Create a match profile

A match profile is the complete configuration governing how one entity type is matched. This page creates one and activates it.

Before you start

• An entity type with its attributes defined. See Model your data.
• Standardization rules configured for the attributes you will match on. The matcher compares canonical values, so standardizing first materially improves results.
• Permissions per step: match-profile.create for the profile itself, then match-profile.blocking-strategy.create, match-profile.comparison-level.create and match-profile.survivorship-rule.create for its parts, match-profile.maintain to build blocking keys and frequencies, and entity.match to run the verification dry-run.

1. Create the profile

The profile's overview tab, where it is created and later activated.

Create the profile with its thresholds and prior. Thresholds are on the log2 evidence scale — see how probabilistic scoring works before choosing them.

Start deliberately conservative. A profile that under-merges leaves work in the review queue; one that over-merges corrupts golden records and needs unmerging.

2. Add blocking strategies

The Blocking tab. Each strategy derives its own keys; a record is a candidate if any of them matches.

Blocking gathers candidate pairs. Each strategy derives a key from one or more attributes; records sharing a key are compared.

Design for recall: a pair no strategy proposes can never match, whatever the evidence would have been. Several narrow strategies work better than one broad one — each catches a different failure mode.

Strategy shape	Catches	Misses
Exact postal code plus surname initial	Typical households	Anyone who moved
Phonetic surname plus birth year	Spelling and transcription variants	Name changes
Email local part	Shared contact details	Records with no email

Use several. Overlap is the point.

3. Add comparison ladders

The Ladders tab. Each rung carries its comparator and two probabilities; the weight is derived from them.

For each attribute that carries identity, define an ordered ladder from strictest to loosest. The engine takes the first level that fits.

Every ladder must end in a catch-all level that matches anything, which is what records a disagreement. Without it, a mismatch contributes nothing instead of counting against the pair.

Mark a level deterministic only when agreement on it alone should settle a match — a national identifier, a passport number. Deterministic levels short-circuit scoring entirely, so reserve them for identifiers you trust without corroboration.

4. Add survivorship rules

The Survivorship tab, where the winning value per attribute is decided.

Decide which value wins per attribute when sources disagree. Attributes with no rule fall back to the engine default.

For multi-value attributes such as email or address, the aggregation strategy preserves every distinct value rather than collapsing to one — usually what you want for contact details.

5. Activate

Activating deactivates any other profile for the same entity type, atomically. Exactly one profile is active per entity type at a time; two would make results non-deterministic.

6. Prepare the data

Before matching, the platform needs blocking keys computed and value frequencies snapshotted. Run profile maintenance to build them.

Frequency data is what powers the adjustment that suppresses common values, so matching before it exists produces systematically worse results.

7. Verify

The Sandbox tab. Score a pair without writing anything, and read the contribution of every attribute.

Run a dry-run match on a record you know has a duplicate. It scores candidates and returns the pairs it would create, writing nothing.

Check that:

• The known duplicate appears among the candidates. If not, the problem is blocking — no threshold change will help.
• Its composite weight lands in the tier you expected.
• The per-attribute contributions look sane: high-value attributes carrying the weight, common values contributing little.

If a known duplicate is absent, add a blocking strategy that would catch it, then re-run maintenance.

Next

• Tune thresholds against labelled data
• Review potential matches
• Match profile reference

---

Last verified against commit 315eb047 (2026-08-03)