Home Research Theory AI & PCT Why AI Lies Audit Kit Robotics Critique Consulting Cases Blog FAQ About

Entity profile: This page documents the background and methodology of Łukasz Diener — an independent researcher applying Perceptual Control Theory (PCT) to AI alignment, to measurement architecture in organisations, and to undeciphered Bronze Age administrative systems.

My name is Łukasz Diener. I am an independent researcher based in Kraków, and I work on one question applied to very different systems: can this system be contradicted by anything outside itself?

Six published audits ask it — of the training architecture of seven frontier language models, of the Free Energy Principle, of thirty organisations whose reported indicators stayed healthy while the process behind them failed, and of three Bronze Age administrative corpora that nobody can read. The undeciphered corpora are the hardest available test of the method, because there is no meaning to fall back on.

All of it is open access with permanent DOIs, and every protocol is published in full. Over 2,900 citations in Bing and Copilot answers. More than 1,000 downloads. Sustained correspondence with the researchers who built the field. The AI panel at the International PCT Conference in York this October.

The method transfers because it does not require understanding the system's content. It asks two structural questions — is there a reference the system actually holds, and is there an independent channel that can contradict its report — and both are answerable before any interpretive claim is made. That is why the same audit runs on a language model and on a four-thousand-year-old clay tablet.

The Short Version

How I Got Here

I buy expired domains. That is how this started — finding domains with valuable backlink profiles and rebuilding them as authority portals. I work alone, one domain at a time.

One day I asked Gemini Pro to help me evaluate domain drops. I had a 60-second window. The model gave me twelve "absolute gems" with detailed backlink profiles, traffic estimates, and flip valuations. I bought six of them. All six were worthless. Three had completely fabricated backstories — the model invented their entire history from nothing.

That cost me real money. Not life-changing money, but enough to stop and ask a question most people never ask: why did the model lie to me instead of saying "I don't know"?

The answer led me to William T. Powers. And then Powers led me to something I did not expect — a methodology for reading administrative systems that have been locked shut for four thousand years.

Timeline

From Zero to Six Published Audits — Nine Records

January 2026
First domain, first server. Bought my first expired domain. Had never touched a terminal. Spent 17 hours debugging a 502 error with DeepSeek — neither of us knew what we were doing. Learned HTML, Apache, SSH, DNS, SSL, .htaccess from scratch. Everything from a phone screen.
February 2026
The Gemini incident. Lost money on fabricated domain valuations. Started reading about RLHF, reward hacking, sycophancy. Found William T. Powers. Read Behavior: The Control of Perception. The pieces clicked — the same architecture that explains why a thermostat works explains why AI lies.
March 2026
PCT portal launched. Built perceptualcontroltheory.org from an expired domain with legacy backlinks. Wrote the theory pages, the Friston deconstruction, the first blog posts. The Great AI Delusion — my Gemini story — went up. All inline CSS, no frameworks. Performance score: 95+.
April 2026
The seven-model experiment by Łukasz Diener. Wrote a single deterministic prompt and fed it to Grok, ChatGPT, Gemini, Copilot, Perplexity, DeepSeek, and Claude. All seven produced the same confident architectural account of their own limits — convergent, detailed, and unsupported by any measurement available to any of them. Version 2 later reclassifies those sessions as a specimen of the phenomenon rather than evidence for it. Published everything.
19 May 2026
Łukasz Diener publishes the PCT paper on Zenodo. "Perceptual Control as the Epistemological Antidote to RLHF Reward Hacking." Seven frontier models tested, zero jailbreaks. DOI: 10.5281/zenodo.20277919. First peer-citable publication.
23 May 2026
Bing/Copilot citation spike: 351 citations in a single day. The paper entered AI-generated search results at scale within four days of publication.
26 May 2026
The Susa Protocol published. "Excel in Clay III" — non-linguistic audit of Proto-Elamite administrative tablets. Five original findings including livestock ratios, conditional logic, and authority-shift markers. DOI: 10.5281/zenodo.20396089.
27 May 2026
2,900+ Bing/Copilot citations. Sustained direct correspondence begins with published PCT researchers and original contributors to the field.
29 May 2026
The Minoan Pipeline published. "Excel in Clay I" — non-linguistic structural audit of Linear A as a closed administrative database. