Benefits of edi: Software Outcome Claims Index

Every row is a verbatim sentence read from the named source's own page through our fetch service on the date shown, hash-pinned to the stored read (the sha of the page's visible text; the raw page is mirrored to R2).

4 of the 20 named sources that could be read published a usable figure; 16 published none and are recorded as not published, never filled in; 2 could not be read (blocked, dead or unreachable) and count nowhere.

The headline figures (median, minimum, maximum, quartiles) are over ONE figure per source in % (a source's median where its page yielded several accepted sentences), so N counts sources, never sentences; 57 candidate sentence(s) were refused by a reviewer with the reason recorded.

The fuller derivation this dataset is chartered to grow into (not yet what the figures above are; they are each source's own published figure as read): one row per vendor per claimed outcome: the number claimed (cost per invoice, percent touchless, times faster, ROI), the unit, whether a customer count or a commissioned study stands behind it, and the vendor's own name against it - then the derived figure nobody publishes: the SPREAD of claims per category and the distance between the vendor claims and the independent measurement (the European Commission's e-invoicing work, the national statistics), so a page can say 'vendors claim between X and Y; the measured figure is Z' instead of repeating one press release

57 candidate sentences on these pages were refused by a reviewer; each refusal and its reason is recorded in the review ledger.