Assign tiers and apply monotonicity + range filtering to all calibrated probes.

Query one landmark and judge. Returns (landmark_name, correct, response).

Run one probe through all 6 landmarks in parallel.

Create a judge function using a local Ollama model.

Effective capacity ratio: 10**(predicted - actual log_params).

Get probes that haven't been tested on this landmark yet.

Leave-one-out median multiplicative error in predicted params (10**slope*acc+intercept).

Query OpenAlex for author info. Max 2 retries with backoff.

Score a researcher probe response. The response should mention the researcher's field. Accept if: - Response contains the gold field or a

Use an LLM as judge for complex matching cases. This is the fallback when simple matching fails. The judge LLM evaluates whether the response

Search the web for entities matching a tier's frequency band, then build probes from the discovered facts. Args: client: OpenRouterCl

Record a calibration result for one probe on one landmark.

Test a model in both normal and thinking modes.

Test one probe on one landmark model. Returns (correct, response_snippet).

Test a single probe on a single model.

Escape special LaTeX characters in a string.