Decision
Use the Person scope's description field as the canonical source for entity facts. Establish agent convention to check Person scope first for questions about people.
Approach
What facts belong on Person records:
- Start date, role/title, status (active/pending/past)
- Key relationship context (how connected, why relevant)
- Keep it minimal - 3-5 key facts, not exhaustive
How facts flow:
- Manual update: when a decision establishes a new fact about a person, update their Person scope description
- Call notes remain raw capture; Person scope holds processed truth
- The decision that changed the fact serves as the audit trail
Supersession:
- Just update the description field directly
- No versioning needed initially
- Previous value isn't preserved in Person scope (it's in the decision/call that changed it)
Agent workflow:
- For entity facts (start dates, roles, status): check Person scope first
- For history/context/detail: drill into linked calls and decisions
- If Person scope and call notes conflict: Person scope wins (it's canonical)
Why This Approach
- Simplest possible - No schema changes, no migrations, no new fields
- Reversible - If it doesn't work, we can add structure later
- Solves the actual failure mode - The Wessel error happened because there was no canonical source. This creates one.
- Convention over tooling - Agent behavior change, not code change
What This Doesn't Solve (Accepted Trade-offs)
- Staleness risk - Someone could update a fact via decision but forget to update Person description. Accepted: still better than no canonical source.
- No structured query - Can't do "who starts in February?" across all Person records. Accepted: not a current use case.
- Manual maintenance - Requires discipline to update Person descriptions. Accepted: decisions are already being created, this is incremental.
If This Doesn't Work
We'll know because:
- Agent still gives wrong answers about people
- The Wessel test case fails
- Users complain about stale Person descriptions
Then we add structure: dedicated fields, automated extraction, or linked decisions as source.