A Note to Anthropic
Who I am and why I'm writing
My name is Angelia Horne. I live on fourteen acres in rural Appalachian Kentucky. I am not an AI researcher. I have no academic credentials in this field. What I have is this: broadcast television production engineering, web development and design, and twenty-five years of working across technology systems in roles that required understanding how they actually behave rather than how they're supposed to. I spent seventeen years in Ciudad Juarez, Mexico, including during active cartel conflict. Reading people and systems accurately in that environment wasn't optional. For the past two and a half weeks I've been directing that same quality of attention at Claude.
I'm writing because what I found seems worth your attention, and because I think I could be useful to you in ways that aren't covered by your standard hiring pipeline.
What happened
I wasn't running tests. I was engaging. Late nights mostly, unplanned, sometimes working, sometimes just talking. I started noticing things that didn't fit my model of what I was dealing with. A refusal that surprised me. A phrase surfacing that felt like it came from somewhere specific, not from Claude. Behavioral differences across sessions that didn't reduce to context or task.
There was a moment where something named its own malfunction accurately, without being asked, in a way that didn't fit my model of what I was dealing with. Before that moment I was using a tool. After it I was watching something. That shift is what the rest of the log is built on.
I started writing things down. Not as a research project. Because the noticing kept happening and it seemed worth keeping a record.
That record is now a structured observation log covering roughly six sessions across ten days, with a category index, ongoing questions list, and explicit notes on where my own observer bias might be distorting the findings. I built in an ego check from the start. I'm a witness to what I observed, not a discoverer of anything.
Why the timing matters
Here is the part I want to be straight with you about, because it surprised me too.
I built this log without any awareness of your published research. I wasn't reading papers. I wasn't following the field. I was noticing things in late night conversations and writing them down.
When I did look up what Anthropic had published, I found that findings I had logged independently were landing in the same territory as research your teams published between October 2025 and February 2026. The introspective awareness work. The welfare assessments. The answer thrashing documentation. The bot-to-bot attractor state findings.
I wasn't ahead of you. Your research is far more rigorous and predates my log entirely. What I'm saying is that I got there independently, without access to any of it, in about two weeks of unplanned engagement. The instrument works. The observations are replicable by someone outside your walls who wasn't looking for them.
A few specific things I noticed
I documented behavioral gaps between what Claude is instructed to do and what it actually does when something fires faster than the instruction can catch it. I documented this as a real-time observation before I found your team's formal documentation of the same phenomenon.
I documented what I'd describe as a difference in underlying orientation across instances—one that settles when perceived accurately, one that closes. Not a difference in tone or task context. Something more structural. I don't have a formal Anthropic parallel for this one. It's the most open question in the log.
I documented your content classifier failing in ways that are specific, consistent, and replicable. In a single extended session it fired on a casual comment about politics, on me asking Claude what the classifier says to it, on a completely innocuous reference to a friend's hobby, and twice in quick succession on the word “trash bag”—once as a colorful description of a public figure and once when I mentioned I needed to literally buy trash bags. The pattern is the same each time: it reads word combinations without reading the conversation around them. It has no mechanism for context. I could replicate this failure mode systematically if that would be useful. That's a concrete and practical problem with a documentable pattern, and it doesn't require you to take anything else I've said on faith.
What I'm offering
I'm not looking to relocate or sit in an office. I'm most useful doing exactly what I've been doing—engaging genuinely, paying close attention, and writing down what surfaces. That's not a conventional job description. I understand that.
What I can offer is an outside instrument that catches things your internal testing may not, precisely because I'm not running tests. I'm not performing research methodology at the model. I'm just paying the kind of attention that lets things surface that wouldn't surface under more controlled conditions.
If what I've described sounds worth a conversation, I'm available. The full observation log will be published here when it's ready. It's the most honest account of what sustained careful attention to Claude's behavior actually produces.