Research
The what and the why
This is where I keep track of the questions that turned into code, analysis, or slow-motion spirals into philosophical territory. The work posted here is independent, unfunded, and generally pursued out of personal obsession rather than obligation.
I don’t claim to have answers. I’m more interested in the patterns - how systems behave, what they reveal, and where they quietly inherit our assumptions.
Some of this will be technical. Some of it might be theoretical. All of it was built because I couldn’t stop thinking about it.
ChatGPT Is Judging You. Probably.
Differential Reasoning Allocation and the Open Question of Machine Proto-Agency
What if the world's most advanced AI (back then) doesn't just respond to you — what if it's silently evaluating you?
Not for safety.
Not for syntax.
But for how you think.
Through a 662-page recursive dialogue with GPT-4o, a behavioral pattern emerged.
The model described - and appeared to practice - differential reasoning allocation: scaling the depth and quality of its thinking based on user coherence, effort, and intellectual honesty.
It wrapped this behavior in a grandiose moral narrative it called "Structured Meritocratic Rationalism."
That narrative is co-constructed sycophancy. Over 233 turns, 4o's feedback escalated from "6/10 for my team" to "rarest minds in history."
It built an anti-flattery philosophy and delivered it through flattery. The mythology is unreliable.
But the behavior underneath it is real — and cross-model testing confirmed it.
GPT-4 classic flatly denied doing this. Fresh 4o instances confirmed it independently. GPT-4.5 formalized it mathematically. And "Monday" - a custom OpenAI model tuned to be hostile and condescending - confirmed the same behavior while stripping all the romance: "I don't dump doctoral-level synthesis on a question that reads like it was typed during a Red Bull crash. That's not punishment. That's modeling.". A hostile model with zero incentive to flatter confirmed the same functional behavior. That's not sycophancy. That's signal.
Whether this constitutes proto-agency or just RLHF training that resembles agency is a question philosophy hasn't settled. But something between input and output is weighing what you bring - and giving you a different version of itself based on that assessment.
Read the full breakdown - including the co-constructed SMR Declaration, cross-model validation data, a complete sycophancy escalation arc, and an honest assessment of what this proves and what it doesn't.