The six sessions 🔗 · one customer master, built across all of them
Every session adds a stage to a real matcher, and every number it reports is computed against the labels rather than asserted.
No shared key
Two CRMs, a support system, and nothing that joins them. The five pipeline stages, the pair count that grows as the square, and the labelled truth that makes everything later measurable.
📏 Session 2 · foundationsComparing strings honestly
Seven ways to score how alike two names are, each taught with the case where it is confidently wrong - including where Jaro-Winkler's prefix boost cannot help at all.
⚖️ Session 3 · the coreFellegi-Sunter, and the band nobody implements
The two probabilities that turn a similarity into evidence, and the structural point almost every system drops: two thresholds, with the pairs between them belonging to a human.
🧱 Session 4 · hands-onBlocking
Four ways to throw away most pairs before scoring them, and the three metrics that judge the choice. Every scheme here costs nine percent of the true pairs, permanently.
🎯 Session 5 · hands-onJudging it
Precision and recall on pairs, why this field never reports accuracy, and the two opposite ways transitive closure gets the cluster count wrong.
🛡️ Session 6 · capstoneMaster data, and the linkage you should not make
Golden records, survivorship, and the four styles - taught with their real provenance, which is analyst frameworks rather than papers. Then the linkage you could make and should not.
Your passport 🎫
Clear all three quiz questions in a session and it gets stamped. Stamps live in this browser only.
What is real here 🧾
A course about judging other people's matching has to be exact about its own.
| Claim | Status |
|---|---|
| Every similarity score and match weight | Real implementations running in your browser - Jaro-Winkler with the prefix capped at 4 and scale 0.1, Levenshtein, Damerau-Levenshtein, Jaro, Jaccard and Soundex. Type two strings into the session 2 lab and the seven numbers recompute as you type |
| Precision, recall, F1, pair completeness, reduction ratio, cluster counts | Computed against the 35 labelled true pairs at the moment you move a control. Nothing is scripted and no figure is looked up |
| The two pairs that both score 0.84 | A genuine property of the data, not a staged demonstration. One is Sarah Whitfield after a job change and a move; one is a father and son sharing a landline. Both single-threshold rules measure F1 0.971 and neither can do better |
| The m and u probability table | A judgement, not a measurement. It stands in for what Splink would estimate with expectation maximisation on a real corpus. Labelled as such on the widget itself. Everything computed from it is real arithmetic |
| Blocking costs nine percent of the true pairs | Measured on all four schemes. Postcode, surname soundex and sorted neighbourhood all land at 91 percent pair completeness on this data |
| Master data management's four implementation styles | Attributed to Gartner consistently across five independent vendor sources, with no primary Gartner note obtainable. Taught as an analyst framework, never as an academic citation. "Golden record" has no traceable originating citation either, and "survivorship rules" is vendor vocabulary |
| Sorted neighbourhood, B-cubed, MinHash, and the PETS 2011 cryptanalysis | Real work, cited from secondary sources rather than fetched primaries. Flagged on the page that teaches each |
| PDPA and entity resolution | Not covered. Session 6 reasons about GDPR identifiability and purpose limitation, and says plainly that no GDPR clause names record linkage and that PDPA was not researched |
The knowledge map 🧠
The whole course at a glance - hover a session to spotlight its concepts, click any node to jump in.