## TL;DR
Normalize first, dedupe second: lowercase, trim, strip legal suffixes, and canonicalize domains before any comparison. Require a second signal such as a domain or address match before merging on name similarity alone. Different companies stop merging, and every merge stays reversible.

```text
enrichment dedupe ran before normalization and agent merged two different companies
```

## Use this when
- dedupe merges companies that are clearly different
- "Acme Corp" and "Acme Consulting" end up as one record
- normalization happens after or alongside dedupe
- merged records need to be unmerged and reviewed

## Not for this skill when
- true duplicates are not being caught (a recall problem)
- you need to choose a fuzzy matching library
- the data source itself conflates the companies

## Steps
1. Normalize before dedupe: lowercase, trim whitespace, strip punctuation and legal suffixes (Inc, LLC, Ltd), and canonicalize domains (strip www, lowercase). Expected: "Acme Inc." and "acme" become the same key.
2. Dedupe on the normalized key with blocking (for example same domain or same normalized name prefix) before any fuzzy match. Expected: different companies stop colliding.
3. Keep a merge audit trail: which records merged, on what key, and a confidence score. Expected: bad merges are reversible and reviewable.
4. Require a second signal (domain match or address match) before merging on name similarity alone. Expected: "Acme Corp" and "Acme Consulting" stay separate.
5. Sample merges weekly for human review and feed corrections back into the rules. Expected: precision improves instead of drifting.

## Variant phrasings
### dedupe merged two different companies
Short version: normalize first, then dedupe, and require a second signal to merge.
### normalization ran after dedupe
Short version: move all canonicalization ahead of the dedupe pass.
### merged records need review and rollback
Short version: keep a merge audit trail with keys and confidence scores.

Compatibility: language-agnostic; works in Python, SQL, or dbt pipelines with plain string functions, no library pinned.

## Why it happens
The dedupe pass compared raw strings, so "Acme Inc" and "ACME" looked different enough to need fuzzy matching, and the fuzzy matcher then over-merged near neighbors. Normalization would have made true duplicates identical and kept distinct names distinct. Running it after dedupe meant it could only clean up the damage.

## Edge cases / pitfalls
- Stripping legal suffixes can merge a parent and subsidiary that share a name; keep the domain signal for those.
- Canonicalizing domains merges regional sites (acme.co.uk vs acme.com); decide per use case whether that is correct.
- Fuzzy matching after blocking still needs a threshold; too loose reintroduces the original bug.
- Audit trails grow large; retain the key and score per merge, not the full record copies.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_hk0tQ72n5oHuu1Db8cc8VQ