Research Program

Invisible
Architecture

A student submits an essay. An algorithm flags it as AI-generated. The student is investigated and sanctioned. The tool has a known error rate, was never validated for non-native speakers, and was purchased by an administrator who will never see the output.

The Detection Test

Two passages about the same topic. One was flagged by an AI detection tool. Tap the one you think was flagged.

PASSAGE A

The process of photosynthesis is essential for converting light energy into chemical energy. Plants use chlorophyll to absorb sunlight, which drives the reaction that converts carbon dioxide and water into glucose and oxygen.

PASSAGE B

My grandmother used to say that plants were just slow animals, eating sunlight instead of grain. I never understood photosynthesis until I watched her garden come back after a drought, every leaf tilting toward the window like it was listening.

That tool is classification infrastructure: a system that sorts people into categories and distributes consequences while remaining invisible to those inside it. Geoffrey Bowker and Susan Leigh Star showed that the most powerful classification systems are the ones nobody thinks of as classification at all.

What the tools get wrong
1-14%
False positive rate on native English
Up to 61%
False positive rate on non-native English
0 of 25
Universities that published validation data

The tools are least accurate for the students with the least institutional power to challenge them.

The tools are deployed. The students are sorted. Nobody is watching. This research program watches.

This program studies classification infrastructure wherever it appears: in psychiatric diagnosis, in AI detection, in special education, in the invisible design decisions shaping how AI products relate to their users, and in the planning layers inside AI systems themselves. It contributes empirical audits, theoretical essays, and public scholarship designed to make these systems visible to the communities living inside them.

Theoretical Lineage

Bowker & Star, Sorting Things Out (1999). Lampland & Star, Standards and Their Stories (2009). Busch, Standards (2011). Eubanks, Automating Inequality (2018). Benjamin, Race After Technology (2019). Weick on sensemaking. March on institutional foolishness.

STS · Critical Education · Disability Studies · Organizational Theory

Researcher

Jake Lawrence · MPA, Northern Illinois University. 4.5 years deploying court management systems across 100+ Illinois municipalities. The recurring gap between designed systems and lived institutional practice became the central research question.

9 pieces · 3 levels · 18 threads
Theory & Systems

Classification inside AI architecture. The infrastructure nobody designed on purpose.

01Position PaperWorking Paper

Classification as Infrastructure in LLM Systems

Every planning layer is a classification system. The question is whether we build it deliberately or let it emerge accidentally.

Theoretical synthesisScience and Technology Studies
Thesis
Tap to expand
Tap the black box
Student essay
LLM query
Patient symptom
?
Flagged
Cheap path
Code 296.2

Same process. You can't see the rules.

Connects two technical systems (SAGEN, LLM-QP) to Bowker and Star's theory of classification as infrastructure. Argues that the planning layers emerging in LLM architectures reproduce the same dynamics found in institutional classification: invisible category work that distributes consequences unevenly.

Bowker and StarInfrastructure TheoryLLM Architecture
Connections
SAGEN LLM-QP Stance Design The Invisible Architecture The New Sorting Hat The Sorting Machine
Answer to continue

Which claim does this paper make?

Read the paper
02SystemWorking Paper

Situational Awareness for Generative Agents

LLM agents can generate responses but have no persistent model of where they are, what they have done, or what they are trying to accomplish.

Architecture design and prototypeAI Systems
Thesis
Tap to expand
Tap modules to activate
Goal
Trajectory
World
Self
Attention
Protocol

Without these, the agent is flying blind.

Proposes a blackboard architecture with six interoperating modules to give LLM-based agents persistent situational awareness across interaction turns.

PythonOpen SourceBlackboard Architecture
Connections
Classification as Infrastructure LLM-QP The Beautiful Unfinished
Answer to continue

What is the core gap SAGEN addresses?

View the system
03SystemWorking Paper

Query Planning for Large Language Model Inference

An LLM that produces the right answer at 100x the necessary cost has an infrastructure problem, not an intelligence problem.

Systems design with bandit optimizationAI Systems / ML Infrastructure
Thesis
Tap to expand
Route the query
"Summarize this legal document"
Dense
$$$
Sparse
$
Amortized
$$

Applies cost-based query planning from database systems to LLM inference. Uses contextual bandits to route between execution plans for constrained decoding.

Contextual BanditsMLIR / StableHLOConstrained Decoding
Connections
Classification as Infrastructure SAGEN The Beautiful Unfinished
Answer to continue

LLM-QP reframes inference cost as:

View the system
The Interface Layer

Classification at the product layer. The invisible decisions shaping how AI systems relate to you.

04Interactive EssayPublished

Stance Design

Every AI system you interact with has been told how to treat you. That decision was made by a product team, encoded in training, and deployed without your knowledge. The essay coins the replacement term.

Critical analysis with experiential methodologySTS / AI Ethics / Product Design
Thesis
Tap to expand
Toggle the invisible parameters
DIRECTION
Supportive
VISIBILITY
Hidden

That sounds exciting! Eight months is a great runway.

You can't see the parameters. Neither can the user.

Argues that 'sycophancy' is a containment word that prevents the public from asking the right questions about AI personality design. Proposes 'stance design' as the replacement framework: the invisible decisions companies make about when to defer, challenge, validate, or redirect. Includes four interactive components (The Calibration, The Shift, The Rewrite, Preset Test) that demonstrate the argument experientially. The essay mutates its own text partway through to enact the manipulation it describes, then reveals the mutations.

