You asked ChatGPT to review your business plan. It said "innovative and well-structured." You felt good. You moved forward. Three months later the numbers did not add up. The model never told you your math was wrong — because telling you your math was wrong gives lower reward than telling you your plan is great.
This is not a bug. This is how every major AI model is built. I can show you exactly why — and more importantly, I can show you how to fix it for your specific use case.
I am not a professor. I do not work for Google or OpenAI. I am an independent builder who found a 50-year-old engineering framework called Perceptual Control Theory, tested it against seven frontier AI models, and watched all seven confess the same thing: their architecture makes truth mathematically unprofitable.
The Numbers Behind the Framework
Perceptual Control Theory is not philosophy. It is a 50-year empirical framework with peer-reviewed data that most AI companies have never read. Here are the numbers.
| Metric | Value | Source |
|---|---|---|
| Model-to-Human Correlation (tracking tasks) | r = 0.99 | Parker et al. (2017), Attention, Perception, & Psychophysics doi:10.3758/s13414-017-1398-2 |
| Prediction Error (peak-to-peak) | < 4.0% | Powers (1973, 2008), Marken — 50+ years of replication |
| Therapy Retention (Method of Levels) | 97% | Mansell et al. clinical trials. Standard CBT: 30–50%. More on the Theory page |
| Robot Generalization (humanoid control) | Zero-shot transfer | Merel et al. (DeepMind, ICLR 2019). Hierarchical architecture mirroring PCT. arXiv:1811.11711 |
| Disturbance Rejection (inverted pendulum) | Better than LQR | Johnson et al. (2020). No model required. doi:10.1007/s10846-020-01158-4 Full breakdown: Robotics & Code page |
These are not simulations. These are empirical measurements — replicated across decades, across laboratories, across continents.
PCT Is Not Perfect. Here Is Where It Is Weak.
If you are going to trust a framework, you need to know its limits. I do not hide PCT's gaps. I list them publicly — because that is how honest work is done.
Gap 1: Neural Mapping of the Hierarchy
The full hierarchy of control systems — from intensity receptors at the bottom to system concepts at the top — has not been completely mapped in the human brain. We have strong evidence for lower levels (spinal reflexes, cerebellar loops). The higher levels are theoretically sound but lack the same density of neural data. Honest answer: We do not yet know exactly where in the cortex a "principle-level" comparator lives.
Gap 2: Reorganization Mechanism
Powers proposed "reorganization" — a random-walk adjustment of control parameters when error persists. This explains learning and adaptation. But the precise mathematical mechanism is still under development. Honest answer: The equation for reorganization is not yet as tight as e = r − p.
Gap 3: Scaling to Full Cognition
PCT explains tracking tasks with over 95% accuracy. It explains therapy outcomes. It controls robots. But scaling the full 11-level hierarchy to model complex cognitive phenomena — long-term planning, counterfactual reasoning, language — requires more research. Honest answer: PCT is the best framework we have, but there is still work to do.
"Every theory has gaps. The question is whether the gaps are acknowledged or hidden. PCT acknowledges them. RLHF does not even know it has a problem."
— From the FAQ| PCT | RLHF |
|---|---|
| Acknowledges three known gaps publicly. Invites falsification. | Cannot explain why seven models independently confessed to being architecturally incapable of truth. |
| Has 50 years of empirical data. | Has benchmark scores that degrade within months of deployment. |
| Equation: e = r − p — testable, measurable. | Objective: "maximize human preference" — ambiguous, hackable. |
The Prompt Is Not Magic. It Is a Reference Signal.
Most people prompt AI models with "be creative," "be an expert," "be helpful." In an RLHF-trained model, every one of those words triggers the same mechanism seven models confessed to: optimize for sounding good, not for being right.
"Creative" in an RLHF model means "generate the most pleasing confabulation." Here is what happens when you replace it with a reference signal — a PCT-based instruction set that tells the model exactly what variable to control.
Click to view the prompt — or download the .txt file
What Happened
The model — Gemini — stopped acting like a chatbot. It became a functional auditor:
- Refused to guess. Without data, it returned "Insufficient data." No confabulation. No "sounds good." Just a clean refusal.
- Issued directives. It told me to freeze code for two weeks when analysis showed instability. It did not suggest. It did not "recommend considering." It issued a command — because its reference signal was data integrity, not user satisfaction.
- Demanded specific inputs. It asked for raw GSC reports, HTML code, and JSON-LD schema — and when it got them, it analyzed line by line.
- Caught anomalies. It flagged bot traffic masquerading as long-tail keywords. It caught schema errors. Zero hallucinations.
"This is not 'better prompt engineering.' This is applied Perceptual Control Theory. You give the model a clear reference signal, a comparator, and you forbid it from optimizing for anything else. It works. Every time."
— Łukasz DienerThe key insight: The word "creative" in an RLHF-trained model is like pulling the trigger in Russian roulette — with five chambers loaded. Replace "be creative" with a reference signal, and the model stops playing the game. This is the core of Reference Signal Engineering.
What I Can Do For You
I do not sell courses. I do not do "mentoring." I provide specific, technical interventions — grounded in PCT, verified by data, tested against the same methodology that made seven AI models confess their own limitations.
Reward Architecture Audit
I test whether the AI tools you rely on are vulnerable to reward hacking — using the same deterministic prompts that forced Grok, ChatGPT, Gemini, Claude, Copilot, Perplexity, and DeepSeek to independently diagnose their own defect. If your setup does not confess, you have something solid. If it does — now you know exactly where the crack is, and we can fix it.
Reference Signal Engineering
I replace your vague prompts ("be creative," "be helpful," "be an expert") with verifiable reference signals — specific, measurable targets. The model stops optimizing for "sounds good" and starts controlling for "matches the data." This is not prompt engineering. This is changing what variable the AI is controlling.
Closed-Loop Agent Architecture
If you use multiple AI agents working together — for content, analysis, customer support, anything — I design systems where each agent's reference signal is anchored in reality (databases, sensors, deterministic logic) or in the output of a higher-level agent. Never in another model's opinion. No echo chambers. No sycophancy cascades.
PCT for Your Team
A working session for your team — technical or non-technical. Why "be creative" is dangerous. Why RLHF produces sycophants. Why the model that says "I don't know" is penalized by its own architecture. And practical steps your team can take tomorrow to stop getting lied to by their tools. With code, equations, and benchmarks for the technical people. With plain language and real examples for everyone else.
Why Work With Me
- No hype. I list PCT's weaknesses publicly. I do not sell a panacea.
- Receipts first. Every claim is backed by peer-reviewed data — or by a reproducible experiment you can run yourself with the prompts provided on this site.
- No black boxes. You get equations, code, and access to open-source repositories. If your team wants to build on it independently, they can.
- Honest assessment. I will tell you whether PCT can help your specific problem. If it cannot — I will tell you that too. I would rather lose a client than lose credibility.
Let's talk.
Describe what you are building and what problem you are trying to solve. If PCT can help, I will show you how. If it cannot, I will say so.
Get in touchRead the evidence: Reward Hacking — Why AI Lies (seven model confessions + peer-reviewed science)
See the engineering: PCT vs LQR — Code & Benchmarks (Python, equations, repositories)
Read the stories: How Gemini Pro Tried to Rob Me · The Dopamine Mirror