Every question goes to the cheapest model that can settle it
The Agentic Ontologist orchestrates four memory-augmented agents to integrate two ontologies: KROMA matches classes, extraction finds missing relations, reasoning places classes that have no match, and validation removes matches that contradict each other. This is one recorded run on the OAEI Mouse–Human anatomy track, with every decision and what it cost.
How a question gets settled
A controller tries the free tiers first and pays for the API model only when they are not sure. Below, for the matching agent, how many questions each tier settled and how often it was right.
Evidence first
Exact label and synonym scores settle the easy pairs with no model at all.
Then a local model
What evidence cannot settle goes to a model running on this machine, at no cost.
The API last
Only the hardest questions reach the paid model, in batches, until the budget is spent.
Matching questions by tier
Tallied from the recorded decisions of the routed run. Right means the reference alignment agrees.
Questions that reach a later tier are the ones the earlier tiers were least sure about, so the tiers are not equally hard. See Results and cost for quality against spend for every routing policy.
The workflow
One controller, three tiers, and four agents. Each card shows the agent's headline result and how its questions were settled.
Controller
Settles each question at the first tier that is sure enough, in batches, and stops paying for the API tier when the budget runs out.
Every decision
The two classes, the evidence, the answer, the tier that settled it, and whether the reference agrees. Filter by agent, tier, or outcome.
Results and cost
For each agent, quality against API spend for every routing policy. The routed policy is highlighted. Hover a point for its numbers.