Odd Ecosys: grounded assistance
The intended input is a reviewed lead fact bundle: relevant business information, selected website observations, and the user’s campaign constraints. The intended output is a structured draft that can be checked, edited, and approved by a person.
The current code includes local-model drafting and an optional AWS Bedrock Claude adapter for subject-line refinement. This is implementation evidence, not a claim that a production Claude service is currently enabled.
Direct Claude API: planned evaluation
We plan to evaluate the first-party Claude API for source-grounded message drafting, structured review, and explanations of unsupported claims. The application should record which provider produced an output and preserve errors or missing evidence explicitly.
Verdict: a measured research role
A potential Claude integration would provide a reasoning or review role inside the experimental panel. We would compare its contributions against alternative configurations, including disagreement quality and verified task outcomes. This is a research direction, not an existing benchmark result.
Data and approval boundaries
- Send only the context required for the selected workflow.
- Keep provider credentials on the server.
- Make the provider and data flow clear to the user.
- Require human review before consequential outbound actions.
- Keep unresolved claims visible instead of presenting them as facts.