• integrated_82#Q3.5

Model, provider, version identifiers and decoding settings matter because generative behaviour is produced by a run-time configuration, not by the model label alone. The foundations paper is explicit that outputs vary with model updates, contextual differences, and decoding settings, while the evidence review defines configuration provenance as including model provider and version, prompt identifiers, decoding settings, and toolchain configuration. In RAIDT terms, these are not peripheral engineering details; they are part of the evidentiary conditions under which a run should be judged.

This matters across the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Auditability and Traceability require reviewers to reconstruct what happened. Dependability requires repeated-run and sensitivity testing, which is impossible to interpret if temperature or related settings are unknown. Interpretability is also affected, because a fluent answer may reflect stochastic sampling rather than a stable reasoning pattern. Recording identifiers and decoding settings allows reviewers to compare like with like, isolate whether a change came from a provider release or from sampling choices, and judge whether reliance on the output was justified.

It also connects to influence methods as governance interventions. RAIDT argues that behaviour-shaping mechanisms must be logged and reviewed as governed configuration. When these fields are absent, the run-level evidence pack cannot support a defensible score profile, and the anchors 1=missing / 3=partial / 5=audit-ready become difficult to apply consistently because the causal conditions of the run remain obscure.

Take the cybersecurity alert triage scenario. A team uses the same provider and model family on two occasions to summarise evidence and recommend next steps. In the second run, the response becomes more speculative and less stable. Without the provider/version identifiers and decoding settings, reviewers might wrongly blame the analyst, the prompt, or the model family in general.

If the run-level evidence pack shows that the provider deployment changed and that temperature was increased for the second run, the difference becomes interpretable. Reviewers can rerun the task under controlled settings, examine dispersion across repeats, and decide whether the weaker dependability score came from the deployment change, the sampling choice, or both. That is exactly why RAIDT treats configuration capture as governance evidence rather than operational trivia.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10