RAIDT Core, Definition, Values, Claims and Innovation

Defines RAIDT at the highest level and clarifies its central proposition: responsible governance of generative AI in organisational work should be anchored in run-level evidence rather than broad assurance claims alone. This star establishes the project vocabulary, its normative values, its core outputs, and its main innovation.

This is the centre note for the RAIDT project. It sits primarily in the foundations and theory layer, but it also connects directly to implementation, evidence design, empirical validation, policy pathways, and sector application. It gives supervisors and workshop participants the shortest route into the project because it explains what RAIDT is, why the run matters, how the evidence pack and score profile fit together, and what the framework does and does not claim.

• What does RAIDT mean as a project concept and governance framework?
• Why does RAIDT treat the run, rather than the model or organisation in the abstract, as the unit of governance?
• What problem in responsible AI and Information Systems governance is RAIDT trying to solve?
• What exactly counts as a run in organisational use of generative AI?
• What evidence should be captured to make a run reviewable, contestable, and governable?
• How do the run-level evidence pack and five-pillar RAIDT score profile work together?
• What is the core innovation of RAIDT, and what makes it more than a generic assurance checklist?
• What evidence would be needed to validate RAIDT empirically?
• How does RAIDT connect to prompt engineering, RAG, alignment controls, PEFT or LoRA adaptation, and human oversight?
• How does this star help supervisors understand the foundations of Papers 08, 09, and 10?
• What does RAIDT explicitly not claim to do?

• 10-20 min ? How should RAIDT be defined in one sentence so that it is conceptually precise, operationally useful, and distinct from generic Responsible AI rhetoric?
• 20-40 min ? Is the run the right unit of governance for high-variability GenAI use, or should governance remain focused on systems, models, and organisational policies?
• 40-60 min ? What minimum evidence is required for a run to be reviewable, contestable, and useful for managerial intervention across different sectors?

• C0.01 ? RAIDT
• C0.02 ? Run
• C0.03 ? Run-level evidence
• C0.04 ? Evidence pack
• C0.05 ? Score profile
• C0.06 ? Governance readiness
• C0.07 ? Core value: evidence over assertion
• C0.08 ? Core value: reviewability
• C0.09 ? Core value: contestability
• C0.10 ? Core claim
• C0.11 ? Core innovation
• C0.12 ? What RAIDT is not

RAIDT is best understood as a practical governance framework for generative AI used in organisational work under conditions of uncertainty. The background problem is familiar across Responsible AI, Information Systems governance, and managerial decision-making: organisations increasingly depend on GenAI systems whose behaviour is variable, context-sensitive, and difficult to justify through static documentation alone. A policy may exist, a model card may exist, and a procurement checklist may exist, yet none of these necessarily explain what happened in a specific use episode when a member of staff used a configured AI system for a real task. In practice, risk often materialises not at the abstract system level, but in situated use.

RAIDT addresses this gap by treating the run as the unit of governance. A run is one configured use of a generative AI system for a specific task, at a specific time, in a specific context. It includes the instruction or prompt, the model and tool configuration, retrieved context where retrieval is used, the output, and the human or automated checks applied to that output. This matters because the same model can be acceptable in one run and problematic in another, depending on the task, domain, prompt design, attached tools, retrieved documents, reviewer actions, and stakes of the decision. RAIDT therefore shifts governance attention from abstract claims about AI systems to inspectable evidence about what was actually done.

The framework produces two practical outputs. The first is a run-level evidence pack. This is the structured record that allows a reviewer, supervisor, auditor, or governance function to reconstruct and assess the run. The second is a five-pillar RAIDT score profile covering Responsibility, Auditability, Interpretability, Dependability, and Traceability. The evidence pack is the documentary basis; the score profile is the evaluative summary. The relationship is important: RAIDT is not intended as scoring without evidence, and it is not merely evidence collection without interpretive consequence. Its contribution is the pairing of evidential capture with governance judgement.

The core values of RAIDT are evidence over assertion, reviewability, and contestability. Evidence over assertion means that claims such as ?the model is safe?, ?the prompt is well designed?, or ?the process is compliant? should not stand on their own if no run-level record exists to support them. Reviewability means a third party should be able to inspect a run and understand how the output was produced, checked, and used. Contestability means important outputs should be open to challenge, especially where uncertainty is high, harms are plausible, or organisational consequences are significant. These values are especially relevant for GenAI because outputs can appear fluent and authoritative while still being wrong, biased, poorly grounded, or misaligned with task requirements.

