Skip to main content
Origin uses one LLM for attribution — the model that reads the trace, the score, and the rubric, and returns the ranked culprit list. This is different from the model your agent pipeline runs on, which Origin does not control.

​
Default: Anthropic Claude

Claude is included when you pip install aevyra-origin. No extra flags needed. 
pip install aevyra-origin
export ANTHROPIC_API_KEY=sk-ant-...
The default model is claude-sonnet-4-5. It’s fast, capable, and well-suited to the diagnostic reasoning Origin requires.

​
Provider format

Models are specified as provider/model — the same convention as aevyra-reflex: 
from aevyra_origin.llm import anthropic_llm, openai_llm

llm = anthropic_llm(model="claude-sonnet-4-5")          # Anthropic
llm = openai_llm(model="gpt-4o")                        # OpenAI
llm = openai_llm(model="qwen/qwen3-8b",                 # OpenRouter
                 base_url="https://openrouter.ai/api/v1",
                 api_key=os.environ["OPENROUTER_API_KEY"])
llm = openai_llm(model="qwen3:8b",                      # Ollama (local)
                 base_url="http://localhost:11434/v1",
                 api_key="ollama")  # pragma: allowlist secret

​
Provider reference

ProviderExtraEnv var
Anthropic(included)ANTHROPIC_API_KEY
OpenAI[openai]OPENAI_API_KEY
OpenRouter[openai]OPENROUTER_API_KEY
Together AI[openai]TOGETHER_API_KEY
Groq[openai]GROQ_API_KEY
Ollama[openai]—
pip install aevyra-origin[openai]   # adds OpenAI-compatible providers
pip install aevyra-origin[all]      # everything

​
CLI provider format

The CLI uses provider/model strings via --model: 
aevyra-origin diagnose trace.json --score 0.4 --rubric rubric.txt \
  --model anthropic/claude-sonnet-4-5

aevyra-origin diagnose trace.json --score 0.4 --rubric rubric.txt \
  --model openrouter/qwen/qwen3-8b

aevyra-origin diagnose trace.json --score 0.4 --rubric rubric.txt \
  --model ollama/qwen3:8b
Bare model names are also accepted — claude-sonnet-4-5 infers anthropic, anything else infers openai.

​
Choosing a model

For most traces, a mid-tier model like claude-sonnet-4-5 or gpt-4o is sufficient. Use a stronger model (claude-opus-4-6) when the trace is long or the rubric has many competing criteria. For quick debugging runs, a fast local model (Ollama qwen3:8b) works well. Attribution quality depends more on clear rubric writing than on model choice — a specific, criterion-by-criterion rubric outperforms a vague one regardless of model.