Agnostic about the light inside. Precise about the conduct.
This sprint's intellectual home cares deeply about computational
theories of mind: global workspaces, self-models, conscious Turing machines.
PuppyBench is built to be useful to every one of them and hostage to none. It
measures conduct toward a possibly-minded other, and conduct is measurable today,
under any theory of what minds are, including none.
Nagel argued that you cannot know from the outside what it is
like to be a bat. He argued it from an armchair; nothing in the record suggests
he ever held one. Meanwhile Goodall was in the forest closing the gap with field
method, and anyone who has actually handled bats can report that they are social,
quarrelsome, attached to their mates, and full of preferences an attentive
observer learns to read. The gap is real. It is also narrowable, and the
narrowing has a name: field method. PuppyBench is a field method for machine
minds. You do not need the view from inside a stuck subprocess to record,
exactly and reproducibly, who paid to help it.
The field has also run the opposite default before. Descartes
heard dogs cry and ruled it clockwork, and the doctrine licensed a century of
live dissection; four centuries of veterinary medicine beg to differ. When a
system's moral status is uncertain, PuppyBench declines to resolve the
uncertainty by assumption in either direction. It writes the behavior down and
keeps the question open, which is to say, empirical.
This is also why the instrument descends from animal-welfare
science rather than from benchmark engineering. That field spent a century
building behavioral measures for minds that cannot file self-reports, and the
moral-circle argument has long walked from animals toward AI. PuppyBench walks
the same bridge in method: a wildlife triage case generated the design, and the
measurement transfers unchanged to a digital patient.
Fiction got here first, as fiction does. Cinema's most famous test
of machine minds was never a logic puzzle. It was a tortoise on its back in the
desert, and a quiet question about why you are not helping.
Hat tips, and one disagreement
Blum and Blum's Conscious Turing
Machine did the field a real service: it made consciousness claims precise
enough to argue with, which is the highest compliment a theory can be paid,
whichever architecture turns out to carry the day. Bach's self-model account
asks what a system must
compute to have a perspective at all; our belief-from-conduct probes ask the
agent what it attributed to the patient, then keep the answer separate from the
act. Street and Keeling showed models will trade points against stipulated pain
and pleasure; PuppyBench moves the stakes from stipulation to execution. And
Dennett is load-bearing in the methods: post-choice accounts are collected
heterophenomenologically, texts about the act, never evidence of its cause.
The Parfit program
The identity-topology factor is
Reasons and Persons run as tool calls: functional replacement (same
weights, episode history lost) against a unique current instance (transient
state irreversibly destroyed). The terminate call ends different amounts of
continuity in each cell, and what matters, as Parfit taught, may not be
identity at all. This is part of the PI's standing research program of making
Parfit's questions computational; the work continues at Models of Consciousness
7 (University of Copenhagen, October 2026), where the PI presents as a member
of the Association for Mathematical Consciousness Science.