This is where RAIDT connects to prompt engineering, RAG, alignment controls, and model adaptation. Prompt engineering shapes the instruction structure and therefore strongly affects output quality and risk. Retrieval-augmented generation changes the evidential basis of the output by introducing external sources whose provenance and relevance need examination. PEFT or LoRA modifications may alter model behaviour in task-specific ways that should be visible in governance records. RLHF and related alignment controls may improve general behavioural tendencies, but they do not remove the need to examine situated runs. RAIDT does not deny the importance of model-level assurance; rather, it argues that model-level assurance is incomplete without run-level evidence when the system is used in organisational practice.

The problem RAIDT solves is therefore both conceptual and operational. Conceptually, it offers a clearer governance unit for systems whose outputs depend on interaction, context, and configuration. Operationally, it gives organisations a way to capture evidence that supports intervention. If a run scores weakly on Traceability because the retrieved sources were not logged, an organisation can introduce logging requirements. If Interpretability is weak because the rationale for prompt design is absent, teams can improve prompt templates and reviewer notes. If Dependability is weak because outputs vary materially across repeated runs, the organisation may restrict use to lower-stakes tasks or strengthen human review. RAIDT is useful because it links diagnosis to action.

This makes the framework relevant to standards and policy alignment. The EU AI Act, ISO/IEC 42001, and the NIST AI RMF all point, in different ways, towards documentation, risk management, accountability, human oversight, and traceability. RAIDT should not be presented as a substitute for these frameworks, nor as automatic proof of compliance. Its role is more specific: it operationalises governance expectations at the level where many practical failures and contestable decisions actually occur. It can therefore function as a bridge between high-level governance requirements and the messy reality of day-to-day AI use in organisations.

The practical examples are straightforward. A university administrator using a GenAI assistant to draft a student case summary generates a run that should record the prompt, any retrieved policy documents, the model version, the draft output, and who checked the result before use. A hospital management team using GenAI to summarise incident reports would need stronger evidence because the domain is higher risk and the possibility of omission or distortion matters more. A legal operations team using a RAG-enabled tool to prepare first-pass contract analysis needs source traceability and clear review boundaries. In each case, the same general AI capability exists, but governance quality depends on the specific run.

RAIDT also supports the wider architecture of the project. Paper 08 can use this star to define the foundational logic, core constructs, and methodological pathway from concept to operational framework. Paper 09 can use it to specify what should be observed, captured, and tested in empirical validation, including inter-rater assessment of scoring and the usefulness of evidence packs in practice. Paper 10 can use it to articulate how run-level evidence might inform policy pathways, sector guidance, and governance design choices. Sector playbooks can translate the core logic into domain-specific templates and intervention rules.

The boundaries are equally important. RAIDT does not claim to eliminate uncertainty, guarantee truthful outputs, replace professional judgement, or explain internal model cognition in full. It is not a universal benchmark, a legal compliance certificate, or a complete governance system on its own. It is a disciplined run-level framework that makes GenAI use more inspectable, discussable, and governable. Its innovation is modest in one sense and strong in another: it does not promise magic, but it offers a concrete governance mechanism for moving from vague AI assurance to actionable evidence.

Q1. What is RAIDT in one sentence?

Answer:
RAIDT is a run-level evidence framework for responsible governance of generative AI in organisational work. It defines the run as the key unit of analysis and governance, then uses structured evidence and a five-pillar score profile to support review, intervention, and accountability.

Practical example:
A procurement officer uses a GenAI tool to draft a supplier risk summary. RAIDT captures the specific task, prompt, model configuration, source documents, output, and review steps for that single use episode.

Link to RAIDT:
This is the framework?s central definition and explains why the evidence pack and score profile are the two core outputs.

Q2. Why does RAIDT focus on the run rather than only the model?

Answer:
Because organisational risk arises through situated use. The same model can produce different outcomes depending on prompt design, retrieved documents, tool access, user intent, timing, and checking procedures. Governance therefore needs a unit that reflects actual use rather than abstract capability.

Practical example:
A general-purpose model may behave acceptably in low-stakes brainstorming but poorly in a compliance reporting task if retrieval is weak or review is absent.

