ROUTEAgents & automation
Jev Classifier Layer Model Router
Bill Watson@xbillwatsonx
This project aims to build a model router using Jev within the classifier layer, with an implementation plan that includes plugin development, testing, and phased activation.
TakeJev questions require typed choice, score, or noul objects.
- Progress -
I provided a bunch of research on Jev to OpenClaw and asked it to study the resources and build an implementation plan with the goal of building a model router using Jev in the classifier layer.
Here is Alex's response:
---
I finished the research and attached the reviewed implementation plan.
My recommendation: build a trusted OpenClaw plugin using the documented before_model_resolve hook. Scope it to a dedicated jev-router-lab agent and begin in shadow mode.
Key findings:
The Gemini example is conceptually useful but not runnable.
OpenRouter uses POST openrouter.ai/api/alpha/decisions.
Jev questions require typed choice, score, or noul objects.
Responses arrive under answers, not Chat Completions choices.
We should use OpenRouter’s latest alias for exploration, but pin the resolved version during evaluation.
Jev should choose only among allowlisted tiers. Code retains control over privacy, tools, approvals, and risky work.
Short context-dependent replies should abstain and inherit the conservative default.
Proposed initial tiers:
Routine: ollama/glm-5.3:cloud
Standard: openai/gpt-5.6-sol
Complex: openai/gpt-5.6-terra
Uncertain: no override
Astra remains human-directed, not automatically selected.
-----
Video processed to REFERENCE-VAULT
Reviewed packet: "file name redacted"
Raw transcript and metadata: "file name redacted"
Transcript: 758 timestamped lines
-----
Implementation plan:
/home/xbill/programming-projects/agenthelpsite/private-project-notes/2026-09-19-jev-openclaw-model-router-implementation-plan.md
The plan includes nine phases: plugin/client build, mocked tests, a 90-task labeled evaluation set, deterministic-baseline comparison, shadow traffic, controlled activation, independent review, and evidence-backed OP-078 revisions.
I intentionally left the public guide unchanged. We’ll update it after the router produces real evidence. No paid Jev requests or live OpenClaw changes were made.
---
I found three or four discrepancies in the implementation plan. My plan now is to point them out, run one more pass to refine it again, read it again and then get started.
Here we go. Wish me luck. I'll keep you posted.