Dot Loom
An open, provider-pluggable runtime that decides how much inference an AI request needs, spending extra model calls on verification only for higher-risk work and reporting the policy, cost and latency for every run.
Dot Loom is not a model itself; it sits in front of one, choosing between a direct answer, a verifier-editor pass, or a longer draft-critic-finalizer chain depending on estimated risk and a configured cost ceiling. Every run records the policy chosen, its call budget, the reason for any escalation, cost, latency, token counts and a receipt.
The same workflow runs against its own Dot models, OpenRouter, other OpenAI-compatible endpoints, Ollama, LM Studio-compatible local servers, or deterministic mocks for testing. It ships an evaluation harness with adversarial backend and code-review test suites plus a mixed-difficulty routing suite, independent strategy-blinded model judging, and shareable HTML, SVG, Markdown and JSON benchmark reports.
The project describes itself as a testable alpha runtime suited to experiments, demos and architecture review rather than a production replacement for commercial multi-agent systems.