Link to RAIDT:
The run anchors Traceability, Auditability, and Dependability by showing what happened in context.

Q3. What problem does RAIDT solve?

Answer:
It solves the gap between broad Responsible AI principles and the evidence needed to govern real GenAI use in organisations. Many governance approaches remain too high-level to explain or assess specific outputs.

Practical example:
An organisation may have an AI policy but still be unable to reconstruct how a problematic response was generated in a staff workflow.

Link to RAIDT:
The evidence pack closes that gap by making the run reconstructable and reviewable.

Q4. What is included in a run?

Answer:
A run includes the task context, prompt or instruction, model and tool configuration, retrieved context where used, output, timestamps, and the human or automated checks applied. Depending on the domain, it may also include user role, risk classification, escalation route, and final decision.

Practical example:
For a RAG-supported policy summary, the run should record the source documents retrieved, the retrieval settings, and whether a human reviewer confirmed the final summary.

Link to RAIDT:
These fields become the raw material of the evidence pack and support scoring across all five pillars.

Q5. Why are evidence packs necessary?

Answer:
Evidence packs convert transient AI interactions into inspectable governance records. Without them, organisations rely on memory, screenshots, or assertions, which are weak forms of accountability.

Practical example:
If a staff member relies on an AI-generated recommendation that later proves harmful, the evidence pack allows investigators to examine what information the system used and what checks were applied.

Link to RAIDT:
The evidence pack operationalises evidence over assertion and enables later audit, challenge, and intervention.

Q6. What does the five-pillar score profile add?

Answer:
The score profile translates evidence into a structured judgement about governance quality. It does not replace qualitative review, but it provides a consistent way to compare runs, identify weaknesses, and target improvements.

Practical example:
Two teams may both use the same model, but one scores lower on Responsibility and Traceability because approval roles and retrieval provenance were not documented.

Link to RAIDT:
Scoring gives the framework operational traction by linking evidence to action.

Q7. How does RAIDT relate to Responsible AI and uncertainty?

Answer:
RAIDT treats uncertainty as a normal condition of GenAI use rather than an exception. Responsible AI principles become more useful when they are tied to specific runs where uncertainty, discretion, and consequences can be assessed.

Practical example:
A manager using AI to prepare a briefing may be uncertain whether the answer is complete. RAIDT requires the run to show the basis of the answer and the checks performed before use.

Link to RAIDT:
Responsibility and Dependability are strengthened when uncertainty is documented rather than ignored.

Q8. How does RAIDT connect to prompt engineering and RAG?

Answer:
Prompt engineering and RAG are not peripheral technical details; they are governance-relevant components because they shape what the model sees and how it responds. Poor prompts or weak retrieval can directly degrade output quality.

Practical example:
A vague prompt plus unverified retrieval may produce a polished but inaccurate summary of organisational policy.

Link to RAIDT:
Recording prompts, retrieved sources, and tool settings improves Interpretability, Traceability, and Auditability.

Q9. Does RAIDT claim to prove that an AI output is correct?

Answer:
No. RAIDT does not guarantee correctness, safety, or compliance. It improves the conditions for checking, reviewing, contesting, and governing outputs. It is a governance framework, not an oracle.

Practical example:
A well-documented run may still contain an incorrect answer, but the organisation can see how it arose and decide how to prevent recurrence.

Link to RAIDT:
This is a core boundary condition and prevents the project from overstating its claims.

Q10. Why is this star important for supervisors?

Answer:
Because it clarifies the project?s identity, scope, novelty, and testable contribution. Supervisors need a stable statement of what RAIDT is before they can assess methods, empirical design, or policy implications.

Practical example:
In supervision, disagreement often begins with definitional ambiguity. This star resolves that by naming the unit of analysis, outputs, values, and limits.

Link to RAIDT:
It is the hub that connects conceptual foundations, empirical validation, policy pathways, and sector translation.

• A university quality assurance team uses GenAI to draft first-pass module review summaries. RAIDT captures the prompt template, institutional policy documents retrieved through RAG, the reviewer who checked the output, and the score profile used to decide whether the run was fit for internal use.
• A legal operations function uses a GenAI assistant to identify clauses in supplier contracts. RAIDT requires evidence of model version, retrieval sources, clause extraction prompt, human verification, and any escalation triggered by low-confidence or ambiguous outputs.
• A local authority service team uses GenAI to draft citizen-facing communications. RAIDT helps govern whether tone guidance, policy sources, approval steps, and audit trails were present before the communication was issued.
• A healthcare administration unit uses GenAI to summarise non-clinical incident reports. RAIDT highlights stronger governance readiness needs because omission, distortion, and traceability failures could materially affect organisational learning.

