S3.05 - Reconstructability

S3.05 ? Reconstructability

flowchart LR
    A[Fragmented records, memory dependence,
incomplete logs after a GenAI run] --> B[RAIDT
run-level evidence framework]
    H[Run fields
identifier, prompt version, model config,
retrieved context, output, review, retention] --> C[[Reconstructability
determine what happened from stored evidence]]
    B --> C
    C --> D[Evidence pack]
    C --> E[RAIDT score profile]
    C --> F[Reviewer reconstruction]
    D --> G[Governance readiness, contestability,
audit readiness, organisational learning]
    E --> G
    F --> G

? Star S3 - Run-Level Evidence Logic

Star context: Explains the proof-object logic inside RAIDT by asking whether a run can be determined from evidence after the event, so that review, comparison, challenge, and governance judgement do not depend on memory or informal explanation.

Academic picture

Definition / background

Reconstructability is the ability to determine what happened in a run from stored records, even if the original user is unavailable. In RAIDT, this is a core evidential requirement inside run-level evidence logic: a run should leave behind enough structured material for another competent reviewer to understand the task, the conditions of use, the output produced, and the human and organisational handling around that output.

Conceptually, reconstructability sits between mere record keeping and full replay. It is stronger than simply having some artefacts, because the artefacts must be sufficient to support meaningful reconstruction. It is different from replayability, because RAIDT does not always require the run to be reproduced exactly; rather, it requires the run to be examinable after the fact. It is also different from comparability, because comparison across runs depends first on each run being reconstructable on its own terms.

Within GenAI governance, reconstructability matters because many organisational questions arise after a run has already happened. A manager, auditor, supervisor, regulator, or affected colleague may need to ask what the user was trying to do, what context was provided, which model configuration was active, what output was generated, what checks were applied, and why the output was accepted, edited, rejected, or escalated. If those questions cannot be answered from evidence, governance remains too dependent on memory, trust, or retrospective narrative.

Inside RAIDT, reconstructability belongs directly to the run-level evidence framework because it determines whether a run can become a credible proof object. It underpins the evidence pack, strengthens the basis of the five-pillar score profile, and helps translate the project?s central commitment to evidence over assertion into an operational governance test.

Why this concept matters

Reconstructability solves a fundamental governance problem: organisations often know that GenAI use should be reviewable, but they do not always know what minimum evidence is needed for review after the event. Without reconstructability, a run may leave traces without leaving enough proof. That creates confusion between having records and having usable evidence.

The concept also prevents a common failure mode in responsible AI practice. An organisation may have policy statements, training material, model documentation, or procurement assurance, yet still be unable to explain one disputed or high-impact run. RAIDT treats this as a serious gap because governance credibility depends on what can be shown when a real case is examined, not only on what was promised in advance.

If reconstructability is weak, several risks follow: incident review becomes partial, disputes become harder to resolve, score profiles become less defensible, and organisational learning is reduced because the run cannot be revisited in a disciplined way. Reconstructability therefore helps move governance from principle-level aspiration to operational evidence.

Key idea: Reconstructability matters because RAIDT can govern a run only if the run can later be determined from evidence rather than recalled from memory.

What this item enables

It enables after-the-fact determination of what happened in one GenAI run.
It enables independent review by someone other than the original user or operator.
It enables a run to function as a proof object rather than as an anecdote.
It enables evidence-pack assembly from structured run records rather than informal recollection.
It enables more credible five-pillar scoring because judgements can be traced back to artefacts.
It enables challenge, contestability, and supervisory questioning when outputs are disputed.
It enables cross-run learning because comparable cases can be examined on a reconstructable basis.
It enables retention and review policies to be tied to governance need rather than to ad hoc documentation habits.

Practical example / likely audience question

Audience question

What minimum does reconstructability need?

Answer

The concern behind this question is whether reconstructability is being defined so broadly that it becomes impractical. The direct answer is that reconstructability needs enough evidence to let another reviewer determine what happened in the run without relying on the original user?s memory. In RAIDT terms, that usually means identifiers, prompt/version, model/configuration, retrieved context, output, review and retention data.

A practical example is an enterprise productivity workflow in which a member of staff uses GenAI to draft a client-facing summary. If only the final text is retained, the organisation may know that GenAI was used but may not know which prompt template was used, what source material was supplied, which model version generated the draft, whether retrieval brought in additional context, who reviewed the result, or how long the evidence should remain available. In that situation, reconstructability is weak even though some records exist.

RAIDT handles this issue better than a generic AI governance approach because it defines reconstructability at the run level. It does not ask only whether policies or logs exist in general. It asks whether this specific run can be rebuilt into an evidential account that supports review, challenge, and scoring. That makes the concept both practical and auditable.

Practical example in RAIDT terms

Consider a finance setting in which a bank employee uses a GenAI assistant to draft an internal credit-risk briefing before a lending decision meeting. The GenAI use case is plausible and efficient, but the run-level issue arises when a reviewer later notices that a key risk factor was omitted from the briefing. The governance question is not merely whether the organisation had an AI policy. It is whether this specific run can be reconstructed in enough detail to determine what happened.

The evidence needed would include the run identifier, the prompt or template used, the prompt version, the source financial documents provided, any retrieved contextual material, the model and configuration, the generated draft, the employee?s edits, the reviewer?s annotations, the decision about whether the draft was accepted or rejected, and the retention record showing whether the evidence remains available for review. Responsibility is affected because reviewers need to know who initiated and signed off the run. Auditability is affected because the run must be examinable later. Interpretability is affected because the relationship between instructions, evidence, and output must be intelligible. Dependability is affected because repeated omissions may reveal an unstable process. Traceability is affected because each artefact must connect back to the specific run.

In governance-readiness terms, reconstructability improves the organisation?s position because the disputed case can be analysed as an evidence-based event rather than as a vague recollection. The evidence pack can be assembled credibly, the score profile can be justified, and lessons can be fed back into prompt controls, review thresholds, and retention policy.

Detailed link to RAIDT

Reconstructability links to RAIDT in four ways.

First, it supports RAIDT?s core idea that governance should attach to what happened in one actual use event rather than to abstract claims about systems in general.

Second, it is inseparable from the run because the run is the unit of governance and reconstructability is the test of whether the run remains examinable after completion.

Third, it strengthens both the evidence pack and the RAIDT score profile by ensuring that evidence-based judgement is grounded in recoverable run records.

Fourth, it supports reviewability, contestability, audit readiness, and organisational learning because a reconstructable run can be revisited, questioned, compared, and improved.

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

Link to the five RAIDT pillars

Responsibility

Reconstructability supports Responsibility by showing whether accountable roles, decisions, and review actions can be tied back to the run rather than left implicit.