Supplier-data extraction: calibration, validation and review

Supplier-data extraction was part of my historical work on Bridge Sourcing, an own project that is paused as of October 2026. This note describes calibration, validation and review methods from that work. The original numerical headline has been retired because its cohort, denominator and measurement window have not been reconciled for publication.

01 · Define the output before choosing the model

Supplier pages mix contact details, product descriptions and incomplete claims. Give each field a definition, preserve its source, and distinguish missing information from a guessed value. A useful record is one a reviewer can verify.

02 · Build a labelled calibration set

Collect representative pages and review the expected fields manually. Include missing fields, ambiguous names and formatting variations. Keep that fixture set versioned so a change can be compared against a known input.

03 · Validate outside the model

Apply deterministic checks to structured output. Check required fields, types and formats; retain an unknown value when the source does not support a claim. Validation belongs in the application, alongside retries and explicit failure handling.

04 · Preserve evidence for human review

Attach the relevant source to extracted fields. Escalate ambiguous records for review rather than silently treating a plausible answer as verified. Review decisions should be recorded separately from generated output.

05 · Report comparable results

A numerical accuracy claim needs a defined field population, scoring rule, denominator, sample and date. The methods above are retained as engineering history. This note makes no current accuracy, supplier-quality, completed sourcing transaction or client-outcome claim.