What generated records can support a complaint investigation?

MirrorFirm could generate a complaint record together with the material needed to investigate it: customer messages, transaction histories, internal requests, supporting documents and proposed findings. The useful unit is the relationship between these records, rather than a folder of independently plausible documents.

An allegation can link to evidence that supports it, evidence that contradicts it and questions that remain unanswered. A finding can identify the records it relies on and the limits of what those records establish. A proposed remedy can then link to the finding it is intended to address.

This structure can preserve uncertainty. A dispatch scan may establish that a parcel left a depot without establishing that the customer received it. A model should not convert the existence of one operational event into proof of a different event.

A release can include a generated evidence graph alongside readable documents. Labs can use the graph for supervised reasoning tasks, retrieval tasks or investigation environments. The buyer should ask which relationships are explicit in the input and which the model must infer, because that choice materially changes the difficulty and purpose of the task.