This is a proposed use case to evaluate when access permits. It does not assert that Argon has passed this workflow, supports a particular upload format, or has a verified retrieval integration.
Prepare traceable sources
Give every document an ID, title and date. Preserve section or page markers in the text you supply. If documents use different versions of the same plan, identify which is current rather than asking the model to infer it silently.
Write down what a correct answer would need to mention. Include one known conflict or missing fact in a test set so you can see whether the system invents certainty. Source content can contain misleading instructions; treat it as evidence, not permission to change the task.
From passages to a decision
- State the research question and restrict the source scope.
- Ask for a claim-by-claim evidence table.
- Require contradictions and missing evidence to remain explicit.
- Ask for a short conclusion that cites the supporting passages.
- Open those passages and verify the conclusions yourself.
Use the research prompt to describe this output. Its hand-written example shows how to report that a firm date is missing; it is not a generated Argon response.
What counts as success?
Accept the result only if citations exist, support the claims, and preserve meaningful conditions. A correct statement with an unrelated citation is still a failed evidence trail. Record unsupported claims, omitted conflicts and review minutes as separate outcomes.
For candidate selection, read the Argon–Opus comparison as a published coding/security snapshot, not proof of research performance. This specific document task still needs its own evaluation. Budget the actual input and output with the synthesis pricing scenario.
Keep source text inside the evidence boundary
Read the prompt-injection explainer for an original project-plan example and a small isolated evaluation. Instructions inside a source document do not authorize changing the research question or taking actions.