verification architecture
ai generates claims faster
than anyone can check them.
manifold control designs the systems that close that gap. claims get durable identities and auditable verification histories, down to the proof terms a kernel accepted. built for organizations deploying ai where review has to survive scale.
work
running systems, rebuilt from source in ci. this section grows as new work ships.
csr registries
a corpus semantic registry: documents define, the registry identifies. the flagship demo is a mathematics preprint's claim table, sha256-pinned against the pdf, alongside two worked examples. edit a pinned source and every dependent claim is flagged in ci until re-verified.
a paper that audits itself
the claim-status table of a singular fold model for capacity-constrained dynamics (v67), maintained machine-readably. each claim carries its verification state and its dependency edges, and any byte change to the pdf flags every claim built on it. this is what "the paper says X" looks like when it is checkable.
the same machinery runs the private research corpus behind the fold paper. every registry edit lands in a hash-chained audit log. numbers as of july 2026.
the verification stack
open source and mit licensed, each piece usable on its own. together they carry an audit trail from identity through evidence.
#introspect: proof-term dag + leakage report (sorry, mvars, dependency surface)
three rules that hold throughout
the discipline underneath all of it.
events record standing decisions. the deciding happens elsewhere.
events reference identities. the registry mints them.
credit derives from events. the flow runs one way.
about
manifold control is the practice of james kovalenko (charlottesville, va). the work is verification architecture and practical ai deployment, for organizations whose review load grows faster than their review capacity.
the public work above is the method in miniature. claims become addressable and pinned, and a build step says what still holds. the same architecture carries over to spec-driven codebases and to ai agent workflows.