Lesson 04 — Three suppliers in twelve disguises

Implant supplier names drift across spelling variants — what looks like ten suppliers is often three

Hospital implant catalogues — and the billing extracts that reference them — typically record supplier names freehand at charge time. There is no dropdown bound to a supplier registry. Documenters type what they remember, in whatever spelling convention is muscle memory that day. Over a few years the same supplier accumulates spelling variants.

Five common drift modes:

• Bracket-style toggling: Acme Schweiz GmbH ↔ Acme (Schweiz) GmbH
• Legal-form fluctuation: XYZ GmbH ↔ XYZ AG (sometimes tracking a real corporate change, often an entry error)
• Whitespace and punctuation: Johnson & Johnson ↔ Johnson&Johnson ↔ Johnson+Johnson ↔ Johnson + Johnson
• Casing: GmbH ↔ Gmbh ↔ gmbh
• Department suffixes: Vendor ↔ Vendor CRM ↔ Vendor Implantat

What it does to the analytics

Any analysis that counts distinct suppliers per material — or measures supplier concentration — overcounts the supplier base by 1.5–3× depending on the catalogue’s age. A material that looks multi-sourced may actually be single-sourced from one vendor whose name was entered three different ways.

Two specific signal pathologies follow:

• signal_multi_source_materials over-fires: a single-supplier material under three spelling variants is reported as 3-source.
• thesis_supplier_concentration_risk under-fires: a “diversified” supplier base measured at the catalogue level may be three suppliers in twelve disguises. The risk concentration that the thesis is supposed to surface is hidden by the noise.

What changed in the analysis

The fix is data-side, not workflow-side. Hospitals cannot ask busy OR teams to type vendor names consistently — that’s a speed/accuracy tradeoff they always lose at the point of charge.

The solution lives between charge events and analytics: a deterministic mapping from raw spelling to canonical supplier identity, maintained as a Subject Matter 50cents per tenant and applied in silver before downstream signals run. Three pieces:

1. Family detection in silver. Group supplier names by lowercase first word; surface families with ≥ 3 distinct variants as drift candidates.
2. Canonical-name mapping per tenant. A small per-tenant YAML maps each known variant to its canonical form. Maintained by analysts, not surgeons.
3. Signal application post-mapping. Multi-source counts and concentration measures read the canonical names, not the raw ones.

What stays interesting

A new spelling variant appearing for an established supplier may indicate a real corporate change — acquisition, legal-form switch. Treating every variant as drift can mask a real signal. The mapping discipline knows the difference because the variant is deliberately added, not auto-collapsed.

Why this is in a Lessons journal

Because the catalogue’s structural truth — supplier identity — is not in the catalogue. It is reconstructed downstream by analysts. The data-entry workflow that produced the catalogue could not afford the time cost of canonical lookup at charge time. The reconstruction must happen somewhere; the question is where in the pipeline it lives. The Lesson is that it lives in silver, codified, per-tenant, deliberately maintained — not as a one-time cleanup but as a standing discipline.

Sources

• Wisdom: smebit_implant_supplier_spelling_drift.yaml
• Provider: a hospital procurement director with medical-supplies oversight
• Anchor signal: signal_multi_source_materials
• Anchor thesis: thesis_supplier_concentration_risk
• Date Wisdom captured: 2026-05-08