Seven directions. Status is stated plainly: done, in progress, or planned.
A verification layer that runs alongside a language model and checks its claims against an explicit ontology using Answer Set Programming (Clingo/Potassco). The design borrows from SAT solvers and conflict-driven nogood learning. Status: architecture defined, prototype next.
111 algorithms encoded against abstract and mathematical attributes, principles mined with Clingo and stress-tested by three independent scanners, and the admissible design space enumerated exactly: known algorithms occupy 95 of 556,340 admissible profiles. Status: draft paper, see Papers.
A hypothesis: for increasingly capable systems, full verification may cost an order of magnitude more than development, turning the "alignment tax" into the dominant cost. The 10x–20x range is a hypothesis, not a measurement. Status: essay in preparation.
Six studies (Sep 2026) on how sentence-embedding models encode truth and negation, logical connectives, a 37-item logic taxonomy, arithmetic, and 60 ordered scales, with held-out splits, bag-of-words baselines and shuffled-target controls. One finding: there is no shared AND/OR/NOT operator across forms. Status: being merged into one paper.
Explainable reasoning via ASP plus ontologies, the common thread of the two lines above.
Language-model "connectome" and entropy-compression experiments from March 2026 . Status: unpublished; to be re-run cleanly before any claims.
Autonomous research loops, and architectures where reasoning is inspectable rather than emergent, so the explainable part can be reused across models. Status: design notes.