The framework
The CiteAbility™ framework evaluates whether information is dependable, evidenced and attributable enough to support an AI-generated answer. It does not score truth; it assesses whether the evidence behind a claim deserves confidence.
Seven dimensions
No single signal proves integrity. The framework combines seven lenses so that strengths in one area cannot hide material weaknesses in another.
Is the source itself credible and appropriate for the subject?
We look at whether the domain, publication or repository has a track record on the topic, whether it is the kind of source a reasonable expert would consult, and whether it carries the independence and topical depth the claim requires.
Is the person or organisation behind the information qualified and dependable for this claim?
Authorship, institutional affiliation, declared expertise and prior reliability all matter. A source is only as strong as the entity standing behind the specific claim being made.
Can the claims be substantiated, and do cited sources actually support them?
Citations must exist, resolve correctly, and say what they are claimed to say. We check for missing references, broken links, misattribution and quotations taken out of context.
Does the source show direct knowledge, testing or original evidence rather than mere repetition?
Second-hand reporting has its place, but original research, first-hand observation, measured data and disclosed methodology carry more weight than unsourced repetition.
Is the information accurate, complete, specific, current and appropriately qualified?
We assess whether the content is internally consistent, sufficiently detailed for the claim, free of obvious factual errors, and framed with the right caveats rather than overstated certainty.
Can an AI system reliably discover, parse and attribute the information?
A source can be authoritative yet inaccessible to automated systems. We consider crawlability, structured data, licensing barriers, rate limits and whether the content can be attributed back to a stable location.
Considered together, should the information be relied on and cited?
This final reasoning layer weighs the dimensions against each other. A strong score in one area cannot paper over a material failure in another. It is where human-judgment-style reasoning is encoded into the assessment.
Acts as the final reasoning and risk layer — it can override an otherwise strong assessment when something material fails.
Decision outputs
Every assessment resolves to a single decision about whether the information should support an answer. The language is deliberate: it tells builders what to do next, not just what score was reached.
Evidence and source integrity are sufficient.
Useful evidence exists, but material caveats remain.
Integrity is insufficient for confident use without more evidence.
A material integrity failure is present.
In practice
Citation Integrity™ is embedded into products and APIs, not delivered as a one-off audit.
Sources, pages and claims are assessed as they enter a knowledge base or retrieval pipeline, so weak evidence is flagged before it can influence an answer.
During answer construction, each candidate citation is checked against the framework so only well-supported evidence is surfaced to the user.
Completed answers are sampled and re-assessed to catch drift, measure improvement and satisfy governance and compliance requirements.
Responsible claims
Citation Integrity™ is designed to reduce measurable integrity failures in AI-sourced evidence. It does not eliminate AI hallucinations or guarantee that cited content is true. It validates specific, measurable improvements in citation and evidence quality, and it flags cases where the evidence is too weak to support a confident answer.
We work with AI builders and information owners to embed the framework into products, APIs and governance workflows.