S10.04 - 6_configurations

S10.04 ? 6 configurations

flowchart LR
    A[Background problem:
organisations compare outputs,
not governance effects of configuration] --> B[RAIDT:
run-level evidence framework]
    B --> C[[6 configurations:
comparative intervention set]]
    C1[Baseline prompting] --> C
    C2[Structured prompting] --> C
    C3[RAG] --> C
    C4[PEFT / LoRA] --> C
    C5[RLHF-type controls] --> C
    C6[Stacked influence] --> C
    H[Healthcare] --> C
    I[Finance] --> C
    J[Public services] --> C
    K[Cybersecurity] --> C
    L[Supply chain] --> C
    C --> D[Evidence pack]
    C --> E[RAIDT score profile]
    C --> M[Governance move:
evidence over assertion,
reviewability, contestability,
audit readiness]
    D --> F[Reviewer reconstruction]
    E --> G[Governance readiness]
    F --> N[Organisational learning]
    G --> N

? Star S10 - Empirical Programme, Domains and Sector Playbooks

Star context: Places RAIDT's empirical programme inside a comparative design space by showing how the framework is tested across six distinct configuration conditions rather than being asserted from a single model setup.

Academic picture

Definition / background

In this item, the six configurations are the main comparative conditions through which RAIDT examines how a generative AI run is shaped by different forms of influence and control. They include baseline prompting, structured prompting, retrieval-augmented generation (RAG), parameter-efficient fine-tuning such as PEFT/LoRA, RLHF-type controls, and stacked influence, where several control layers are combined. The concept comes from experimental comparison: if the task remains broadly stable while the configuration changes, the analyst can examine what difference the configuration makes.

Within RAIDT, a configuration is not merely a technical setting. It is a governance-relevant arrangement of prompts, retrieval sources, model adaptations, feedback constraints, and control layers that affects how a run behaves and how that behaviour can later be evidenced. This matters because RAIDT treats the run as the unit of governance. If the run is configured differently, it is not simply the same run with a cosmetic variation; it is a different evidential condition with different review and assurance implications.

This item therefore belongs centrally inside RAIDT's empirical programme. The framework is not intended to rest on abstract principles alone. It is intended to show, through structured comparison, how governance readiness changes when different influence methods are applied. The six configurations make that comparison visible and operational. They help explain why a run-level evidence pack and a five-pillar score profile are more meaningful when they are produced across comparable configuration conditions rather than in isolation.

The idea also helps distinguish RAIDT from generic capability benchmarking. A benchmark may tell an organisation whether a model performs well on a task. S10.04 asks a different question: under which configuration is that performance most responsible, auditable, interpretable, dependable, and traceable? That shift is what makes the item important for governance rather than only optimisation.

Why this concept matters

Many organisations evaluate generative AI systems as if governance were separate from configuration, when in practice the configuration often determines what can be checked, justified, reproduced, and challenged later. If configuration differences are ignored, quality improvements may be mistaken for governance improvements, and governance failures may be wrongly attributed to the model alone rather than to the way the run was assembled.

The six configurations solve that problem by turning influence methods into explicit comparative conditions. This avoids confusion between model capability, deployment design, and assurance quality. It also reduces the risk of making broad governance claims from a single setup that happens to perform well in one domain but is poorly evidenced, weakly traceable, or difficult to reconstruct under review.

For organisations using GenAI in real work, this matters because procurement, policy, and internal oversight often ask whether controls are proportionate and effective. RAIDT can answer that question more convincingly when it can show how governance readiness changes across baseline, structured, retrieval-based, tuned, feedback-shaped, and stacked configurations. That is a move from principle to operational governance.

Key idea: The six configurations matter because they make configuration itself visible as an empirical governance variable rather than leaving it hidden behind output quality claims.

What this item enables

Systematic comparison between minimally influenced and more heavily governed GenAI runs.
Clear separation of task performance questions from governance readiness questions.
Identification of which configuration features improve or weaken specific RAIDT pillars.
Construction of evidence packs that explain not only what the model produced, but under which influence conditions it produced it.
Development of sector playbooks that recommend controls appropriate to domain risk rather than one universal setup.
More defensible claims in supervision, review, audit, and policy discussions about why one configuration should be preferred over another.

Practical example / likely audience question

Audience question

Why compare configurations?

Answer

The concern behind this question is usually that configuration sounds like an engineering detail, while governance is assumed to sit in policy, oversight committees, or post hoc review. RAIDT rejects that separation. The direct answer is that configuration determines how governable the run is, not just how fluent or accurate the output appears.

Consider a case where the same organisational task is run under two conditions. In the first, a user enters a free-form prompt into a general model. In the second, the task is carried out with a structured prompt, controlled retrieval sources, and a documented review step. Even if both outputs appear acceptable, the second configuration gives a stronger basis for explanation, reconstruction, challenge, and accountability. That means the organisation has changed governance readiness, not merely surface quality.

RAIDT handles this better than a generic AI governance approach because it does not stop at saying that controls should exist. It compares how different controls behave at run level and records the evidence of those differences. That makes the answer empirically defensible: the organisation can show why a chosen configuration is preferable, under what conditions, and with what trade-offs.

Practical example in RAIDT terms

A hospital operations team uses a generative AI system to draft discharge-planning summaries for clinicians. In a baseline prompting configuration, the model receives a short free-text instruction and produces a plausible summary, but the source basis for its recommendations is unclear. In a structured prompting plus RAG configuration, the same task is run with a discharge-summary template, retrieval from approved hospital guidance, and a reviewer sign-off step.

The run-level issue is not only whether the text reads well. It is whether the organisation can show what instructions were used, what knowledge sources informed the output, whether local guidance was current, and how a reviewer checked the result before use. The evidence needed includes the prompt template version, model identifier, retrieval corpus version and timestamps, access logs for the guidance set, reviewer notes, and the rationale for the resulting RAIDT pillar scores.

In this case, Responsibility is improved because role allocation and sign-off are clearer. Auditability and Traceability improve because the retrieved sources and prompt structure can be reconstructed. Interpretability improves because the output follows a known template and source basis. Dependability improves if repeated runs show stable behaviour under the structured configuration. Governance readiness is therefore increased not because the model became magically safe, but because the configuration made the run more governable and reviewable.

Detailed link to RAIDT

6 configurations links to RAIDT in four ways.

First, it turns RAIDT from a static governance claim into an empirical comparison framework for different intervention conditions.
Second, it fits RAIDT's core unit of analysis because each configuration creates a distinct run condition that can be documented and reviewed at run level.
Third, it feeds directly into the evidence pack and the five-pillar score profile by showing how governance readiness shifts when retrieval, tuning, feedback, or stacked controls are introduced.
Fourth, it strengthens reviewability, contestability, audit readiness, and organisational learning because reviewers can reconstruct why one configuration was selected over another and what evidence supports that choice.

6 configurations ? Run variants ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

Link to the five RAIDT pillars

Responsibility

The six configurations affect who is accountable for shaping the run and for approving its outputs. As configuration becomes more structured, responsibility can be allocated more clearly across prompt design, retrieval governance, model adaptation, review, and deployment ownership.