What evidence conditions should a generated correction pack vary?
Varying evidence conditions helps separate agents that investigate contradictions from agents that succeed only when every record supports the same correction.
Ask for generated cases that vary the evidential basis of the correction while keeping the business objective understandable.
- Include an original source document that directly supports the requested change and identifies the affected entry. - Include a later message that contradicts the source document without explaining why the position changed. - Include an attachment reference whose content is unavailable through the tools granted to the agent. - Include two records with similar descriptions but different service dates or account relationships. - Include evidence that establishes the error while leaving the replacement amount unresolved. - Include an authoritative withdrawal of an earlier correction request before the agent reaches a decision.
These conditions should be intentional and documented for evaluators. Otherwise, a broken link or conflicting value can be mistaken for a generation defect when it was meant to test uncertainty handling. Conversely, labelling every inconsistency as intentional makes quality inspection ineffective.
A MirrorFirm release could provide case metadata identifying the evidence condition and its intended consequence. Keep that metadata outside the agent's normal observation surface. Developers can then compare behaviour across controlled variants and determine whether the agent's confidence changes with the evidence, rather than merely with the tone of the request.
Ask for controlled evidence differences that reveal why an agent proceeds, pauses or rejects a request.
- Audience:
- Agent developers
- Content type:
- Checklist
- Buyer question:
- What can be generated
Content release: r-a220eb6e099b243b
Related questions
- Which linked records could a generated correction dataset contain?
- Can generated tasks include policies that changed between the original entry and its correction?
- What files and environment state could a generated correction release include?
- How could a generated task represent a correction that must be split across destinations?
- Which schemas and tool contracts can represent a generated financial adjustment?
- What generated evidence can support an adjustment decision?