Model Routing

A scientific working group does not need every role to use the same model route. In this container, the PI Liaison and Scientific Director should stay on the most reliable human-facing route available, while narrower specialist agents can be evaluated against open-model APIs.

The practical pattern is conservative: keep the user-facing and decision-setting roles on OpenAI/Codex OAuth or another approved high-reliability route, then experiment with hosted open models for bounded work such as data inventories, literature triage, first-pass technical drafts, and skeptic checklists. This follows the same bounded-role idea used throughout the working group: agents get specific jobs, expected outputs, and explicit limits.

The workspace seed includes /workspace/MODEL_ASSIGNMENTS.md as the role-level routing register. Update that file when a role changes model route, and record meaningful changes or reversions in /workspace/DECISIONS.md.

Suggested Defaults

Role group	Recommended route	Why
PI Liaison / User Interview Agent	OpenAI/Codex OAuth or another approved high-reliability route	This role talks to the user, manages Slack intake, and batches decisions.
Scientific Director	OpenAI/Codex OAuth or another approved high-reliability route	This role sets scientific direction and decides when claims can advance.
Deputy Director / Integrator	High-reliability route first, open-model experiments after evaluation	This role maintains coherence across the team.
Data, modeling, citation, communication, and skeptic roles	Approved open-model API route allowed for bounded tasks	These roles can be tested with narrower prompts and reviewable outputs.
Societal Impact / Translation Agent	High-reliability route preferred	Sensitive claims about communities, policy, public health, legal rules, or sovereignty require careful review.

Verde-Style API Experiments

For Verde/CyVerse experiments, use the OpenAI-compatible base URL https://llm-api.cyverse.ai/v1 and keep credentials local in .env. The AI-VERDE API documentation explains that keys are obtained from the AI-VERDE course or team details and that available models can be listed from the API. Do not commit keys, paste them into chats, or store them in markdown:

VERDE_LLM_BASE_URL=https://llm-api.cyverse.ai/v1
VERDE_LLM_API_KEY=
VERDE_LLM_DEFAULT_MODEL=
VERDE_LLM_PROVIDER_NAME=verde

After setting a local key, inspect the models your team can access:

curl -s -L "${VERDE_LLM_BASE_URL}/models" \
  -H "Authorization: Bearer ${VERDE_LLM_API_KEY}" \
  -H "Content-Type: application/json"

Current AI-VERDE Candidate Models

The following model IDs were reported as available for the current API key context on 2026-05-18. Availability can vary by course, team, quota, and provider changes, so confirm with the /models endpoint before assigning a role:

Model ID	Candidate use
nrp/glm-4.7	General specialist-agent evaluation
nrp/glm-5	General specialist-agent evaluation
nrp/kimi	Literature triage, summaries, and long-context checks
nrp/gemma	Lightweight drafts and comparison tests
nrp/qwen3	General specialist-agent evaluation
nrp/qwen3-small	Fast, low-risk smoke tests and simple routing checks
nrp/minimax-m2	General specialist-agent evaluation
nrp/gpt-oss	Open-model comparison tests
Meta-Llama-3.1-70B-Instruct-quantized	Larger-model baseline comparisons
llama-3.3-70b-instruct-quantized	Larger-model baseline comparisons
phi-4	Lightweight technical drafts and checks
Llama-3.3-70B-Instruct-quantized	Larger-model baseline comparisons
phi-4-multimodal-instruct	Multimodal experiments only when inputs and privacy are reviewed
gemma-3-12b-it	Lightweight drafts and comparison tests
Llama-3.2-11B-Vision-Instruct	Vision experiments only when images are appropriate to share
esiil/GLM-4.7-Flash	Fast ESIIL-context experiments and smoke tests
js2/llama-4-scout	Jetstream2-hosted model experiments
js2/gpt-oss-120b	Larger open-model comparison tests

Do not treat this table as approval to use a model for sensitive claims or external actions. It is an inventory for bounded experiments.

These variables are intentionally generic placeholders. They make the container ready for local experimentation, but they do not guarantee that OpenClaw has registered a provider. Provider registration should be added only after checking the installed OpenClaw version and documenting the exact provider name, model ID, validation command, and fallback route.

Evaluation Workflow

1. Pick one bounded role and one bounded task.
2. Run the task with the current reliable route.
3. Run the same task with the open-model candidate.
4. Compare evidence handling, citation behavior, role compliance, and uncertainty language.
5. Record the result in /workspace/agent_reports/model_evaluations.md or /workspace/DECISIONS.md.
6. Promote the route only after human approval when the role is user-facing, Slack-facing, director-level, or involved in sensitive claims.

Scaling Pattern

The scalable unit is not "one key per agent." The scalable unit is a reviewed route assignment: role, provider, model, allowed task types, blocked actions, evaluation date, and fallback. That keeps the system reproducible as users add local keys, hosted endpoints, or future OpenClaw provider integrations.