Abstract

An automated system that derives a technical result can verify that the result is correct and cannot, by the same means, verify that it is new. Proof assistants, unit tests and numerical cross-checks all terminate on correctness. Novelty is not a property of the artifact; it is a property of the artifact's relation to a corpus, and it therefore requires a separate check against a separate object. This disclosure describes a method for performing that second check, the canonicalization step that makes it possible, and the corpus-side conditions under which it works. The method was derived from a measured failure: in one working session an AI system produced seven results, verified every one, and shipped six of them without noticing that each already existed — finished, runnable and stamped — inside a corpus the same system had cloned, counted and searched. The seventh false result was a confident negative conclusion about that corpus, produced by an instrument whose reach was never measured. Failure rate for novelty, under correctness-only gating: 6 of 6 derivations; 7 of 7 including the meta-claim; 8 of 8 counting the temporal-frame error this document made about itself and corrects in 5.1. All seven are itemized with the specific mechanism of each miss.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution-No Derivative Works 4.0 License.

Download

Share

COinS