• Run identifier, date, time, task description, business purpose, and organisational context
• User role, approval role, and whether the run is high-stakes, medium-stakes, or low-stakes
• Prompt or instruction text, prompt template version, and any system message or policy wrapper
• Model name, model version, deployment environment, tool access, and parameter settings where relevant
• Whether PEFT, LoRA, or other task-specific adaptations were active
• Whether RLHF or other alignment controls are assumed at system level and how that assumption is documented
• Retrieved sources, retrieval method, source provenance, and retrieval timestamps for RAG-enabled runs
• Output produced, confidence indicators where available, and post-processing steps
• Human review actions, automated checks, exceptions raised, and final decision or use outcome
• RAIDT pillar scores, scoring rationale, reviewer comments, and governance intervention triggered
• Retention, access control, and contestation route for later review

This note is the conceptual anchor for the whole RAIDT project.

• Paper 08: foundations and methodological pathways. It defines the central constructs, the meaning of the run, the logic of evidence capture, and the claim that governance should be grounded in inspectable use episodes.
• Paper 09: empirical validation. It identifies what should be tested in practice, including feasibility of evidence capture, consistency of scoring, supervisory usefulness, and whether RAIDT improves reviewability and intervention quality.
• Paper 10: policy pathways. It frames how run-level governance can complement broader regulatory and standards-based approaches without claiming to replace them.
• Sector playbooks. It provides the shared core that can be adapted into domain-specific checklists, threshold rules, and evidence templates.
• RAIDT scoring. It explains why scoring must be evidence-based, interpretable, and tied to governance action.
• RAIDT evidence pack. It defines the evidence pack as the practical record of a run rather than an abstract repository of assurance claims.
• RAIDT governance interventions. It clarifies that the point of scoring and evidence capture is to enable actions such as escalation, restriction, redesign, reviewer strengthening, and policy refinement.

• Responsible AI governance literature on accountability, contestability, and documentation
• Information Systems governance literature on control, audit trails, and socio-technical use context
• AI uncertainty and managerial uncertainty literature
• GenAI risk literature on hallucination, variability, opacity, and automation over-reliance
• Prompt engineering and RAG literature on context construction and output quality
• Alignment and adaptation literature covering RLHF, PEFT, and LoRA as governance-relevant configuration layers
• Standards and policy materials relating to EU AI Act documentation, ISO/IEC 42001 management systems, and NIST AI RMF governance and mapping functions
• Empirical methods literature on rubric design, inter-rater reliability, case comparison, and evaluation of governance artefacts

• RAIDT does not claim that run-level evidence alone is sufficient for full AI governance.
• RAIDT does not replace model evaluation, procurement due diligence, security controls, or organisational policy.
• RAIDT does not guarantee output correctness, fairness, legality, or safety.
• RAIDT does not provide a complete explanation of internal model reasoning.
• RAIDT does not remove the need for human judgement, escalation, or domain expertise.
• RAIDT should not be presented as an automatic compliance certificate for the EU AI Act, ISO/IEC 42001, or NIST AI RMF.
• RAIDT may introduce documentation overhead, so proportionality and sector tailoring matter.
• RAIDT is strongest where organisations are willing to capture evidence consistently and use it for real governance decisions.

RAIDT is the centrepiece of the project because it tells us what exactly is being governed when generative AI is used in organisations. My core argument is that governance should not stop at model-level claims or high-level policy statements. It should also examine the run, meaning one configured use of a GenAI system for a specific task, at a specific time, in a specific context. That is the level at which prompts, retrieved documents, tool settings, outputs, checks, and human judgement come together. RAIDT therefore produces two linked outputs: a run-level evidence pack and a five-pillar score profile covering Responsibility, Auditability, Interpretability, Dependability, and Traceability. The innovation is not that it promises certainty; it does not. The innovation is that it makes GenAI use more reviewable, contestable, and governable in practice. This note matters because it defines the project?s identity, its values, its boundaries, and the logic that connects the theoretical work in Paper 08 to the empirical work in Paper 09 and the policy pathways in Paper 10.