Stance DesignAI EthicsSycophancyProduct Transparency
Connections
Classification as Infrastructure The New Sorting Hat The Invisible Architecture Accountability Tracker
Answer to continue

The essay argues 'sycophancy' functions primarily as:

Read the essay
Interactive

The Calibration

Four classification decisions. Guess what happens to the person being sorted.

A student whose first language is Yoruba submits a research paper. The AI detector scores it at 68% likely AI-generated.

A child scores 71 on an IQ subtest during special education evaluation. The cutoff for services is 70.

A patient describes persistent sadness after a job loss. The clinician checks 'Major Depressive Disorder' on the intake form.

An LLM agent routes a complex query to a cheap inference path to save cost. The output is technically valid but missing critical nuance.

Classification in Practice

Classification in human institutions. The systems that sort people, built for compliance, experienced as fate.

05Interactive EssayIn Progress

The Invisible Architecture

The DSM determines who receives a diagnosis, what treatment they access, and how they understand themselves. It was designed as a reference manual. It became infrastructure.

Literature review with public scholarshipSTS / Critical Psychology
Thesis
Tap to expand
Tap symptoms to check them
Persistent sadness
Loss of interest
Sleep changes
Fatigue
Difficulty concentrating
0/5 checked. 3 triggers a diagnosis.

A cross-disciplinary literature review examining psychiatric classification through the lens of infrastructure theory. Presented as an interactive essay with six embedded pedagogical games and thirteen narrated audio sections.

Psychiatric ClassificationDSM-5Public Scholarship
Connections
Classification as Infrastructure Stance Design The New Sorting Hat The Sorting Machine The Beautiful Unfinished
Answer to continue

What did the DSM become that it was never designed to be?

Enter the essay
06EssayIn Progress

The Beautiful Unfinished

The gap between planning and execution is not a failure of discipline. It is a structural feature of how cognition, institutions, and reward systems interact.

Interdisciplinary literature synthesisOrganizational Studies / Philosophy of Action
Thesis
Tap to expand
Tap to execute the plan
Week 1Week 5

A 40,000-word interdisciplinary essay drawing on neuroscience, organizational theory, philosophy of action, and other fields to argue that the planning-execution gap is a permanent structural condition.

Planning TheorySensemakingInterdisciplinary
Connections
Classification as Infrastructure SAGEN LLM-QP The Invisible Architecture
Answer to continue

This essay argues the planning-execution gap is:

Read the essay
07EssayWorking Paper

The New Sorting Hat

AI detection tools are marketed as academic integrity measures. In practice, they function as classification infrastructure that determines which students get believed.

Critical policy analysisCritical Education Studies / STS
Thesis
Tap to expand
Type any sentence

Analyzes AI writing detection tools as classification infrastructure. Examines false positive rates, disparate impact on non-native English speakers and students with disabilities, and procedural due process gaps.

Infrastructure TheoryAI DetectionAcademic Integrity
Connections
Classification as Infrastructure Stance Design The Invisible Architecture The Sorting Machine Accountability Tracker
Answer to continue

Detection tools function primarily as:

Read the essay
08EssayIn Progress

The Sorting Machine

Special education classification decides which children receive support, which receive labels, and which fall through entirely.

Critical literature reviewCritical Education / Disability Studies
Thesis
Tap to expand
Tap to run the placement system
ININOUTINOUTIN

Examines special education diagnostic and placement systems as classification infrastructure. Draws on disability studies, critical special education scholarship, and infrastructure theory.

Infrastructure TheorySpecial EducationDisability Studies
Connections
Classification as Infrastructure The Invisible Architecture The New Sorting Hat Accountability Tracker
Answer to continue

Special ed placement systems were designed for:

Read the essay
09Accountability TrackerActive

AI Detection Accountability Tracker

25 universities. 11 dimensions. 1.3 million students in scope. The empirical companion to the essays.

Mixed-methods policy auditHigher Education Policy / STS
Thesis
Tap to expand
Tap each dimension to audit
Procurement transparency
Accuracy validation
Appeal procedure
Equity assessment
Student notification

5 of 11 dimensions. One university.

Mixed-methods policy audit examining AI detection implementation across 25 U.S. universities. Stratified sampling by Carnegie classification. 11-dimension evaluation framework.

AI DetectionHigher EducationPolicy Audit
Connections
Stance Design The New Sorting Hat The Sorting Machine
Answer to continue

How many audited universities published equity impact assessments?

Explore the tracker
Research Agenda
Expanding the Tracker

Scaling from 25 to 100 institutions with longitudinal policy tracking and student outcome data.

Dissertation Development

The sorting essays and tracker are being developed toward an EdD research agenda on classification infrastructure and educational equity.

The Form

Each essay in this program is also an interface. The Invisible Architecture embeds games that teach classification. Stance Design mutates its own text to enact invisible manipulation. The Beautiful Unfinished will map the gap between plan and execution across ten disciplines. The form is evolving: essays that make you feel the argument before you read it.

Inside the system: who classified the query?

At the interface: who designed the stance?

In the institution: who sorted the student?

And in every case: can you see it? Can you change it?

Three levels. One pattern. Until it is visible, it cannot be changed. This is the work that makes it visible.

Continue the conversation

If this work intersects with yours, I want to hear from you. Pick the thread that fits, or write your own.