S3.03 - Proof_object

S3.03 ? Proof object

flowchart LR
    A[Fragmented documentation and weak governance claims] --> B[RAIDT - run-level evidence framework]
    B --> C[[Proof object - structured reviewable object for one run]]
    C --> D[Run-level evidence pack]
    C --> E[Five-pillar score profile]
    C --> F[Reviewer reconstruction and challenge]
    D --> G[Audit readiness]
    E --> H[Governance readiness]
    I[Healthcare, finance, education, public services, enterprise use] --> C

← Star S3 - Run-Level Evidence Logic

Star context: Places the proof object inside RAIDT's run-level evidence logic. It explains how evidence from a single run is assembled into a reviewable object that supports reconstruction, comparison, challenge, and governance judgement.

Academic picture

Definition / background

A proof object in RAIDT is the structured evidence object that shows how a specific run occurred and on what basis governance claims about that run can be examined. It is best understood as the reviewable form of a run-level evidence pack: evidence is gathered, organised, and presented so that another party can reconstruct the run, inspect whether relevant controls were present, and challenge weak or unsupported claims.

The term matters because governance often fails at the point where an organisation must move from saying that safeguards exist to showing how those safeguards operated in a real use episode. In that sense, a proof object is not the same as a raw log, a single evidence file, or a policy statement. A raw log may be too technical, fragmented, or incomplete. A policy statement may declare intent without demonstrating execution. A proof object brings together the evidential components required for review at the level of one run.

Within RAIDT, this concept belongs inside run-level evidence logic because RAIDT treats the run as the unit of governance. The proof object therefore ties governance assessment to a concrete run in a concrete context, rather than to generic claims about a model, vendor, or organisational principle. It supports both of RAIDT's practical outputs: the run-level evidence pack itself and the five-pillar score profile derived from that evidence.

The concept also helps distinguish between related items in Star S3. An evidence object may be any individual evidential element. A proof object is the assembled, review-ready evidential whole. It exists so that reviewers can move from isolated artefacts to a coherent basis for judgement about responsibility, auditability, interpretability, dependability, and traceability.

Why this concept matters

The proof object solves a central operational problem in generative AI governance: even when evidence exists, it is often not assembled in a form that supports scrutiny. Without a proof object, an organisation may have prompts in one place, outputs in another, policy references elsewhere, and no stable way to show how the run should be interpreted. Review then becomes slow, inconsistent, and vulnerable to selective disclosure.

This concept also avoids a common confusion between having documentation and having reviewable proof. Documentation may describe a process in general terms. A proof object shows what happened in one run and how that run can be assessed. That distinction matters in high-stakes organisational settings, because governance failures often emerge from the difference between intended control and control in actual use.

If the proof object is missing, the organisation struggles to defend claims, compare runs, investigate incidents, or improve practice over time. It becomes difficult to challenge decisions, difficult to support an audit trail, and difficult to show why a score profile was justified. RAIDT uses the proof object to turn governance into something inspectable, contestable, and operational.

Key idea: A proof object matters because it converts dispersed run evidence into a reviewable governance object that can justify, challenge, and improve decisions about GenAI use.

What this item captures

The evidential basis for claims about how one run was configured and executed.
The relationship between raw run artefacts and a reviewer-ready evidence pack.
The presence, absence, or weakness of governance controls at run level.
The basis on which a RAIDT score profile can be assigned and defended.
The information needed for reconstruction, comparison, challenge, and organisational learning.
The distinction between assertion that controls exist and evidence that they operated in practice.

Practical example / likely audience question

Audience question

Does proof object mean legal proof, or a claim that RAIDT can conclusively prove that a system was compliant or correct?

Answer

The concern behind that question is understandable, because the word 'proof' can imply certainty, legal sufficiency, or mathematical demonstration. In RAIDT, that is not the meaning. A proof object is a governance proof object: a structured, reviewable evidential object that allows a run to be examined and governance claims about that run to be tested.

The direct answer is no. RAIDT does not claim that a proof object automatically proves legality, truth, or ethical adequacy in an absolute sense. Instead, it provides the evidence needed for accountable review. For example, if a team uses a large language model to draft a policy briefing, the proof object may contain the prompt, output, model identifier, time stamp, user role, task purpose, review notes, and control metadata. That does not guarantee the briefing is correct. It does allow a reviewer to see how the run happened, whether safeguards were applied, and whether the governance judgement is defensible.

RAIDT handles this better than generic AI governance approaches because it ties the object to one run rather than relying on broad claims such as 'the system is governed' or 'the vendor is compliant'. The proof object therefore supports practical scrutiny, not abstract reassurance.

Practical example in RAIDT terms

Consider a healthcare setting in which a clinician uses a generative AI assistant to draft a discharge summary from structured notes. The run-level issue is not simply whether the model is approved in principle, but whether this specific run was conducted under the right controls for a clinically sensitive task.

The proof object for that run would include the prompt template, the contextual data supplied, the model and version used, the time and user role, any policy restrictions, the generated draft, the clinician's amendments, and evidence of human review before release. Additional metadata might record whether patient identifiers were masked, whether the task was marked as high risk, and whether the model's output triggered any warning or escalation rule.

The affected RAIDT pillars are Responsibility, because role allocation and human sign-off matter; Auditability, because the run must be reviewable; Dependability, because the output must be reliable enough for clinical use; and Traceability, because the origin and handling of the output must be reconstructable. Interpretability may also be relevant where reviewers need to understand why the model was used and how the output was assessed. In this setting, the proof object improves governance readiness because it gives supervisors, auditors, and clinical governance leads a concrete basis for checking whether the run was controlled appropriately.

Detailed link to RAIDT

Proof object links to RAIDT in four ways.

First, it operationalises RAIDT's core idea that governance should be based on evidence rather than assertion.
Second, it ties that evidence to the run as the unit of governance, so review is grounded in a specific use event.
Third, it provides the structured basis from which the evidence pack is assembled and the score profile is justified.
Fourth, it enables reviewability, contestability, audit readiness, and organisational learning because the evidential basis of judgement is visible to others.

Proof object -> Run-level evidence -> Evidence pack -> RAIDT score profile -> Governance readiness

In practice, the proof object is the point at which raw artefacts become governable evidence. Without that intermediate object, RAIDT would have data but a weaker basis for assessment and institutional learning.

Link to the five RAIDT pillars

Responsibility

The proof object supports Responsibility by showing who initiated the run, for what purpose, under which role, and with what human oversight or approval expectation.