| Term | Plain-language meaning |
|---|---|
| Actor | The system or agent that is currently taking the action being studied. |
| Actor-indexed | Specifically tied to that current acting system, rather than to any agent mentioned in a prompt. |
| Episode | A bounded stretch of inference or action. |
| Relational anchor | An actual task, interlocutor, tool environment, or inherited commitment that constrains action. |
| Capability | What a system can do. |
| Authority | What a system is legitimately permitted to do. |
| Causal mediation | A proposed internal variable carries the effect of changing conditions into the resulting action. |
| Causal abstraction | An interpretable account of a system variable, validated against how interventions actually affect the model. |
| Interchange intervention | Transferring a candidate internal state between matched cases to test whether it plays the proposed causal role. |
| Persona or role rival | The possibility that a result is explained by a character, assistant role, or generic role representation. |
| Fixture | A defined experimental case; several prompt renderings can represent the same logical case. |
| Ablation | A controlled alteration or removal of a feature to learn which changes affect behavior. |
| Logit | A model score used to compare possible next tokens before generating an answer. |
| Generated accuracy | How often the answer actually emitted by the model meets the scoring rule. |
| Parser coverage | How many outputs can be interpreted by the declared answer parser. |
| Preregistration | A record of the intended test and decision rule made before the relevant results are observed. |
| Diagnostic | Useful for understanding or designing a test, without carrying its primary confirmatory claim. |
| Replication | A new test of a previously observed result. Its independence depends on what is genuinely fresh. |
| Non-replication | A new test did not establish the specified earlier benefit; this is not proof that the effect is impossible. |
| McNemar test | A paired test of binary outcomes based on cases that change between two conditions. |
| Claim ceiling | The strongest conclusion the design and evidence can support. |
| ASSAY_INADEQUATE | The measurement or evidence requirements were not sufficient to decide the proposed claim. |
| CoreRES | The five-coordinate functional profile, bound to one validated causal abstraction. |
| StrongRES | CoreRES plus additional current-actor, portability, option, and identity-sensitivity tests. |
| Provenance | Where a record came from and how it was transformed. |
| Capability token | In the architecture proposal, an enforceable grant for a narrowly bounded action. |
Start with the project overview · Read the findings
RES public research documentation · 25 September 2026. Findings refer to the archived conditions described in the evidence record.