1. Missing provenance
A claim can circulate without a stable record of where it came from, what scope it describes, or when it was observed.
The open web is easy to publish to and difficult to evaluate consistently. Pages change, identities drift, copied material spreads, source context disappears, and confident language can look stronger than the evidence behind it. Generative systems make that evidence problem more visible, but they did not create it.
A machine evaluating a public source may need to answer several different questions: What URL was actually observed? When was it observed? What changed? Who controls the domain? Is the current page consistent with earlier observations? Does one stored source address a specific claim? Are two exact pages materially similar under the available evidence?
Those questions require different evidence. Collapsing them into one universal “trusted” or “untrusted” label hides important uncertainty.
A claim can circulate without a stable record of where it came from, what scope it describes, or when it was observed.
A site can make claims about ownership, authority, or affiliation that are difficult to separate from independently observed evidence.
A page can change after it was first retrieved. Without historical observations, a consumer may not know that the evidence moved.
Large volumes of similar pages can create the appearance of independent support even when the available evidence shows strong content relationships.
A cited page may exist but still fail to support the particular claim attributed to it.
A system may have too little evidence to answer confidently. Treating missing evidence as positive evidence creates false certainty.
A derived score or model judgment can be mistaken for raw observation unless the evidence classes stay explicit.
| Capability | Evidence contribution | Limit |
|---|---|---|
| Evaluate | A GoGuides assessment tied to a defined normalized scope. | Not a universal certificate that the source is factually correct. |
| Web State V3 | Stored objective web-state evidence for supported dimensions. | Not a guarantee that the live site still matches the stored observation at every moment. |
| Change Intelligence | Deterministic comparison of concrete stored observations. | Does not invent publisher motive or semantic importance. |
| Verification | Evidence that an accepted method demonstrated control or authorization. | Does not make every statement on the verified site true. |
| Claim Support | Bounded analysis of whether one fixed stored source addresses a claim. | Not a universal fact-checking oracle. |
| Content Relationships | Conservative comparison of two exact URLs under stored evidence. | Does not prove common ownership, coordination, deception, or intent. |
An outside model controls its own generation, retrieval, citation, and safety behavior. Better provenance and source evidence can reduce uncertainty; they cannot force another system to produce a correct answer.
Proof of domain control answers an identity question. It does not prove every factual claim published by the domain.
Pages change for legitimate reasons. Deterministic change evidence tells a consumer what changed under the supported comparison contract, not why.
Content relationships can be evidence of similarity without supporting a conclusion about intent, ownership, or coordinated behavior.
GoGuides' role is to make evidence and GoGuides-derived interpretation inspectable. The final source-use decision remains with the consumer. That separation is what makes the evidence useful to systems that may have different ranking, citation, retrieval, safety, or eligibility policies.