This is a research and development site, not a product.
What ran at this address is being replaced, and the page you are reading
is the placeholder in between.
Status · under revision
The earlier prototype has been retired. Its successor is under
construction on new foundations, and this address will point at it when
there is something worth showing.
What it is
MotoKage (元影, origin shadow) is a digital twin of Jon Mott —
an attempt to answer a narrow question honestly: can a machine speak in
someone's voice, from their actual thinking, without quietly inventing
the parts it does not know?
Most systems that imitate a person are impressive and unaccountable.
They generalize from a training set and cannot show their work. The
interesting problem is not sounding like someone. It is being
checkable.
How it works
A vault, not a model. The twin's substrate is a
repository of plain text under version control — writing, research,
notes, decisions. Every tool is a client of those files. Nothing
important lives only in a model's weights.
Evidence or nothing. Claims about what a paper says
must quote it, with a page reference, from stored full text. A
deterministic verifier — no model involved — checks each quote
character by character before it is allowed into the knowledge graph.
Whatever fails is quarantined and auditable, never silently dropped.
A glass box. The persona is compiled into a
human-readable file that is committed alongside the rest: mandates
written by hand, plus inferred heuristics that each cite the evidence
they were drawn from. It can be read, diffed, argued with, and
corrected. There is no hidden prompt.
Voice stays earned. Only Jon's own writing enters
the style corpus. Machine-drafted prose is labeled from birth and never
graduates into it, so the twin cannot end up imitating itself.
Verified quote ≠ true claim. Everything the system
asserts carries a tier: proven, proposed, or confirmed by a human.
Proposals are never presented as facts.
Why the rebuild
The first version proved the idea was worth having and taught us where
it was thin. The current work rebuilds it on a stricter foundation: one
substrate, one bill, verifiable claims, and a persona that has to show
its sources. Slower, and much harder to fool.