DarkMatch
Match people, not records.
DarkMatch finds the same customer across nicknames, typos, moves, married names and missing fields, and keeps apart the people who only look alike. It layers on top of your current matching and recovers the connections rules leave behind.
June 2026 benchmark
1,000,000 records, 250,000 real people, identical input to both systems.
Precision when matching: 98% vs 96%. Methodology
DarkMatch is DarkMath’s identity resolution product. It combines rules-based matching, probabilistic scoring and semantic matching to decide whether records describe the same person or household, then returns resolved IDs with confidence scores, sources and an audit trail of every merge.
Same name, same street, two different people.
Rules fail in two directions: they miss records that belong together, and they merge people who don’t. Step through what each method does with five ordinary records.
Three stages, each catching what the last one missed
DarkMatch runs as a cascade. Cheap, certain decisions happen first. Expensive judgment is saved for the records that need it.
In the June benchmark, rules-based matching alone found 55% of real connections; adding the semantic layer raised that to 89%, on the same records.
Augmented rules
Rules-based matching with learned weights and confidence thresholds instead of a blunt pass or fail.
Probabilistic matching
The Fellegi-Sunter model, fit with expectation maximization, estimates match probability where no shared key exists.
Semantic matching
Records become vector embeddings, so variants with the same meaning converge. Language-model review is applied only to the hardest boundary cases, where uncertainty is highest.
What DarkMatch returns
One Golden Record per identity: a single, auditable profile with every merge explained.
Clean records
Duplicates removed, names and addresses standardized, bad records flagged.
People and households
Individual and household IDs that stay stable as your data changes.
Fresher contact data
Address and contact updates from licensed reference data, where permitted.
Confidence and sources
A score, the contributing sources and the reason for every link.
Embeddings
Vector representations ready for search, models and DarkMirror audiences.
Net-new coverage
A report of the identifiers we can add to your file, signal by signal.
Test it on your data before you buy.
The benchmark is ours. The test is yours. Most match tests run in about two weeks, and the report is yours to keep.
Test sets come from the Corruptor, our controlled perturbation engine. It introduces realistic errors such as typos, transposed digits, nicknames, format changes, missing fields and moves into copies of records with known answers, so every method can be scored against the truth.
Send a sample
Hashed or raw values, your choice. Vector-only is available too.
We build known answers
Controlled errors go into copies with known identities. Your originals are never changed.
Same input, every method
DarkMatch and your current method or vendor process identical files.
Standard report
Match rate, precision, recall, F1, duplicates found, household sizes and net-new identifiers.
Choose how to connect
Batch files, API, in your environment, or a dedicated clean room.
Plans
From one rooftop to a national program. Tier names, limits and pricing are being finalized.
Business
Self-serve portal, coming soon
- Upload, dedupe and household your customer files
- Files up to about 10 million records
- Standard match report
Pro
Semantic matching
- All three matching stages
- Persistent IDs and contact updates
- API and scheduled feeds
Enterprise
Your environment or a clean room
- In-your-cloud deployment on anonymous IDs
- Dedicated clean room option
- Custom volumes, terms and support
DarkMatch questions
What data formats does DarkMatch accept?
CSV, JSON and Parquet files, plus direct connections. Data can arrive by upload, SFTP, Amazon S3, Snowflake sharing or REST API. No proprietary formats or custom connectors are required.
How does DarkMatch handle the John Smith problem?
Common names at shared addresses break rules-based matching in both directions. DarkMatch compares the whole profile, including contact history, vehicles, moves and generation signals, so it links a son’s scattered records while keeping his father separate.
Can DarkMatch work with our existing matching system?
Yes. It is designed to layer on top of your current method. Many teams run it on the records their system could not resolve, or as a validation layer, before moving more matching logic over.
How do you prove accuracy?
With known answers. We introduce controlled errors into copies of records whose true identity is known, run the same input through each method, and score precision, recall and F1. Originals are never changed.
Does DarkMatch need personal data?
Not always. Tests can run on hashed values, and vector-only ingestion lets you convert records to embeddings inside your environment so raw personal data never leaves it.
What is a household?
A group of people who share a residence and often share purchase decisions. DarkMatch builds households from resolved individuals, so a household is only as accurate as the identities inside it.
Find out what your current matching misses.
Send a sample. Get a standard report on your own records.