Examples

The repo ships four runnable examples in examples/ — each a single Python file using the SDK against the bundled cards corpus. No database, no API keys.

pip install -e ./sdk-py
python examples/verify_hypothesis.py   # or any example below

Example	Shows	What it does
verify_hypothesis.py	the cross-check engine	A well-formed vs a dimensionally-broken hypothesis → NONE vs HIGH. Walked through in Verify a hypothesis.
browse_cards.py	corpus access	Load, count, filter, and read cards from the corpus.
validate_card.py	schema validation	Parse a valid card; reject a malformed one with structured errors — the in-process version of the ajv-cli check.
use_mcp_tools.py	the MCP tools	Call cards_list / cards_get / ops_get / hypothesis_crosscheck as plain functions — exactly what an MCP client invokes over the protocol.
verify_llm_output.py	LLM in the loop	Ask a model (Ollama or any OpenAI-compatible endpoint) to propose a law, then run the engine on its output. Needs a model endpoint.
usce_check.py	finished-output check	Range-check a finished result’s numbers against a card’s validation envelopes (within → pass, outside → HIGH).

Test a model (Llama or any other)

To run a model through the verification, there are two paths:

• Quick, self-contained — examples/verify_llm_output.py asks a model (local Ollama by default, or any OpenAI-compatible endpoint) to propose a law, then runs the engine on its output. No benchmark prompts needed:

  ollama pull llama3.1:8b           # have Ollama running
  python examples/verify_llm_output.py

• Full A/B benchmark — the eval harness scores a model control (alone) vs treatment (with Lemma tools) over the HumanEval-Sci prompt set:

  cd eval/humaneval-sci
  HUMANEVAL_SCI_PROMPTS_DIR=/path/to/prompts pnpm smoke-ab --ollama --model llama3.1:8b

  Adapters (Ollama, Gemini, Anthropic) live in eval/humaneval-sci/runner/adapters/.

Not covered by these examples

• rag_lookup needs a Postgres + pgvector backend — see the MCP server page.
• To run the server itself over MCP, or browse the corpus with the lemma CLI, see Quickstart and the Python SDK.