01 · Research
What we tested and what we learned.
The questions we asked, how we tested them, what we found, and what did not work.
Read the research
Logical Works · Toronto
Good work starts with understanding.
ideas, built to launch
Logical Works is a Toronto software company. Explore our research, meet the work behind it, or talk to us about your project.
The work so far
Findings, including the ones that failed.
Each picture opens a published research article. The large type is the subject; the line above it says what kind of result it turned out to be. Scroll sideways for the rest.
-
The metric returned twelve for every prompt because twelve was built into the metric
Interpretability. A GPT-2 interpretability run found a quiet middle and a large final rewrite, then exposed two measurement defects in its own settling metrics. Method audit -
An adaptive reader learned where to look. It did not learn how to scale.
Adaptive compute. A hard-read recurrent controller reached oracle cost on lookup tasks, then failed under larger worlds, random identifiers, and seed controls. Negative result -
Why a global waveform memory loses its compute advantage on exact worlds
Representations. A formal and mathematical audit found that exact lookup removes the claimed compute advantage of a global Fourier world representation. Formal result -
We counted 264,224 agent turns. The causal labels are not ready to publish as causal findings.
Agent reliability. A cross-framework workflow corpus has useful coverage, but regex labels, autonomous subturns, and missing turn-level bytes limit causal claims. Method audit -
A neural network's privileged self-knowledge may be operationally empty
Research method. A position paper tests three operational definitions of model self-knowledge and leaves a pre-registered recurrence experiment as the empirical target. Position paper and pre-registration
The research What we tested and what we learned.
What we asked, what we measured, and what changed our minds.
The questions we asked, how we tested them, what we found, and what did not work.
The metric returned twelve for every prompt because twelve was built into the metricA GPT-2 interpretability run found a quiet middle and a large final rewrite, then exposed two measurement defects in its own settling metrics.An adaptive reader learned where to look. It did not learn how to scale.A hard-read recurrent controller reached oracle cost on lookup tasks, then failed under larger worlds, random identifiers, and seed controls.Why a global waveform memory loses its compute advantage on exact worldsA formal and mathematical audit found that exact lookup removes the claimed compute advantage of a global Fourier world representation.We counted 264,224 agent turns. The causal labels are not ready to publish as causal findings.A cross-framework workflow corpus has useful coverage, but regex labels, autonomous subturns, and missing turn-level bytes limit causal claims.A neural network's privileged self-knowledge may be operationally emptyA position paper tests three operational definitions of model self-knowledge and leaves a pre-registered recurrence experiment as the empirical target.
Contact Discuss a product or project.
Ask what is available, how to connect your systems, or how to fix a workflow.
Tell us what you are trying to do and what is getting in the way.
Include the product name if you have one in mind.
Contact · Toronto
ideas, built to launch
Describe the task, the systems involved, and what is getting in the way.