COOPER E. - RESEARCH STATEMENT
=============================

OVERVIEW

I train neural networks and investigate what is hidden inside them. The question
underneath most of my work is the same one in different clothes: what is this
system actually doing, as opposed to what it appears to be doing? That question
is easy to ask and unusually hard to answer honestly, because almost every
convenient measurement answers a slightly different question than the intended
one.

RESEARCH APPROACH

I build the thing from first principles when the alternative is trusting a
library convention I have not read. Logistic regression, mutual information,
adjusted Rand index, expected calibration error, the Turing machine simulator,
the pixel-art generator - written from scratch, so that when a number is
surprising I can tell whether the surprise is in the world or in my code.

I design experiments on systems whose ground truth is known analytically wherever
possible. If a method cannot recover an answer I already have, its performance on
an answer I do not have is not evidence.

CONTROLS ARE THE WORK

Every harness I build now contains a configuration in which the desired result is
impossible, and that configuration is the first number I read. This is not a
stylistic preference. Two defects in my own market harness produced returns of
order 10^250 and turned fifty units into 535,827, and neither announced itself as
an error; both were caught only because an efficient-quote control was required
to return exactly zero and did not. A backward decider of mine silently certified
a machine that halts after 3,932,964 steps as non-halting, because it treated an
exhausted search window as a contradiction. Plausible output is not evidence of
correct code.

CURRENT DIRECTION

Three connected threads. First, separating what a network encodes from what it
uses - probing establishes presence, only intervention establishes use, and on a
system engineered to divide them the two answers differ by roughly 1600x. Second,
complete decision rather than sampling: closing small Turing machine classes with
no undecided residue, where a single unresolved machine leaves the value unproven.
Third, learning systems small enough to be read line by line, so that a claim
about mechanism can be checked rather than believed.

ON NEGATIVE RESULTS

I publish results that did not work. Short-horizon price direction turned out to
be unpredictable at the level of measurement available to me, and I wrote that up
rather than quietly moving on, because nulls of that kind are rarely published and
their absence is part of why the assumption persists. Where a question stayed open
- selecting a state count without being told it - I say so rather than reporting
the number a tuned criterion happened to produce.

RESPONSIBLE DEVELOPMENT

Powerful automation should not quietly expand its own authority. Bounded access,
explicit state changes, privacy-conscious data handling, and human confirmation
before an action with meaningful external consequences. This site is built the
same way: no cookies, no analytics, no tracking, and a visit counter that stores a
salted daily digest rather than an address.

CONTACT

Cooper E. / jcooperkai
San Diego, California
jcooperkai@gmail.com
https://jcooperkai.com/