The Workbench · Thisness

pip install 'thisness[ui]'
thisness ui identity.duckdb

The Workbench runs on your machine against a store file. Nothing leaves it unless you configure a model provider. Its layout follows Twenty's record table, show page and side panel, and its colors and type are Twenty's design tokens. The explorer on this site is a read-only copy over real exported data.

Entities

Each entity type has a list, largest first, with open-review counts and search. An entity's page shows:

Values. Every attribute's values, how many observations support each, and from which sources. The value shown is the best supported, so "Zalbalt Consulting LLC 24/29" reads as twenty-four of twenty-nine observations.
Observations. The member records with their sources. Click one for its own history.
Relationships. Recorded relationships and possible ones, with the signals behind them.
History. The ledger for this entity, one item per commit: created, observations joined, merged, split, related. Click a commit for the decisions it cites.
Context. Exactly the text a language model would be given about this entity, within a budget, with local-only attributes withheld.

The review queue

Reviews are ranked by impact. A review is a board: one column per current entity, one card per record (R1, R2, ...). Drag cards between columns to regroup, then answer:

All the same entity, Different entities, Related (with a relation), or Save grouping.
Every answer needs a reason and is stored with it. A wrong answer is undone like any other commit.

Below the board, the comparisons (C1, C2, ...) are listed strongest first, each with its levels and weight. Why this was raised shows the decision that opened the review.

If a model is configured, Ask the model requests a suggestion from the review assistant. It appears above the board with a rationale that cites the board's own labels ("R1 and R3 share a postcode, C2"), and the comparisons it relied on are marked. Apply to the board moves the cards to its grouping for you to confirm. Whether reviewers' answers matched the suggestions they were shown is recorded (engine.assistant_agreement()).

Possible relationships

Entity pairs linked by relationship signals (a shared brand, a group's sibling companies, a household) and not yet related. They are grouped into neighborhoods of at most thirty entities, largest first, fifty to a page. Open a pair to relate the two entities, or to merge them if they are the same.

Model calls

Every model invocation: what was sent, what came back, tokens, cost and latency, with totals per run. With --record, the same exchanges are in a cassette you can replay.