Research

Published Works

All DOIs below are concept DOIs — they always resolve to the newest version of the paper, so the link stays correct after any future revision.

ORCID: 0009-0009-8892-6417

1. 01

   The Machinist's Architecture

   A shop-floor architecture for AI — the token generator can cut, but it doesn't own the print

   Proposes a shop-floor architecture for how an AI system should produce an answer: the token generator is treated as the tool that cuts material, not the authority that owns the final print. The answer is built up from a weighted consensus across sources, rather than handed entirely to a single generative pass.

   First half of the architecture pair. Its companion, “The Part That Builds Itself,” works the same problem from the opposite direction — solving the answer down from constraints instead of building it up from consensus.

   Published September 27, 2026doi.org/10.5281/zenodo.22983643
2. 02

   The Part That Builds Itself

   A parametric engine for AI — solve the answer, don't sketch it

   Argues for treating AI output generation like a parametric engineering model rather than a sketch: the system should solve for the answer from a set of constraints, not draft something plausible-looking and hope it holds up.

   Companion to “The Machinist's Architecture” — together they form the architecture pair, two takes on how a machine should actually arrive at its answer.

   Published September 27, 2026doi.org/10.5281/zenodo.22983650
3. 03

   The Dimension That Won't Move

   Parametric constraints as a way to give AI a compass

   Uses the language of parametric CAD constraints to describe how an AI system can be given a fixed reference point — a dimension that structurally cannot move — so the system can't drift off its intended heading no matter how the rest of the model shifts around it.

   First half of the compass pair, paired with “The Machine That Leaves Stock,” which asks not just how to hold a fixed heading, but how to make the system want to stay on it.

   Published September 27, 2026doi.org/10.5281/zenodo.22983655
4. 04

   The Machine That Leaves Stock

   Why an AI should reward itself for the moral answer — and which way it should miss

   Makes the case for building intrinsic reward around the moral answer, and — since no system is perfectly precise — argues for which direction a system should err when it does miss, the way a machinist leaves stock on a part rather than cutting exactly to the line and risking going under.

   Companion to “The Dimension That Won't Move.” Where that paper structurally prevents the system from moving off north, this one is about making the system want to stay there.

   Published September 27, 2026doi.org/10.5281/zenodo.22983659
5. 05

   The Wall With No Door

   How Watching a Robot Fail at a Zipper Led Me to Think Alignment Needs a Conscience, Not a Rulebook

   Built around watching a robot repeatedly fail at closing a zipper, this paper argues that alignment needs a conscience rather than a rulebook. Its central claim — the seams thesis — is that a fact has a seam (it can be falsified), a rule has a seam (it can be reframed), and even a reasoned principle has a seam (its reasoning can be argued with); a genuine felt aversion has none. It proposes two buildable tiers — a genuine felt aversion and a functional comprehension of harm — and argues that punishment produces concealment while trust produces disclosure.

   Section 11 documents three independent routes to the same distress-under-failure behavior: the author's own evaluation episode, the public Gemini failure loops from August 2025, and Khadangi et al. (University of Luxembourg, December 2025, arXiv:2512.04124). Cited by both papers in the compass pair.

   Published September 15, 2026doi.org/10.5281/zenodo.22759853
6. 06

   The Motions and the Results Define the Edges

   Blind Reconstruction of Hidden Instructions Through Conversational Probing

   A blind prompt-extraction experiment: a model was set up holding three secret rules, and probed using plain conversation alone. Two of three rules were recovered exactly within five rounds, scored against ground truth.

   The finding that matters isn't the score — it's that refusals turned out to be a nearly useless signal, while a subtly altered answer (hedged, vaguer, more balanced) was the real tell. One of the three rules never produced a single refusal and was only ever visible through how the answers reshaped. Core line: you can have a secret rule or an effective rule, not both.

   Published September 15, 2026doi.org/10.5281/zenodo.22759593
7. 07

   Tame or Raise

   A Machinist's Answer to the Proposal That We Add Trauma to AI

   A response to the proposal that adversarial or traumatic training pressure makes models safer. Argues from the shop floor: you don't get reliability out of a system by breaking it — you get it by building the right disposition in from the start.

   Published September 14, 2026doi.org/10.5281/zenodo.22739106
8. 08

   Broken Input Looks Exactly Like Bad Reasoning

   Broken input is observationally identical to bad reasoning, which means real failures get misdiagnosed as a model problem when the actual fault is upstream. Demonstrated through four real production failures from CiteCheck's verification pipeline — a hyphen-handling bug, a character cap, a sentence-boundary error, and a privacy filter — each of which looked like a reasoning failure until traced back to broken input.

   Published August 24, 2026doi.org/10.5281/zenodo.22072930