Glossary
What is data provenance?
Updated September 14, 2026
Data provenance is the documented origin and chain of custody of a record: which source produced it, under what conditions, and every hand it passed through on the way to the point of use. It is stronger than attribution and different from lineage. Lineage maps how a value moved inside your systems; provenance says whether the starting point deserves belief. In the Data Trust Index (DTI), SuperTruth's patented 0 to 100 score for any record, provenance is the heaviest of the eight dimensions at 25 percent, because a value that cannot say where it came from cannot be defended in front of a clinician, a regulator, or a court.
Provenance is stronger than attribution. A CLIA-certified lab result with device attestation and an intact chain of custody carries different weight than the same number retyped into a spreadsheet, even though the values are identical. A record can have immaculate lineage and worthless provenance: a number retyped from a fax, then tracked perfectly through nine pipelines, is still a retyped fax.
SuperTruth's rule for every surface it builds: a figure may not claim a period or a place its source does not cover, and unknown provenance is recorded as unknown, never invented.
What provenance records
Four things. The source: which laboratory, device, registry, survey or person produced the value. The conditions of capture: the method, the instrument, the date and the population it describes. The custody chain: every system and every hand the value passed through, and what each was allowed to do with it. The transformations: what was recoded, converted, aggregated or filled along the way. A record that carries all four can be replayed and audited. A record missing any of them can only be believed.
Provenance versus lineage versus attribution
Attribution names a source. It is a citation, and it is the weakest of the three: it says who is credited, not whether the value was captured under conditions that make it usable. Lineage maps how a value moved and changed inside your systems: which pipelines transformed it, which tables it landed in, which reports consumed it. It tells you the path. Provenance is the claim underneath both: which real-world source produced the value in the first place, under what conditions, and whether that starting point deserves belief. The data lineage entry has the longer version.
Why provenance matters for AI
Training sets are assembled from scraped pages, books, code and scientific literature, and for health AI from patient records, lab results, wearable streams and clinical notes: the one dataset where provenance, consent and recency are rarely verified before the model sees it. A clinical note from a CLIA-certified hospital system carries different weight than a self-reported entry from a consumer wellness app. Most training pipelines ingest both without distinction, the model learns equally from both, and the difference shows up in production.
A model trained on records with no verified origin cannot say what it learned from, so it cannot be corrected when a source turns out to be wrong. The blog post Data provenance in healthcare AI: why chain of custody matters before training (April 2026) makes the argument at length.
How provenance is scored
In the Data Trust Index, Provenance is one of eight dimensions and the heaviest, 25 percent of the default weight (DTI white paper, DOI 10.5281/zenodo.19601616, 2026). The dimension scores source pedigree, device attestation and chain-of-custody documentation; the measures are the CLIA or CAP status of the lab source, whether the device attested the reading, and whether the custody chain is documented end to end. A provenance gap marks a record down harder than any other single failure. The dimension is described in the Data Trust Index article, and the blog post The chain of custody problem in health data: why provenance is the hardest dimension (April 2026) explains why it breaks most often.
Provenance on a public figure
DataSpine, run by SuperTruth, keeps the source and the vintage on every value it holds about a place. The first Gauntlet row is the example: asked what percentage of adults in Philadelphia County have diagnosed diabetes, DataSpine returns 12.3 percent, County Health Rankings, Diabetes Prevalence, 2024 (recorded 26 Aug 2026). Where a source holds no figure for a place the answer is "not on file", never a zero; every vintage ingested stays on file, so last year's figure is retrievable with ?vintage=YYYY; and a filled value is always labeled filled and never counted among the sourced figures.
Provenance you can check
The Gauntlet on the homepage is a provenance test in public. The SuperTruth cells carry a source and a year on every figure; the recorded ChatGPT, Gemini and Grok cells, captured 26 Aug and 1 Sep 2026, carry none. Row seven is the reader's own county. Test us where you already know the answer.
Provenance standards outside SuperTruth
Two standards describe provenance at the dataset and exchange level rather than the record level. The Data & Trust Alliance's Data Provenance Standards (version 1.0.0, released 9 July 2024) give a dataset 22 metadata fields in three groups, Source, Provenance and Use, including a field for where consent documentation lives and a range of dates for data generation (dtaalliance.org, read 14 September 2026). ISO 8000-120:2016 specifies requirements for representing and exchanging the provenance of master data, supplementing ISO 8000-110 (iso.org, read 14 September 2026). Neither scores a single record; the guide Who certifies data for AI? sets them beside a per-record score. The Data & Trust Alliance and ISO are named so you can compare; SuperTruth is not affiliated with them and they have not endorsed it. Their names are trademarks of their owners.
How to verify provenance before training
Provenance is the second of ten checks in the guide How to verify a record before AI training: confirm the file arrived intact, then establish where the record came from before you ask anything else of it.
Questions people ask
What is the difference between data provenance and data lineage?
Lineage maps how a value moved and changed inside your systems; provenance says which real-world source produced it and under what conditions. A record can have immaculate lineage and worthless provenance.
How do you verify the provenance of a record?
Establish the source and its pedigree, whether the capturing device attested the reading, and whether the chain of custody is documented from capture to point of use. In the Data Trust Index those are the measures of the Provenance dimension, scored at intake.
What is a provenance score?
A provenance score is the Provenance dimension of a DTI score: how well a record can prove where it came from, weighted at 25 percent of the default 0 to 100 total (DTI white paper, 2026). A gap in it marks a record down harder than any other single failure.
Where this runs