S3.10 - Minimum_metadata

S3.10 ? Minimum metadata

flowchart LR
    A[Stored outputs without context] --> B[RAIDT
Run-level evidence framework]
    A2[Weak logging and unclear review state] --> B
    B --> C[[Minimum metadata
Minimum evidential description of a run]]
    C --> D[Evidence pack]
    C --> E[RAIDT score profile]
    C --> F[Reviewer reconstruction]
    C --> G[Contestability and audit readiness]
    D --> H[Organisational learning]
    E --> G
    I[Healthcare triage] --> C
    J[Public-service drafting] --> C
    K[Finance review] --> C
    L[Education support] --> C
    M[Enterprise productivity] --> C

? Star S3 - Run-Level Evidence Logic

Star context: Explains the proof-object logic of RAIDT by showing the least evidential description a run must carry if it is to be reconstructed, compared, reviewed, and challenged rather than merely asserted.

Academic picture

Definition / background

Minimum metadata is the irreducible descriptive record that makes a specific GenAI run identifiable, intelligible, and reviewable within RAIDT. It usually includes the run identifier, timestamp, task or context marker, prompt or template used, model or system identifier, key settings, produced output, and any retrieval or human-review references that materially shaped the result. The concept is ?minimum? not because it aims to collect as little information as possible for convenience, but because it defines the lowest evidential threshold below which a run can no longer be governed responsibly.

Conceptually, this item sits between raw system logging and a full evidential dossier. It is narrower than complete telemetry, because RAIDT does not require every possible machine event to be retained for every run. It is also more operational than a generic documentation requirement, because the metadata is tied to a specific run and must support later reconstruction, comparison, and challenge. In that sense, minimum metadata functions as the entry point into run-level evidence logic.

Within RAIDT, minimum metadata matters because the framework treats the run as the unit of governance. A run-level evidence pack cannot be assembled coherently if the run itself is weakly identified or poorly described. Likewise, a five-pillar score profile is more credible when the score is anchored to a stable record of what was actually run, when, under which configuration, and with what surrounding evidence. Minimum metadata therefore links the basic description of a run to both evidential packaging and governance judgement.

The term also differs from neighbouring ideas such as audit trail, replayability, or evidence readiness. An audit trail may contain a much richer sequence of actions. Replayability may require stronger preservation of prompts, inputs, tools, and environmental conditions. Evidence readiness concerns whether material is available in a form suitable for scrutiny. Minimum metadata is the threshold layer beneath all three: if it is absent, the stronger properties become difficult or impossible to demonstrate.

Why this concept matters

Organisations often retain outputs from GenAI systems without retaining the minimum context needed to govern those outputs later. When that happens, questions that seem basic become unexpectedly hard: Which model produced this? Which prompt version was used? Was retrieval enabled? Was a human reviewer involved? Was the output generated before or after a policy change? Minimum metadata solves this problem by ensuring that each run carries a minimal evidential signature.

This matters because governance failure in GenAI is often not caused by a total absence of activity, but by the inability to connect an artefact to the conditions of its production. Without minimum metadata, organisations cannot reliably compare runs, investigate incidents, justify scores, or respond to challenge. They are left with assertions such as ?the system was configured appropriately? or ?a reviewer checked it? without the run-level evidence needed to substantiate those claims.

For RAIDT, the concept is important because it operationalises proportionality. Low-risk, high-volume runs do not need the same evidential richness as high-stakes clinical, legal, or welfare-related runs. But even in low-risk contexts, some metadata floor is required if the output is to remain governable. Minimum metadata therefore prevents two opposite errors at once: under-documentation that makes review impossible, and over-documentation that makes governance impractical.

Key idea: Minimum metadata matters because it gives every governed run a defensible evidential identity, allowing RAIDT to move from policy statements to reviewable run-level proof.

What this item captures

The stable identity of a run, including a run ID or equivalent unique reference.
The temporal position of the run, such as timestamp, version date, or review window.
The generative conditions of the run, including prompt or template, model identifier, and relevant settings.
The relationship between inputs, retrieval context, outputs, and any attached human review evidence.
The minimum retention anchors needed to support later reconstruction, comparison, scoring, and audit.
The baseline evidential threshold below which a run cannot be treated as governable within RAIDT.

Practical example / likely audience question

Audience question

Is minimum metadata simply burdensome administrative overhead, especially when organisations run large volumes of low-risk GenAI tasks?

Answer

The concern behind this question is that governance requirements can become detached from operational reality. If every run required a forensic-level evidential dossier, many teams would either fail to comply or bypass governance entirely. RAIDT addresses that concern by separating the minimum evidential baseline from richer evidence requirements that apply when stakes, sensitivity, or downstream dependency increase.

The direct answer is that minimum metadata is not intended to be burdensome; it is intended to prevent a run from becoming evidentially anonymous. A lightweight enterprise productivity task might only require a stable run ID, timestamp, prompt template reference, model identifier, output reference, and reviewer flag. A higher-impact run, by contrast, may need detailed retrieval traces, approval records, and retention controls. The minimum is therefore scalable by risk, but not optional in principle.

A practical example is an organisation using GenAI to draft internal policy summaries. If a summary later contains a harmful omission, the team should at least be able to identify which run produced it, when it was produced, which prompt template was used, which model generated it, and whether anyone reviewed it. Generic AI governance often stops at policy commitments such as ?human oversight is in place.? RAIDT handles the issue better by requiring enough run-level evidence to show what oversight actually attached to the specific output under question.

Practical example in RAIDT terms

Consider a public-service organisation using GenAI to draft first-pass responses to housing-support enquiries. A caseworker receives an output that understates an applicant?s eligibility for temporary accommodation, and the response is challenged a week later. The run-level issue is not merely whether the model performed well on average; it is whether this specific run can be identified and reviewed.

Minimum metadata for that run should include the run ID, timestamp, service context, prompt or template version, model identifier, key generation settings, the source references or retrieval pathway if policy documents were used, the output record, and any reviewer sign-off or amendment. That evidence allows the organisation to determine whether the answer reflected an outdated template, an incorrect retrieval source, an unreviewed output, or a process gap.

In RAIDT terms, the affected pillars are especially Auditability and Traceability, with important implications for Responsibility and Dependability. Auditability depends on being able to inspect the evidential basis of the run. Traceability depends on linking the output back to its configuration and review state. Responsibility is strengthened when reviewer roles and decision points are visible. Dependability improves because recurring failure patterns can be detected and addressed. Minimum metadata therefore improves governance readiness by turning an isolated problematic output into a reviewable case with a reconstructable evidential history.

Detailed link to RAIDT

Minimum metadata links to RAIDT in four ways.

First, it supports RAIDT?s core idea that governance should attach to the run rather than only to a model, policy, or organisational statement.
Second, it gives the run enough descriptive structure to function as run-level evidence rather than as a detached output.
Third, it provides the baseline material from which the evidence pack and justified score profile can be assembled.
Fourth, it enables reviewability, contestability, audit readiness, and organisational learning because the run can be revisited with evidence rather than memory.

Minimum metadata ? Run-level evidence ? Evidence pack ? RAIDT score profile ? Governance readiness

This chain matters because RAIDT is not satisfied by broad claims of good practice. The framework requires that governance judgements can be traced back to the minimum evidential description of what actually happened in a specific run.

Link to the five RAIDT pillars

Minimum metadata has its strongest direct effect on Auditability and Traceability, but it supports all five pillars by making run-level claims checkable.

Responsibility

Responsibility is strengthened when a run can be linked to a task, role, reviewer, or accountable decision point rather than being treated as an anonymous system event.