Language models

The engine never needs a model. Where one helps, it is given the contested pairs or the open reviews, with the evidence for each, and its answers are recorded like any other decision: which model, the exact request, the response, tokens and cost.

pip install 'thisness[ui,openai]'      # or [ui,anthropic]
export OPENAI_API_KEY=...              # or ANTHROPIC_API_KEY

Two roles

The adjudicator answers contested pairs during resolution: same, different, related or unsure. In decide mode a "same" becomes a strong edge and a "different" a negative one. A model's "same" yields to its own "different" in the same component and to a person's cannot-link, so one wrong answer cannot pull two entities together over the model's other answers. In shadow mode answers are recorded but change nothing, so a model can be measured before it is trusted.

The review assistant suggests an answer to an open review: same, different, related, a regrouping, or unsure, with a rationale that cites the board's labels. A person accepts or overrides it.

Each profile states which role a model plays (model_role): organizations and taxa decide, people suggest. --model-role overrides it for a run.

thisness resolve --provider openai --model gpt-6-luna --budget 2
thisness recommend --provider openai --model gpt-6-luna --limit 50
thisness ui --provider openai --model gpt-6-luna      # Ask the model, from the Workbench

What a model is sent

The prompts are domain-neutral, with each profile's guidance appended. The review assistant sees records and comparisons under short labels (R1, C1), the same ones the Workbench shows. Attributes a profile marks local_only (a person's date of birth, an address line) are withheld from external providers; contact details go out hashed. With --local, the OpenAI-compatible server is yours and they may be sent.

Cost and control

• --budget stops spending at a limit. There is no default limit, so set one.
• --record cassette.jsonl keeps every exchange; --replay cassette.jsonl answers from it, free and offline, which is how evaluations are re-run.
• Transient errors are retried three times with backoff; a permanent error is not.
• --concurrency sets requests in flight.

Measured

On real UK companies (400 sampled, 3,507 records, identifiers hidden):

As a review assistant on the same benchmark's 177 open reviews ($0.11), 93% of its decided suggestions would have been right if accepted, and accepting all of them would have merged different companies in none. Judging a whole review, a model is far more careful than judging one pair. The numbers and their caveats are in design 07.