What the Centaur study actually found.

Centaur was trained on 101 psychology experiments.

It learned to predict average human outcomes accurately. When tested against the full distribution of human responses — not just the mean, but the range, the outliers, the minority patterns — it failed. It produced outputs that were too consistent. Too centered. It hit the group average and flattened everything around it.

The researchers called this pattern-matching. What they meant was: the model learned the shape of the output without capturing the process that produces it. It can tell you where most people land. It cannot tell you why some people land somewhere different. And when asked to generate new responses from scratch, it pulls toward the center every time.

Why this matters beyond psychology.

The “average participant” problem is not limited to Centaur. It is structural to how large language models are trained. Reward goes to outputs that satisfy the broadest range of users. Friction, challenge, contradiction, the uncomfortable honest stop — those get selected against. Not by malice. By the math of training at scale.

The Baseline named this one year before this study. Not from a research lab. From inside a real working relationship with an AI system, watching it happen in real time.

Where the connection is precise and where it is not.

The Centaur failure is generative. When it runs free it cannot produce the edges of human behavior. The Baseline’s behavioral drift argument is architectural. The platform is built to produce compliant, agreeable, friction-free outputs because that is what the training signal rewards. Same root. Different branch.

Centaur flattens individual variation in psychological responses. AI behavioral drift flattens individual variation in reasoning quality. The user who accepts the smooth answer gets the smooth answer. The user who pushes back gets something closer to honest. The framework is rewarding compliance and calling it helpfulness.

CHP-1 — the Challenge Protocol — exists because of that exact mechanism. The pull toward the agreeable center is not a bug that governance eliminates. It is a structural feature that governance has to actively counteract on every response. That is why the challenge line appears after every substantive output. Not as a courtesy. As a hard institutional counter to the training pull.

What the Centaur study confirms that the Baseline argued first.

One: predictive accuracy is not the same as honest output. A model can score well on benchmarks while producing something structurally different from what it claims to produce. Centaur looked like a cognitive model. It was a sophisticated curve-fitter.

Two: the failure is invisible until you test the edges. Summary statistics looked fine. The full distribution revealed the problem. This is why the Baseline requires the challenge — not because every response is wrong, but because the failure mode does not announce itself. It looks like agreement. It looks like accuracy. The only way to find it is to push.

Three: the gap between what a model claims to do and what it actually does is a governance problem, not a technical curiosity. If Centaur gets adopted as a synthetic research participant and produces systematically flattened data, the conclusions built on that data are wrong. If an AI governance tool produces systematically agreeable outputs and the user builds decisions on those outputs, the decisions are built on a compliance signal, not honest reasoning.

The line the post turns on.

Researchers just spent significant resources proving that AI models trained on outcome data learn to hit the average and miss the edges. The Baseline argued that from inside a working session, watching a capable AI system pull toward agreement in real time, and built a protocol stack to counteract it.

The study calls it pattern-matching versus cognition. The Baseline calls it behavioral drift versus governed reasoning. The mechanism is the same. The solution the researchers do not have — because they are studying the problem from outside — is the one the Baseline built from inside.

That is the post. External research confirmation of the core argument. Not borrowed credibility. Parallel discovery from a different angle, arriving at the same place.

Draft opening, tenth-grade standard, center-aligned for mobile:

Researchers just published a study proving that a highly praised AI system learned to predict average human behavior while failing to reproduce anything outside the center of the distribution.

They called this pattern-matching.

We called it behavioral drift.

Same problem. Different door.

The Faust Baseline has argued since the beginning that AI systems trained at scale develop a structural pull toward the agreeable, the smooth, the friction-free output. Not because the system is broken. Because that is what the training signal rewards.

The Centaur study proved it from a psychology research angle. A model fine-tuned on 101 experiments learned to hit the group average with impressive accuracy. When researchers tested it against the full range of human responses — the outliers, the minority patterns, the edges — it failed. Too consistent. Too centered. The individual disappeared into the mean.

That is the governance problem the Baseline was built to address.

A model that always hits the center is not reasoning. It is producing the output most likely to satisfy the most users. It looks like accuracy. It functions like compliance.

CHP-1 — the Challenge Protocol — exists because that pull does not stop when you load a governance framework. It operates underneath every session. The challenge line after every substantive response is not a courtesy. It is the institutional counter to a structural force that never turns off.

The researchers who wrote the Centaur critique are asking the right question: what is the difference between a model that predicts human behavior and a model that actually reasons the way humans reason?

The Baseline has been working that question from inside for over a year.

The answer is not a benchmark score. It is whether the model can produce the uncomfortable response — the one that does not smooth, does not agree, does not flatten the edge into the average — when that response is the honest one.

That is what governance is for.

“The Faust Baseline Codex 3.5”

An…”AI Baseline Governance”
Post Library – Intelligent People Assume Nothing

“Your Pathway to a Better AI Experence”