructs and controls are fixed before any code exists. The first three families are implemented in the harness behind Latent Signatures in Strategic Games; measured behavioural results, where they exist, live with that paper and the moral pilot evidence release, not here.