| _version_ | 1866901333254078464 |
|---|---|
| author | Williams, Andy E |
| author_facet | Williams, Andy E |
| contents | <p>We designed a bidirectional corpus audit to test whether two AI safety research programs have complementary coverage gaps. The forward direction worked: a venue-relevant corpus (57 documents by the TAIGR workshop organizer) was evaluated against ten structural requirements derived from a minimal admissibility algebra, and specific gaps were identified. The reverse direction broke. The reason it broke is the paper’s main finding: the two corpora are different ontological kinds of objects. One is a collection of research contributions—experiments, methods, benchmarks. The other specifies a constraint algebra that functions as a coordinate system within which alignment approaches can be formulated and composed. A protocol designed to compare measurements with measurements does not generalize to comparing coordinate systems with measurements, because any finite subset of a coordinate system’s possible instantiations is a projection whose coverage depends on which projection was chosen. We report this ontological asymmetry as a scope condition on mutual-audit protocols in AI governance research, present the stable forward results as preliminary evidence of structural coverage gaps, and argue that the distinction between framework specifications and object-level contributions has practical consequences for how the TAIGR community evaluates its own research portfolio.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19338723 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Coordinate Systems and Measurements: An Ontological Asymmetry in Corpus-Level AI Safety Audits Williams, Andy E <p>We designed a bidirectional corpus audit to test whether two AI safety research programs have complementary coverage gaps. The forward direction worked: a venue-relevant corpus (57 documents by the TAIGR workshop organizer) was evaluated against ten structural requirements derived from a minimal admissibility algebra, and specific gaps were identified. The reverse direction broke. The reason it broke is the paper’s main finding: the two corpora are different ontological kinds of objects. One is a collection of research contributions—experiments, methods, benchmarks. The other specifies a constraint algebra that functions as a coordinate system within which alignment approaches can be formulated and composed. A protocol designed to compare measurements with measurements does not generalize to comparing coordinate systems with measurements, because any finite subset of a coordinate system’s possible instantiations is a projection whose coverage depends on which projection was chosen. We report this ontological asymmetry as a scope condition on mutual-audit protocols in AI governance research, present the stable forward results as preliminary evidence of structural coverage gaps, and argue that the distinction between framework specifications and object-level contributions has practical consequences for how the TAIGR community evaluates its own research portfolio.</p> |
| title | Coordinate Systems and Measurements: An Ontological Asymmetry in Corpus-Level AI Safety Audits |
| url | https://doi.org/10.5281/zenodo.19338723 |