S7.08 - Boundary_conditions

S7.08 ? Boundary conditions

flowchart LR
    A[Over-general AI governance claims
informal use, weak evidence capture,
contexts where expert judgement remains central] --> B[RAIDT
run-level evidence framework]
    H[Practical fields
healthcare, finance, public services,
workflow wrappers, review checkpoints] --> C[[Boundary conditions
where RAIDT is valid, proportionate, and useful]]
    B --> C
    C --> D[Evidence pack]
    C --> E[RAIDT score profile]
    D --> F[Reviewer reconstruction
and contestability]
    E --> G[Governance readiness
and organisational learning]

? Star S7 - Academic Theory and Design Logic

Star context: Positions RAIDT as a design-science, mechanism-based mid-range theory by clarifying the scope conditions under which run-level evidence, evidence packs, and five-pillar scoring are theoretically defensible and practically useful in organisational governance.

Academic picture

Definition / background

Boundary conditions are the scope conditions that specify where a theory, method, or artefact can reasonably be expected to work. In design-science and mid-range theory terms, they define the circumstances under which a contribution is applicable, informative, and empirically defensible rather than claiming unlimited universality. For RAIDT, boundary conditions describe the organisational settings in which run-level evidence can be captured, reviewed, and turned into meaningful governance outputs.

This matters because RAIDT is not a general philosophy of responsible AI in the abstract. It is a practical framework for governing concrete uses of generative AI through the run as the unit of analysis. That means RAIDT works best where there is a defined task, a bounded context, identifiable actors, feasible evidence capture, and some meaningful consequence attached to the use. Typical examples include repeated or high-impact organisational uses, regulated workflows, decision-support tasks, and settings in which reviewability and audit readiness matter.

Boundary conditions differ from limitations in a narrow sense. A limitation describes weakness or incompleteness; a boundary condition defines legitimate scope. The concept therefore protects RAIDT from two errors at once: over-claiming, where the framework is presented as suitable for every form of GenAI use, and under-claiming, where its value in accountability-sensitive organisational work is diluted by comparison with casual or purely personal use.

Inside RAIDT, boundary conditions matter because they determine whether run-level evidence is realistic, whether an evidence pack can be assembled with integrity, whether a score profile can be justified without guesswork, and whether the five pillars can be assessed on evidence rather than assertion. The concept therefore belongs in Academic Theory and Design Logic because it clarifies the theoretical and practical scope of the RAIDT contribution.

Why this concept matters

Boundary conditions solve a basic but often neglected governance problem: organisations need to know not only what a framework does, but where it should and should not be used. Without this clarity, AI governance easily becomes either too broad to be operational or too narrow to be useful. In RAIDT, boundary conditions prevent the framework from being treated as a universal checklist for every interaction with GenAI, while also identifying the contexts in which its evidence-centred logic is most valuable.

The concept also avoids confusion between principle-level aspiration and method-level applicability. An organisation may endorse accountability, transparency, or responsible AI in all cases, but RAIDT requires more than aspiration. It requires a run that can be reconstructed, evidence that can be retained safely, and a workflow in which review and scoring are proportionate to the stakes. Boundary conditions therefore translate broad governance ambitions into a usable decision about when RAIDT should be activated.

If boundary conditions are ignored, several risks appear: evidence demands may be imposed on trivial uses where they create burden without value; important high-stakes uses may be governed too lightly because they are treated as ordinary productivity tasks; score profiles may be generated where evidence is too thin to justify them; and critics may dismiss RAIDT either as unrealistic bureaucracy or as theoretically vague. Clear boundary conditions prevent these distortions.

Key idea: Boundary conditions matter because they show where RAIDT can turn run-level evidence into credible governance rather than symbolic documentation.

What this item explains

The kinds of organisational GenAI use for which RAIDT is appropriately designed.
The threshold conditions under which run-level evidence can be captured safely and meaningfully.
The difference between accountability-sensitive organisational use and low-consequence casual use.
The point at which evidence packs and score profiles are proportionate rather than excessive.
Why some contexts still require human domain judgement that RAIDT can support but not replace.
How scope clarity strengthens the theoretical status of RAIDT as a mid-range, mechanism-based design contribution.

Practical example / likely audience question

Audience question

Where does RAIDT not fit well?

Answer

Low-consequence casual use, contexts where evidence capture is impossible or cases where domain judgement cannot be replaced.

The concern behind this question is whether RAIDT becomes unnecessarily heavy or normatively over-ambitious if applied everywhere. The direct answer is that RAIDT is not intended for every fleeting interaction with a generative model. If a use is inconsequential, ephemeral, or purely personal, the governance value of constructing run-level evidence may be limited. Equally, if the relevant evidence cannot be captured lawfully, technically, or operationally, then RAIDT cannot deliver its full method because its outputs would rest on incomplete reconstruction.

A practical example is the difference between a staff member privately brainstorming slide titles and a clinical team using GenAI to draft discharge communications. The first case may not justify a formal evidence pack. The second does, because the run sits inside an organisational process with real consequences, review obligations, and traceability needs. RAIDT handles this issue better than a generic AI governance approach because it does not merely say that all AI use should be governed; it specifies the conditions under which governance should become evidence-led, reconstructable, and reviewable at the level of the run.

Practical example in RAIDT terms

Consider a public-service setting in which a local authority uses a GenAI assistant to draft case summaries for housing-support decisions. The use case is appropriate for RAIDT because the task is repeated, organisationally consequential, and embedded in a reviewable workflow. The run-level issue is whether each draft can be traced back to the prompt, source case notes, policy guidance, staff review, and final decision rationale.

The evidence needed includes the task purpose, the case type, the prompt template, relevant source materials, model and configuration details, the generated draft, the officer's edits, the reviewer comments, timestamps, and the approved final summary. Responsibility is affected because accountable roles must be clear. Auditability and Traceability are strongly affected because a later reviewer may need to reconstruct why a claimant received a particular outcome. Interpretability matters because the authority must understand how the draft relates to policy instructions and source facts. Dependability matters because repeated use should not produce unstable or misleading summaries.

Boundary conditions improve governance readiness here by showing why RAIDT is a proportionate fit: the task is repeated, the consequences are material, evidence capture is feasible, and human review remains central. By contrast, the same authority would not need RAIDT at full strength for an employee's low-stakes request to brainstorm a workshop icebreaker, because the governance value of run reconstruction would be minimal.

Detailed link to RAIDT

Boundary conditions link to RAIDT in four ways.

First, they connect to the RAIDT core idea by defining the kinds of organisational GenAI use for which evidence-over-assertion is both necessary and practicable.
Second, they connect to the run because they determine when a run can realistically function as the unit of governance.
Third, they connect to the evidence pack and score profile because those outputs are only credible when the underlying run can be evidenced in a proportionate and reconstructable way.
Fourth, they connect to reviewability, contestability, audit readiness, and organisational learning because those governance benefits depend on using RAIDT in contexts where evidence can actually support scrutiny.

Boundary conditions -> Run-level evidence -> Evidence pack -> RAIDT score profile -> Governance readiness

Boundary conditions therefore do not sit outside RAIDT as a disclaimer. They shape whether the framework can operate as intended and whether its outputs should be trusted as governance artefacts.

Link to the five RAIDT pillars

Responsibility

Boundary conditions clarify when responsibility can be meaningfully assigned around a run. If roles, decision authority, or review obligations are unclear, RAIDT may be less suitable or may need stronger workflow design before use.