Every rule in Lima is a small agent with a scope, a schedule and a level of autonomy it has earned.
The capability pages explain how golden records are assembled, structured, governed and served. This section explains how they are kept correct.
A golden record is assembled from its sources once; it is kept correct every day. That work is a loop: measure the data, check it against rules, propose fixes, decide, apply and write back, learn. Lima runs the loop with agents. A rule agent is the smallest kind: one check, bound to a place and a schedule, written as a sentence. Service agents do the wider jobs: profiling a source, finding duplicates, verifying a value against an outside registry, correcting a field, watching for anomalies.
Agents act within a level of autonomy they have earned. At the first level they observe and report; at the last they act alone within a budget while every guard holds. Promotion between levels runs through a gate: a passing evaluation on golden test cases, a wired escalation route and a human stamp. Every action, at every level, lands in the audit trail with its evidence.
Enforcement closes the loop. A record is in flow while checks run, golden once every applicable rule passes, and quarantined the moment a rule fails later. The grid, the headless API, SQL and the write-back each behave according to that state, and quality contracts with consuming systems say what happens when a threshold is breached.
Completeness, validity, consistency, uniqueness, accuracy and timeliness, each with the question it answers and how it is scored on master data.
Concept →What, where and when in one sentence; the eight blocks every agent has; dry runs on live rows.
Concept →Profiler, duplicate finder, verifier, corrector, monitor: what each does, what it produces, what it may do alone.
Concept →The four levels, the promotion gate, guardrails and policy floors, and how a level is taken back.
Concept →Packs, scope, thresholds, schedules, owners, escalation routes and contracts, and how drift is tracked.
Mechanics →Column scores, cross-source comparison, trends and business impact, before and after rules.
Mechanics →From finding to learned outcome: evidence, proposal, decision, the change, the write-back and its acknowledgement.
Mechanics →In flow, golden, quarantined; the publish gate; what each door does; quality contracts with consumers.
Mechanics →Tax ids, registrations, addresses and bank details checked against outside services, with the evidence stored.
Mechanics →Volume, freshness, structure and values watched on every source; learned patterns; root cause through lineage.
Mechanics →The same rules before save at every entry point, so a record is born clean.
Mechanics →Profile on Monday, enforce on Friday; the scorecard the owner reads on day 30.
How to start →| Already on the capability pages | What this section adds |
|---|---|
| Rules at three levels: the language for column, row and business rules | Each rule runs as an agent with a scope, a schedule, a level and an evaluation; packs stamp rules across a scope |
| Stewardship: the queue sorted by confidence and value at stake | Where the proposals come from, what evidence they carry, and which of them the agents may apply alone |
| AI-assisted: models propose, rules validate, people approve | The four autonomy levels in detail, the promotion gate, guardrails and policy floors |
| Golden records: match, survive, cross-reference, unmerge | The three states of a record and the gate that promotes it; cross-source scores feeding survivorship |
| Audit trail: every change and sensitive read, append-only | Every agent action with its guard verdicts, and the write-back acknowledgements beside each change |
| Integration and sync: write-back through released interfaces | Corrections as the main writer, with the acknowledgement stored per system and per field |
On the showcase estate, 14,200 corrections in a month; the rejected 4% trained the next proposals.
Learned patterns for month-end surges and weekend dips replaced fixed thresholds on the showcase estate.
A $75B semiconductor equipment maker, ahead of a 1,000-table planning integration.
The treasury audit on the showcase estate ran as a query over the ledger, with zero payments to an account changed within seven days.
It is part of Lima master data. The rules language, the steward queue, the audit trail and the connectors are the same ones the capability pages describe; this section adds the agents that run the rules, the scoring, the correction loop and the enforcement gate.
Only at the levels they have earned, per action type, inside guardrails and a daily budget, with every action in the audit trail. Most agents run at the propose level; the ones promoted to act alone got there through a passing evaluation and a human stamp, and the level can be taken back at any time.
Inside Lima's own environment. Definitions can be co-drafted with an outside model, and metadata may cross that boundary; row data stays inside. The AI-assisted page has the boundary in full.
It answers the 'data quality and cleansing' and 'stewardship and governance support' duties of the market definition, with the agent model as the distinctive part; the analyst guide maps the vocabulary.
Bring one domain and two systems. In a week the sources are profiled, the ten costliest fields are under rules, the first corrections are written back with their acknowledgements and the publish gate is on.
