A model selector that can abstain: a contributor walkthrough of Koto

Description and use

Koto adds a local-first model selector to a coding client, enabling selection of a local model or abstention when none is suitable. The example shows the classifier’s behavior and abstention logic.

Problem

Determine whether a local model can be selected for a coding task and abstain when no suitable local model exists.

Outcome

The synthetic example demonstrates that Koto can correctly select a local model when evidence exists, abstain when evidence is insufficient, and refuse an ineligible model. It does not execute the model or perform any network calls.

Workflow

• Collect prompt history and optional feedback
• Score the prompt against retained examples
• Classify using a nearest‑neighbor model
• If a local model is eligible and matches the prompt, route to it
• If no eligible local model matches, abstain and keep the commercial baseline

Evidence

The question is small enough to test: when checked examples support a local model for a familiar task, can the selector recommend it—and otherwise leave the permitted commercial baseline in place?

Source

The first query matches an eligible example. The second lacks enough matching evidence. The third matches a model that is not eligible. None is the selector abstaining; it does not itself execute a model or authorize a fallback.

Source