Q173 - What_future_extensions_are_already_visible_in_the_project

Q173 — What future extensions are already visible in the project?

← RAIDT · Star S11 - Boundaries, Limitations and Future Questions · primary item: S11.09 · Future extension: multimodal AI

Appears in sources
Answer

The papers do not yet present a standalone multimodal module, but they do make several future extensions visible in the RAIDT project. Most importantly, RAIDT is designed around the principle that the run as the unit of governance is the relevant object for review. In both the foundations and theory papers, the run-level evidence pack is defined broadly enough to capture prompt records, configuration digests, retrieved context, enabled tools, hashes, timestamps, review logs, and other provenance artefacts for one configured use. Because the framework is centred on reconstructability rather than on text alone, the underlying logic already points beyond single-modality prompting. The same five pillars (Responsibility, Auditability, Interpretability, Dependability, Traceability) and the same score profile can therefore travel to richer evidence settings if equivalent provenance and review traces are preserved.

A second visible extension is theoretical and methodological. The project explicitly treats influence methods as governance interventions, not merely as engineering choices, and it identifies later empirical validation, sector calibration, policy translation, and partial automation of evidence checking as next steps. That matters for multimodal development because it shows RAIDT is intended to absorb new configuration elements without abandoning its design logic. A future multimodal instantiation would therefore be a disciplined extension of the existing artefact: additional media provenance, modality-specific checks, and preserved tool-call sequences would be added to the run-level evidence pack, while governance readiness would still be expressed through the anchors 1=missing / 3=partial / 5=audit-ready. In that sense, the project already signals expansion across domains, standards, and more complex human-AI configurations, even before multimodal AI is named directly.

Practical example

Consider the public-sector eligibility advice scenario already used in the papers. In the current RAIDT logic, a reviewer would preserve the prompt template, retrieved rule text, output, hashes, timestamps, and oversight decision inside the run-level evidence pack, then inspect the score profile across the five pillars.

A future multimodal extension is visible when the same case also includes scanned supporting documents or other non-text evidence supplied by the applicant. The framework would not need to be reinvented. It would need extra provenance fields showing which files were used, how they were transformed into machine-readable form, which tool actions occurred, and what human checks confirmed their suitability. The practical value is that the case remains reconstructable and contestable: the agency can still review one run, compare configurations, and judge whether the added media evidence improved traceability and dependability or simply introduced new ambiguity.

Sources in RAIDT papers
Powered by Forestry.md