C0.07 - Core_value_evidence_over_assertion

C0.07 ? Core value: evidence over assertion

flowchart LR
    A[Traditional assurance claims
policy, ethics statements, supplier promises] --> B[RAIDT
run-level evidence framework]
    H[Practical run artefacts
prompts, inputs, outputs, review notes, timestamps] --> C[[Core value: evidence over assertion]]
    I[Sector contexts
education, healthcare, finance, public services] --> C
    B --> C
    C --> D[Run-level evidence and evidence pack]
    C --> E[RAIDT score profile]
    D --> F[Reviewer reconstruction
reviewability and contestability]
    E --> G[Governance readiness
audit defence and organisational learning]

? Star C0 - RAIDT Core, Definition, Values, Claims and Innovation

Star context: Defines the project identity of RAIDT by insisting that responsible governance of GenAI in organisational work must rest on demonstrable run-level evidence rather than on policy rhetoric, vendor assurance, or retrospective justification alone.

Definition / background

Evidence over assertion is a core RAIDT value stating that governance claims about generative AI should be supported by reconstructable evidence from actual use rather than by general assurances alone. In practical terms, RAIDT asks not only whether an organisation has a policy, a principle, or a confidence statement, but whether it can show what occurred in one specific run, under what conditions, with what controls, and with what review.

Conceptually, this value responds to a familiar weakness in AI governance. Organisations often possess documentation that describes intention, compliance posture, or supplier capability, yet those materials do not necessarily demonstrate what happened in a real use event. "Assertion" in this context therefore includes verbal assurances, policy declarations, unchecked assumptions, and after-the-fact narratives that cannot be tied to traceable artefacts. "Evidence" means the records, metadata, outputs, human interventions, and evaluative notes that allow governance claims to be examined.

Within RAIDT, this value belongs at the core because the framework is built around run-level evidence, evidence packs, and a five-pillar score profile. These elements only have integrity if scoring and judgement are anchored in evidence rather than impression. The value therefore links the philosophical stance of RAIDT to its operational design: the framework treats evidence as the basis for reviewability, contestability, and governance readiness.

This also distinguishes RAIDT from approaches that remain principle-heavy but execution-light. A principle may say that human oversight exists; evidence over assertion asks whether the record shows who reviewed the run, what they changed, and why. A policy may say outputs are checked; evidence over assertion asks whether a reviewer could later verify that this happened in a particular case.

Why this concept matters

This concept matters because generative AI governance often fails at the point where accountability becomes concrete. When an output is challenged, when an incident occurs, or when a supervisor asks how a result was produced, organisations frequently discover that they can describe their governance intentions but cannot demonstrate their governance practice. RAIDT addresses that gap by making evidential sufficiency a core value rather than an optional extra.

The concept also prevents a drift into narrative assurance. Without this value, organisations can overestimate governance quality simply because they have policy language, training slides, approval statements, or supplier documentation. Those materials are useful, but they do not by themselves show whether a specific GenAI use was responsible, reviewable, dependable, or traceable in context.

If evidence over assertion is missing, several risks follow: weak audit defence, poor incident reconstruction, difficulty contesting harmful outputs, inflated confidence in weak controls, and limited organisational learning. In effect, governance becomes hard to test. RAIDT uses this value to convert broad responsibility claims into something operationally examinable.

Key idea: Evidence over assertion matters because RAIDT treats responsible GenAI governance as something that must be demonstrated through run-level proof, not merely declared in principle.

What this item enables

It enables governance claims to be checked against records from a real GenAI run rather than accepted at face value.
It enables evidence packs to contain substantive artefacts instead of only summary statements or compliance language.
It enables the RAIDT score profile to be justified with observable traces, review records, and decision rationale.
It enables supervisors, auditors, managers, and practitioners to reconstruct how a contested output emerged.
It enables organisations to distinguish between documented intention and demonstrated practice.
It enables learning from failures, near misses, and good practice because runs can be compared on an evidential basis.
It enables a governance culture in which reviewability and contestability are expected outcomes of system use.

Practical example / likely audience question

Audience question

Is "evidence over assertion" just a slogan for better documentation, or does it change how RAIDT governs GenAI use?

Answer

The concern behind this question is that many governance frameworks already ask for documentation, so the phrase may sound rhetorical. The direct answer is that RAIDT changes the level at which governance is judged. It does not treat documentation as sufficient merely because it exists; it asks whether the documentation contains enough run-level evidence to support later review, challenge, and scoring.

For example, a financial services team may say that all AI-assisted customer communications are reviewed by staff before release. That is an assertion. Under RAIDT, the governance question becomes: can the organisation show the prompt or task instruction, the generated draft, the identity or role of the reviewer, the edits made, the approval step, and the final communication issued? If it can, the claim becomes evidentially grounded. If it cannot, the statement remains only a governance assurance narrative.

RAIDT handles this better than a generic AI governance approach because it binds the value directly to the run. Rather than asking whether the organisation has adopted responsible language, it asks whether the governance claim survives inspection at the level of one actual use event. That is a materially stronger test of governance quality.

Practical example in RAIDT terms

Consider an education setting in which a university administrator uses a GenAI system to draft reasonable-adjustment guidance for a student support case. The use case appears routine, but the run-level issue is whether the advice was generated from the correct institutional policy, whether sensitive details were handled properly, and whether a human reviewer checked the output before it informed a student-facing decision.

The evidence needed includes the task purpose, the prompt used, the policy documents or notes provided as source material, the tool and version used, the generated draft, reviewer comments, edits to remove unsupported claims, and the final approved text. Responsibility is affected because the institution must identify who reviewed and approved the guidance. Auditability is affected because a later reviewer must be able to reconstruct the run. Interpretability is affected because the administrator must understand how the output related to the source material and instructions. Dependability is affected because the guidance must be consistent and fit for use. Traceability is affected because the run must be linked to time, actor, tool, and downstream action.

Evidence over assertion improves governance readiness here because the university is no longer limited to saying that it has a policy on AI-assisted drafting. It can show whether this specific case followed that policy in practice and whether the output was handled responsibly before use.

Detailed link to RAIDT

Core value: evidence over assertion links to RAIDT in four ways.

First, it expresses the RAIDT core idea that governance quality should be judged by what can be demonstrated about actual organisational use of GenAI.

Second, it links directly to the run, because the preference for evidence only becomes operational when one concrete run can be reconstructed and examined.

Third, it underpins the evidence pack and the RAIDT score profile, since neither output is credible if based only on narrative assurance or self-description.

Fourth, it supports reviewability, contestability, audit readiness, and organisational learning by requiring governance claims to remain open to later inspection.

Core value: evidence over assertion ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

This chain matters because RAIDT is not simply asking organisations to collect more information. It is asking them to ground governance judgement in evidence that can travel from one run into structured review, comparative scoring, and practical improvement.

Link to the five RAIDT pillars

Responsibility

Evidence over assertion strengthens Responsibility by requiring organisations to show who was accountable for a run, who reviewed it, and what obligations were attached to its use.