Maintain traceable logs for inputs, outputs, and model behaviors to support reviews and audits.
Records what was sent, what came back, under which model and prompt version, with what result. The foundation every other solution in this group depends on: without the record, nothing can be evaluated after the fact.
Every production deployment without exception. There is no responsible configuration in which this is optional.
Not applicable — though the retention and masking policy must be designed rather than defaulted.
Logs contain whatever was in the prompt, so they inherit the sensitivity of the source data and require the same protection, residency and retention treatment as the records themselves.
Logging everything including personal or regulated data, creating a secondary store with weaker controls than the primary one. The audit trail becomes the breach.
A generated journal entry narrative logged with its model version, prompt version and input records, so that six months later an auditor can establish exactly what produced it.
Two different absences share this shape. Foundational solutions get built whatever the domain, so no domain links them; the rest are solutions this domain genuinely does not reach for. v_ai_solutions_unlinked separates the two.