AI research source versions: keep a reproducible register
AI research source versioning means recording the exact material a workflow used, not merely saving a bookmark to its current URL. Keep a source register connecting document identity, version, retrieval date and the run that produced a claim. When a source changes, you should be able to distinguish a new observation from a corrected file and trace which conclusions need review.
A URL is a location, not a complete version record
A public page can change at the same address. A downloaded file can be renamed without changing its contents. Record both a stable local identifier and the original location, along with the issuer or publisher, reporting period, publication date and retrieval date. These dates answer different questions and should not be compressed into a single “updated” field.
W3C DCAT Version 3 includes concepts for describing data resources and their versions. The lightweight register below borrows the general distinction between identity, version and relationships. It is an editorial workflow, not a claim that a personal spreadsheet implements the full vocabulary or that a version label establishes the truth of a document.
A register that connects inputs to outputs
| Field | Purpose |
|---|---|
| Source ID and version ID | Distinguish the publication from a particular revision |
| Original URL and publisher | Locate and attribute the material |
| Publication and retrieval dates | Separate when it was issued from when you accessed it |
| Reporting period and scope | Prevent accidental mixing of different observations |
| Optional file hash | Detect byte-level changes in saved files |
| Run ID and derived claims | Identify which outputs depended on this version |
| Relationship and reason | State whether a later file replaces, corrects or supplements it |
A file hash can help detect a changed file; it cannot prove the file came from the claimed publisher or that its claims are accurate. A re-export can change bytes without altering the relevant passage. Keep the original source link and a meaning check alongside any technical identifier rather than treating the identifier as a substitute for evidence.
A hypothetical corrected report
Assume you download an invented report as source S12, version V1, and a workflow produces run A7 with three claims. Later, the issuer posts V2 at the same URL with a corrected cash-flow table. Download V2 separately, record that it corrects V1, and inspect which fields changed. The original bookmark alone would not reveal what A7 actually used.
Suppose only claim C2 depended on the corrected value. Mark C2 for recalculation, then check whether its interpretation changes. Claims C1 and C3 may remain supported, but verify their dependencies rather than assuming the whole memo is either invalid or unaffected. Preserve A7 as a historical output and create a new run from V2. Do not silently rewrite A7 to appear as if it always used the corrected evidence.
Track the actual source packet
An AI run may consume a transformed excerpt rather than the original PDF. Record the relationship between the original file, extracted text, selected passage and supplied packet. If a table was manually transcribed, keep who checked it and where the original table can be found. Otherwise a later reviewer may reproduce the original source but fail to reproduce an extraction error that influenced the answer.
Keep model or tool settings only to the extent needed for your workflow, and avoid copying secrets into the register. Source reproducibility is distinct from deterministic generation: the same packet may produce different wording on a later run. Your record should still make the evidence behind each claim inspectable, even when exact phrasing cannot be reproduced.
Review the register when facts change
A new reporting period usually supplements the history rather than correcting the old observation. A correction revises how an earlier period should be understood. A third-party summary is another source with its own provenance, not automatically a new version of the issuer’s document. State the relationship explicitly and revisit claims whose evidence changed. This prevents “latest” from becoming a blanket instruction to discard context.
FAQ
Must I hash every web page?
Not necessarily. Use identifiers appropriate to the task and available materials. A saved snapshot, source date and clear version relationship may be more useful than an unexplained hash alone.
Does saving all inputs guarantee the same AI answer later?
No. It preserves the evidence packet and supports review, but generation can vary. Reproducible source context and identical generated text are different goals.