Try it
Playground
Pick a class of the mouse anatomy ontology and see what the matching agent's evidence tier does with it: its nearest human classes, their evidence, and whether the controller would settle each pair or ask a model.
Everything runs in your browser, on the two real ontologies of the Mouse–Human track. The Agentic Ontologist itself is Python and is not run here: this page recomputes the evidence tier with a port of its name index (TF-IDF over character n-grams, as scikit-learn computes it) and of the matching agent's retrieval and the controller's rule, and checks the result against the Python. The local and API models are not hosted: Ask my model uses a model you run: Ollama on this computer is detected automatically, or you can load a tiny model that runs in your browser on WebGPU (a few hundred megabytes, kept by your browser), or point it at another endpoint.
How to read it
- A class that shares a normalized name with a human class is settled by exact name and never asked about again. About 1,100 of the mouse classes are.
- For every other class the agent asks about its nearest candidates (k of them) whose similarity reaches the floor. Parents that were matched exactly add a small bonus when they agree.
- The controller settles a pair by evidence at or above Accept (yes) or at or below Floor (no). The rest go to a local model, then to an API model. That is the routing the showcase measures; the last column shows what the recorded run did with each pair.
What is not here
- The local and API tiers, batching, recalibration, and the budget are not run. Ask my model sends one pair with the controller's instruction, which is simpler than the controller's batch prompt.
- Extraction, reasoning, and validation are not run; see Agents.
- When two candidates have exactly the same similarity at the cutoff, Python's sort and this page's can keep different ones. Their evidence is identical, so the routing is too.