Definition
Provenance describes the origins and production history of information. In company research it includes the source material, collection time, research activity, and subsequent review or transformation that produced a finding. A URL is one part of that history, but does not explain which version was read or how the recorded answer was derived.
Good provenance makes errors investigable. If a source was assigned to the wrong company or a question was interpreted incorrectly, a reviewer needs to identify the affected findings. The record should distinguish collection from interpretation and review. Later corrections should preserve enough history to explain both the earlier answer and the current one, rather than making a revised value appear to have always been known.
In origination work
Connect each material claim to its retained source, quoted support where applicable, collection date, and research context. Keep the question definition and accepted value visible so a reviewer can examine the reasoning. A source change, a new fact, and a review correction are different events. Docket's separation of collection and audit supports examining the basis of an answer, while unresolved provenance should remain a limitation rather than being hidden by a polished summary or an unexplained confidence label.
What provenance records
Provenance is the recorded history of a piece of information: where it came from, who or what produced it, when, and by which process. The World Wide Web Consortium's PROV data model expresses this in terms of entities, activities and agents, and relationships among them such as derivation and attribution. The vocabulary is general enough for a research file as well as for web data.
For acquisition research, provenance answers questions a reviewer will ask: what is the source of this ownership claim, when was the page retrieved, was it a primary document or a summary, did a person or an agent extract it, and has it been changed since. A fact without provenance is a claim that nobody can check.
Why it matters more with automated extraction
Language models can extract facts quickly, and they can also produce statements that are not supported by the source. Research on text generation has documented this, and evaluation methods score whether generated statements are attributable to the sources provided. Provenance is the operational answer: each extracted value is stored with the source passage that supports it, so that a reviewer can compare the claim with the text.
Standards for content credentials, such as the specification published by the Coalition for Content Provenance and Authenticity, address a related problem for media files: how to attach verifiable history to an image or document. The same idea, a tamper-evident record of origin and edits, underlies good practice for research evidence, even where the formal specification is not used.
Minimum provenance fields
Store, for each value: the source URL or document identifier, the retrieval date and time, a stored copy of the source as retrieved, the exact passage that supports the value, the method (manual or automated, and which version), and the person or process responsible. Add the entity the value concerns and the period it describes. These fields are modest, and together they make the difference between a database and an evidence file.
Review provenance as part of quality control. Sample records, open the stored source, and check that the passage supports the value. Count and report failures. If a source disappears, the stored copy remains the evidence, which is why retaining a copy matters; web content changes and disappears at significant rates.
Related terms
- Evidence trail — A reviewable chain connecting a research conclusion to its supporting material, interpretation, and decisions.
- Retained source — A saved version of source material preserved as the basis for a dated research finding.
- Grounding — Basing generated research answers on supplied evidence while still checking that the evidence supports each answer.
Further reading
- What makes an answer traceable
- Evidence standards
- Evidence trail
- Retained source
- Grounding
- Article: CRM, research database and evidence store
- Article: the research handoff checklist
Sources and further reading
- PROV-DM: The PROV Data Model, World Wide Web Consortium (W3C). A standard vocabulary for saying which entity was derived from what, by which activity, and when.
- Coalition for Content Provenance and Authenticity, C2PA. The emerging technical standard for signed provenance attached to media.
- Measuring Attribution in Natural Language Generation Models, arXiv (2021). Defines Attributable to Identified Sources (AIS): whether a statement is actually supported by the document cited for it.
- Enabling Large Language Models to Generate Text with Citations, arXiv (2023). A benchmark for whether model-written statements are actually supported by the passages they cite.
- When Online Content Disappears, Pew Research Center (2024). Measures how much of the web cited a few years ago is already gone. The argument for retaining the page, not the link.
To apply these distinctions to your own acquisition criteria, discuss your research brief or inspect the demo platform.