• workshop_dense_100#slide 53

In RAIDT, an output hash is a cryptographic fingerprint of the generated output recorded within the run-level evidence pack. It is not a substitute for the output text; rather, it is an integrity marker that allows reviewers to verify that the stored output linked to a run has not been silently changed. The Foundations paper lists Output hash | SHA256:... | Integrity of recorded output, and the Evidence Review paper places output hashes within the broader requirement for output integrity and retention. This fits RAIDT's basic logic of run as the unit of governance: one configured use should be reconstructable, reviewable, and defensible through evidence rather than memory.

A concrete RAIDT-style example is a hospital discharge-summary run where the evidence pack records the prompt template version, model deployment ID, retrieval snapshot ID/hash, output text, output hash, and clinician review decision. If the organisation later audits that run, the output hash gives a stable pointer to the exact generated artefact that was reviewed. This is especially important when influence methods as governance interventions are in play, for example structured prompting plus retrieval augmentation, because reviewers must be able to connect one particular output to one particular governed configuration.

Why it matters is straightforward: the output hash strengthens Auditability and Traceability within the five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability), supports privacy-aware evidence handling when content access is restricted, and makes the score profile more defensible. In practice, it helps reviewers assess whether a run is closer to the anchors 1=missing / 3=partial / 5=audit-ready on evidential completeness and integrity.

A public-service eligibility advice tool may generate a draft answer explaining whether a claimant appears eligible under current rules. The organisation can preserve the generated answer in secure storage, but the run-level evidence pack should also keep an output hash tied to the retrieval snapshot IDs, prompt version, model deployment, and reviewer action.

Suppose a later complaint alleges that the advice originally shown to the claimant differed from the version produced for audit. The output hash allows the review team to test whether the retained artefact is the same generated output associated with that run. Combined with the stored rule-text snapshot and oversight record, it turns a disputed case into a reviewable evidence object rather than a debate based on recollection. That is precisely why output hashes matter in RAIDT's score profile and governance practice.

• 08-RAIDT_Foundations_M_V50
• 13-RAIDT-Evidence-Review_M_v10
• 18-RAIDT-Technical-Foundation_M_v04