A language model reads the hard pairs a rule leaves open, drafts standardisation and enrichment, writes rules from a sentence and answers questions in plain words.
| Job | What the model does | What a person does |
|---|---|---|
| Hard matches | reads two records with their context and says whether the two describe one company, and why, with a confidence | approves merges of the kinds that matter; sets the thresholds |
| Standardisation | knows that GmbH is a legal form, that St. can be Street or Saint, that a Japanese name leads with the family name | reviews the exceptions the validator flags |
| Rules from sentences | turns 'open orders must stay under the credit limit' into a rule and shows the rows it would flag | reads the rows, adjusts, saves |
| Enrichment | proposes a registry number, an industry code, a parent company from public and licensed sources, with the receipt | approves per source and attribute; the rule then applies within its range |
| Anomalies | flags a lead time that drifted from actual receipts, a value that changed too often, a supplier whose pattern changed | decides what to do; the flag becomes a rule if it recurs |
| Questions | answers 'open orders for Acme in Q3 by plant' with governed SQL and citations | reads the answer with its rows; saves the query |
The ladder is per action type, per domain. A merge stays at level 0 until the steward chooses otherwise; a spelling fix reaches level 1 in a week.
A model needs the shape of the problem: field names, patterns, a rule text, an anonymised pair. Lima sends that. It sends no row data to an external model. Where a company wants a model to read rows, the model runs inside the estate, and the same boundary applies to it as to any user: the twelve security layers and the audit trail.
The trust centre states the boundary in one line: metadata can cross, row data stays.
Any that the company approves: a hosted frontier model for structure and metadata, or a model running in the estate for row-level work. The choice is per action type.
Every proposal carries its reasoning and its evidence, and the decision is logged with it.
Yes, per action type and per domain; the rules and the queue keep working.
Bring two systems that hold the same customers, suppliers or materials. In one working day we connect them, run the match on live rows, review the queue with your steward and publish the first golden view with its API.