Why decision graphs beat documents
June 18, 2026
A document has one fatal flaw as a memory: it can’t tell you when it’s wrong. It sits there looking authoritative long after the decision it describes was reversed. Nobody deletes it, because deleting is work and nobody’s sure.
A decision graph treats each decision as a unit with structure — a claim, the question it answers, the evidence, and its relationships to other decisions. That structure is what lets the system stay honest.
Supersession is the point
When a new decision answers the same question as an old one, that’s a supersession — and the graph records it as an edge, not a deletion. The old decision isn’t erased; it’s demoted and linked, so you can still ask why did we change our minds? and get an answer.
- Live decisions surface first.
- Superseded ones redirect to the current answer.
- Contested ones surface both sides, flagged — because a disagreement is signal, not noise.
Flat context, at any size
The second promise: as the team and the corpus grow, the tokens injected per task should stay flat. We separate a thin awareness layer (always present) from detail the agent fetches on demand — so the floor cost doesn’t climb just because the company did.
That’s the bet: structure beats prose, and honesty beats authority.