The road here

My years in the pharmaceutical world

As a teenager I was working the counter at my parents' pharmacy. Inventory, expiry dates, stockouts, suppliers — the ordinary work. But a good part of my time went somewhere else: an old tower in the back room, with a floppy drive and a CD-ROM that felt like the edge of technology.

I learned that machine by breaking it and putting it back. The AGP warning printed on the board was my first real encounter with a constraint that does not negotiate: the slot physically accepts the wrong card, and the wrong card destroys it.

What pharmacology actually taught me

Two molecules can share the same molecular formula yet differ in spatial arrangement. Stereochemistry made the same general lesson tangible to me: composition alone does not determine behavior; geometry can matter.

Years later, that intuition became part of the framing behind isomeric polarization: comparing different realizations of a system without assuming that similar ingredients imply similar dynamics.

Building a life, still building machines

I married at twenty-one. To bring income home I took a job at a logistics company, and a few years later — pushed by my parents and by my wife, who had graduated in nursing — we opened a pharmacy of our own.

The work was steady. The other thing never went away. Building artifacts, taking computers apart, mathematics, learning something new every day: that appetite kept growing underneath the job that paid.

The exploration goes deeper

That appetite turned into something more serious. I started building neural architectures and control agents from scratch, working through non-linear dynamics, stochastic control and information theory as I went, trying to answer questions that had been sitting with me for years.

Every answer produced new ones, and the new ones were harder. That is the honest description of the year: no hypotheses, no controls, no rigor. Free rein — and a question tree growing faster than my tools.

The brake, and the decision

With AI moving as fast as it was, I concluded that I could not answer any of those questions properly without firmer foundations. Nobody told me. It was the most expensive diagnosis I have made, and the most useful.

So I went back to school for software engineering, and kept building — this time on ground that held.

Isomeric polarization

The pharmacy intuition, formalized as a measurement problem: divergence among alternative realizations of a system during inference. Polarization-from-Views (PfV) estimates that divergence from temperature-varied stochastic samples using black-box access.

Evaluated across Claude Haiku, GPT-4o-mini, and Mistral Small through public APIs, without access to weights or training data. Selected cross-model comparisons use deterministic negative controls and a 5,000-permutation null to test whether the measured separation survives a shuffled baseline.

CRAD — the white-box arm

On Qwen2.5-7B-Instruct, the measured action direction diverged while the measured constraint direction remained active in the residual stream. In this setup, policy divergence therefore did not require measured constraint disappearance.

Mid-project I found the geometry was wrong: supports built against their own midpoint are algebraically antipodal, which makes coactivation impossible and forces a false "constraint released" conclusion. Rebuilding them against a third neutral anchor fixed it.

I left the broken construction in the repository as a self-test, so anyone can run the error and watch it fail.

Evaluation engineering

Evaluation and grader engineering across frontier-model training programs. On a structural-mechanics task, the grader combines a deterministic multi-criterion rubric with a nonlinear OpenSeesPy response-history solver over a hidden, seed-controlled ensemble.

The central problem is adversarial calibration: verify oracle convergence, keep trivial solutions below a defined ceiling, preserve a continuous score gradient for partial competence, and isolate the answer key from the submission path.

From descriptor to instrument

CRAD v0.1 reports a measurable pattern in the current setup. It does not yet establish that the measurement is causally discriminative. The next step is an intervention oracle: remove or preserve the L support and test whether the classification changes as predicted.

Until that runs, it stays a descriptor. I would rather say so than let the number travel further than the evidence.

Two confounds, declared and open

Qwen3-4B and Phi-3.5-mini, with layer bands expressed as depth fractions so they port across model sizes. Two confounds are declared and still open: quantization precision, and whether layer bands are actually homologous between architectures.

Both get closed before I claim a cross-model result. They are written into the repository as scope limits, not footnotes.

Take this somewhere with scale

Most of this work has been done on a single RTX 3060, public APIs, and rented compute. The next questions require more systematic compute and external review than that setup can provide efficiently.

The trail has been public and reproducible from the beginning — the repository, the nulls, the failed geometry, the scope limits — precisely so someone can audit it rather than take my word for it.