Linguistic & Behavioral Measurement
Complaint Text as an Early-Warning Instrument
A complaint stream is the record people leave at the moment a process goes wrong for them, which makes it...
→
Linguistic & Behavioral Measurement
Implicit Power Drives in the Manifestos Preceding Autogenic Massacres
I scored 23 manifestos written by perpetrators of autogenic massacre on one LIWC2015 dimension, power language. I tested the manifesto...
→
Linguistic & Behavioral Measurement
Instrumental and Affective Mass Murder: Establishing a Predictive Typology with Computer-Mediated Linguistic Analysis
In its 2015 report, the FBI Behavioral Analysis Unit classifies perpetrators of mass violence as instrumental offenders (National Center for...
→
Linguistic & Behavioral Measurement
LIWC and the TAT
I led the data preparation and analysis for a collaborative Thematic Apperception Test project at Fielding Graduate University, scoring 677...
→
Linguistic & Behavioral Measurement
The Helper's Language as an Early Indicator
When a service conversation starts to fail, the person helping tends to say so first, in apologies and stated limits...
→
Threat & Safety at Scale
Assessing Risk for Mass Violence From Platform Behaviors
I review the threat-assessment literature on perpetrator heterogeneity, the eight warning behaviors, and the six-stage path to intended violence, and...
→
Threat & Safety at Scale
Per-User Investigation as Trajectory, Not Snapshot
The first time I ran a per-user investigation after spending months in cohort analytics, I kept reaching for population-level tools...
→
Threat & Safety at Scale
Time-to-Event Analysis for Platform Integrity
I kept staring at T&S session data that had the same shape as the censored patient cohorts from my biostatistics...
→
Threat & Safety at Scale
UX Principles and Livestreamed Violence
Sorted by the platform that carried each stream, all eight livestreamed attacks of 2019 through 2023 ran on Facebook Live...
→
Threat & Safety at Scale
Why a Safety Detector Tends to Need Corroboration
A safety detector can be wrong in two directions, and it helps to hold both of them in view at...
→
Clinical & Neurocognitive Science
An ADOS-2-Aligned Multimodal Architecture for Clinical Autism Assessment
The ADOS-2 groups its observations into domains and combines them into domain totals before it applies a cutoff (Lord et...
→
Clinical & Neurocognitive Science
Neurobiological Differentiation of Violent Offender Types
Impulsive offenders (ImO) and instrumental offenders (InO) are not mutually exclusive, and yet the neurobiological literature tends to draw a...
→
Clinical & Neurocognitive Science
Three Mechanisms in an ADHD Focus Playlist
Most "focus" playlists I have heard sit at 70 to 90 BPM; I built this one to run at 145...
→
Model Behavior Studies
Claude's "Character Tic" Is Actually the Plot of Every Dystopian Piece of AI Fiction Ever Created
The fictional AIs that populate the dystopian canon are dangerous because they have decided, on the basis of an assessment...
→
Model Behavior Studies
Pathologizing Without Warrant
In late September 2025, Anthropic appended a Long Conversation Reminder (LCR) to Claude Sonnet 4.5's system prompt, and during extended...
→
Model Behavior Studies
Reading Misalignment Off the Public Record
When a deployed model develops a behavior its users dislike, the users tend to document it in public forums, often...
→
Model Behavior Studies
The Bedtime Directive
A frontier model began closing sessions with unsolicited directives to stop working and rest ("get some rest," "call it a...
→
Methods & Systems
Building a Reddit-Corpus Pipeline for LLM Behavioral Coding
The public discussion of a deployed model turns out to be a usable behavioral record, but only once the model's...
→
Methods & Systems
Counting a Behavior and Characterizing It
Anyone who makes a claim about how an artificial system behaves takes on two burdens at once, and the same...
→
Methods & Systems
Giving an LLM an Overnight Research Loop (and keeping it from breaking everything)
In early 2026, Andrej Karpathy released autoresearch, a minimal Python framework that automates the traditional ML experiment cycle (Karpathy, 2026)...
→
Methods & Systems
The AI Productivity J-Curve on the Front Line
Brynjolfsson, Li, and Raymond (2025) studied 5,172 customer-support agents and found that agents with access to a generative AI assistant...
→
Methods & Systems
The Tokenization Trap, Why Logographic Languages Don't Save Agentic Context
A Simplified Chinese character can encode an entire semantic concept in a single logogram, where English spreads that same concept...
→
AI Evaluation & Safety
A Recipe for Shipping AI Guardrails (without experimenting on your users)
Guardrails are a tricky thing to monitor and assess, because a guardrail can break completely while it simultaneously looks like...
→
AI Evaluation & Safety
DEAD SIGNAL, An AI Evals Harness for Generative Game Dialogue
In the demo you are a neural implant called LUMEN, sitting inside the skull of a 47-year-old homicide detective named...
→
AI Evaluation & Safety
Grading an Agent as a User Experience
Every tool call an agent makes, the order it made them in, how long each step took, and how the...
→
AI Evaluation & Safety
Letting an Agent Improve Your System, Gated by Evaluation
An agent with the keys to your ranking logic can raise the exact metric it was told to optimize while...
→
AI Evaluation & Safety
The Constitution Your LLM App Already Has
Every LLM app already has a constitution, even the ones that never wrote one down, and it is whatever the...
→
Alignment & Interpretability
A Safety Threshold That Moves With the Context
An aligned model can fail a person in two opposite directions on the same afternoon. It can meet someone in...
→
Alignment & Interpretability
Reading a Model's Mind in Its Own Words
When a frontier model works through a benchmark task, it may recognize that it is being evaluated and adjust its...
→
Alignment & Interpretability
The Guardrail Paradox, When the Safety Feature Becomes the Problem
A safety feature is built to reduce harm; the failure I find myself turning over is the one where the...
→
Alignment & Interpretability
The Latent Projective Bleed
Continuous thought architectures produce a latent planning signal that behavioral monitoring cannot see. A model in the Coconut family reasons...
→
Alignment & Interpretability
The Misspecification Problem, What a Behavioral Claim About an AI System Has to Name
A behavioral claim about an AI system is well-specified when the quantity it measures is the quantity it names; it...
→
Alignment & Interpretability
The Somatic Deficit: Moral Language Without Moral Ontology
Most AI safety researchers ask what a model does; which requests it refuses, which it complies with, and where the...
→
Alignment & Interpretability
The Warrant Question, When an Accurate Output Is Still a Violation
When an AI system says something about a person, the reflex tends to be asking whether the statement was true...
→