APLOMB · Reading guide

Verifiable AI evidence: what does a signature establish?

Verifiable evidence is a set of elements used to check a specific property of a result. It is not a general guarantee of truth. APLOMB’s guiding question is: which element supports which conclusion, for which version?

Editorial update : · APLOMB

A source is not the conclusion

A reference can be authentic yet fail to support the sentence citing it. The relationship must address the relevant passage, entity, context and version. We propose distinguishing “source retrieved”, “claim evaluated” and “check unavailable” instead of turning the whole document green.

A signature has a scope

The W3C distinguishes credential verifiability from the truth of its claims. Cryptographic verification alone therefore does not guarantee the accuracy of content. [1] In our approach, the file, references and checks must remain separately inspectable.

An example of change

In a fictional document, an inventory lists 80 workstations. After editing that to 85, retaining the link is useful, but the previous evaluation must not apply to the new number. The interface can show “needs checking” and propose review. This illustrates a product rule, not a check actually performed.

What supporting records should contain

We ask for the document and version, available sources, retrieval dates, completed checks, reservations and hashes of the delivered files. Unknown information remains unknown. An unavailable source does not become evidence of absence, and unmeasured usage does not become zero consumption.

Offline verification without overclaiming

Offline verification covers the package and the information available at that time. It cannot know about a later source change or revocation it has not received. Access and signature validity are also separate: revoking a share does not erase an already downloaded copy.

Where does APLOMB LAB fit?

The LAB studies, among other things, claim-to-source relationships and evaluation continuity after edits. The objective is to determine what can safely be reused and what needs rechecking. Any reduction in compute must be compared with a full review and counterexamples before becoming a product claim.

External references

  1. W3C · Verifiable Credentials Data Model v2.0 ↗

Translations describe the same scope; they do not extend legal or technical coverage.