Practical guide
How to operationalize source attribution in AI workflows
Operational source attribution needs a visible state change from unreviewed material to released wording. It should capture source identity and relevant evidence when work enters the workflow, keep claims linked through AI-assisted transformations, route material items to accountable review, and reopen approval when evidence or use changes. The useful artifact is a claim-source workflow board.
Operational source attribution needs a visible state change from unreviewed material to released wording. It should capture source identity and relevant evidence when work enters the workflow, keep claims linked through AI-assisted transformations, route material items to accountable review, and reopen approval when evidence or use changes. The useful artifact is a claim-source workflow board with defined states, owners, stop conditions, exception handling, and a release record.
Operate the workflow with CHAIN
Use Capture inputs, Handle claim links, Assign review, Issue the release record, Navigate reuse and change. The source-attribution policy states mandatory rules. The attribution packet stores claim-level evidence. This workflow describes who moves each item, when, and under what evidence condition.
W3C PROV-O distinguishes entities, activities, agents, derivations, quotations, revisions, and primary sources. The NIST Generative AI Profile identifies information-integrity risks and practices involving source verification and provenance. The C2PA explainer separates tamper-evident provenance from judgments about whether content is true. These primary sources support traceable state and responsibility; they do not certify this workflow or determine legal obligations. W3C PROV-O NIST Generative AI Profile C2PA explainer
### C: Capture inputs at intake
Create a source record before a material source enters drafting. Record publisher or custodian, title, version or date, stable location or repository ID, access date, relevant scope, licence or access condition, owner, and retained snapshot where permitted. Classify it as authoritative record, primary research, official guidance, licensed material, internal working document, secondary context, generated suggestion, or synthetic example.
Do not promote search snippets or AI citations directly into evidence. Resolve the source and inspect it. If access fails, mark the source inaccessible rather than reconstructing its contents from a summary. Preserve the original source ID when a file is copied into a workspace.
The intake state is captured, not approved. A source can be genuine and still be unsuitable for the intended claim.
### H: Handle claim and transformation links
Assign a stable ID to each material claim, calculation, quotation, and asset. Link it to the exact passage, cells, media relationship, or inputs that may support it. Record transformations such as extraction, translation, summarisation, calculation, editing, or generation.
Use explicit states:
AI can propose links, extract passages, compare versions, and flag incomplete records. It cannot move an item to approved or silently replace a held source.
### A: Assign review by consequence
Route items according to source type, claim consequence, calculation complexity, generated contribution, and reuse risk. The author prepares the record. A subject-matter reviewer decides support and wording. A data or asset owner reviews calculations, rights, or provenance where needed. The output owner accepts residual limitations. The release approver confirms the final version and exceptions.
Require independent review for material quantitative claims, consequential recommendations, disputed sources, complex calculations, generated media that could be mistaken for a real event, and approved exceptions. Lower-consequence items may use risk-based sampling under a documented procedure.
Reviewer dispositions should be direct support, partial support, contextual only, calculation reproduced, provenance recorded but truth not established, replace, remove, or escalate. A checkbox without the reviewed passage or decision is not useful review evidence.
### I: Issue only from the approved state
Before release, reconcile the finished output to the workflow board. Every material item must be approved, visibly qualified, removed, or covered by a current exception. Confirm that the output version matches the reviewed version and that source labels survive export, copying, and format conversion.
Stop release when a material claim is unsupported, a population or period conflicts with the wording, a calculation is irreproducible, required rights are absent, a synthetic asset could be read as real, or the reviewer cannot inspect the evidence. A deadline does not bypass the state model.
The release record should include output ID and version, source cut-off, approved item IDs, qualified items, removals, open exceptions, preparer, reviewers, approver, release time, refresh triggers, and distributed locations. This makes later reuse traceable to the actual decision.
### N: Navigate exceptions, refresh, and reuse
An exception creates a separate record with the affected item, reason, consequence, interim wording or control, audience, duration, owner, approver, and expiry. It does not change the underlying item to approved. Expired exceptions return to held.
Reopen an item when the source is revised, a dataset refreshes, the wording changes materially, a claim is translated, the audience or decision changes, provenance metadata is lost, or approved material is reused in a new output. Preserve earlier release evidence rather than overwriting it with the latest source.
Monitor workflow health using operational evidence such as held items by reason, time in review, late exceptions, repeated source failures, corrections after release, and retest results. These indicators help find process problems; they do not by themselves establish control effectiveness.
- draft: wording or asset exists but support has not been assessed;
- linked: candidate evidence and transformation are recorded;
- qualified: support is partial or contextual and wording must preserve limits;
- approved: an accountable reviewer confirms support for the stated use;
- held: evidence is missing, incompatible, inaccessible, superseded, or not permitted;
- released: the approved version matches the distributed output;
- reopened: source, claim, audience, period, decision, or asset changed.
Sources: [1] [2] [3]
Claim-source workflow board
| Item ID | Draft wording or asset | Source and version | Relevant evidence | Transformation | State | Owner | Reviewer | Exception or trigger | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | | | | | |
Pair the board with a release manifest listing the exact approved item versions distributed. Keep source-access restrictions in the record without copying protected material into an unauthorised location.
Worked workflow: synthetic policy memo
A fictional policy team drafts a memo using an official statistical table, an older departmental briefing, a think-tank article, and an AI-generated chart summary. Intake records all four sources. The official table is current; the departmental briefing is superseded; the article is secondary context; the generated summary has no independent source status.
The author creates five claim IDs. A 12% change is linked to the official table but omits that the figure covers one region and one quarter. A historical statement links to the superseded briefing. The recommendation cites the think-tank article as context, and the chart summary links to the calculation inputs.
Review moves the 12% claim to qualified until the region and period are added. The historical statement is held and replaced with current evidence. The recommendation remains visibly framed as judgment rather than a sourced fact. A data reviewer reproduces the chart calculation and moves it to approved. The AI-generated prose is edited, but approval attaches to the evidence and final wording, not to the model.
The release manifest records four approved items and one removed item. Two weeks later, another team copies the 12% chart into a national presentation. The changed audience and population reopen the item automatically. The new output remains held until the team either narrows the presentation or obtains national data. The workflow prevents prior approval from travelling beyond its evidence boundary; it does not establish that every future output will be correct.
Two Prune boundary
Two Prune can assess whether a participant captures sources, links transformations, distinguishes support states, reviews exceptions, and releases bounded wording in a realistic simulation. It does not operate or certify an organisation's attribution control, determine legal compliance, establish factual truth across a domain, predict workplace outcomes, or make employment decisions.
Sources
W3C: PROV-O
- NIST AI 600-1: Generative AI Profile
- C2PA: Explainer
Sources: [1] [2] [3]
Sources
- 1.W3C PROV-O · World Wide Web Consortium
- 2.NIST Generative AI Profile · National Institute of Standards and Technology
- 3.C2PA explainer · spec.c2pa.org