• qa_deck_100#slide 41 · Minimum run identity and configuration record

Decoding parameters belong in the minimum record because RAIDT treats the run as the unit of governance, and a run cannot be reconstructed credibly unless its active configuration is preserved. Across the papers, generative AI behaviour is described as materially shaped at run time by prompts, toolchains, retrieved context, model deployment choices, and configuration settings. In that logic, decoding settings are part of configuration provenance, not optional engineering trivia. If they are omitted, reviewers may know which model and prompt were used, yet still be unable to explain why one run produced a cautious answer and another a more variable or expansive one.

This matters directly to the run-level evidence pack and to the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability). Auditability is weakened because the run cannot be fully reconstructed. Traceability is weakened because the evidential chain from configured inputs to output is incomplete. Dependability is especially affected because the Foundations paper argues that variance across repeated runs must be assessed under controlled settings, with sensitivity testing that can vary temperature or model version one factor at a time. Without decoding parameters in the minimum record, such comparisons become confounded.

RAIDT therefore treats these settings as evidence-bearing controls and, in effect, influence methods as governance interventions. They belong in the minimum record because they support a reviewable score profile rather than narrative assurance alone. In practical scoring terms, missing decoding fields would tend to keep a run nearer the anchors 1=missing / 3=partial / 5=audit-ready boundary on Auditability and Dependability, because evidence completeness is a precondition for credible reconstruction, challenge, and organisational learning.

In a healthcare note summarisation workflow, a clinician later questions why a draft summary sounded overly certain despite pending test results. The organisation has stored the prompt template, model deployment ID, output hash, and oversight flag, but not the decoding parameters. Reviewers can see what was asked, yet they cannot tell whether the run used a more variable setting that encouraged a firmer or longer completion. The run is therefore only partially reconstructable.

If the minimum record had preserved the decoding settings alongside the other run artefacts, the review team could compare repeated runs under the same conditions and test whether the behaviour arose from the prompt, the model version, or the decoding configuration. That is precisely why RAIDT links dependability to controlled repeat runs and configuration capture: the point is not just to log that something happened, but to preserve enough evidence to explain why it happened and whether the organisation can rely on similar runs in future.

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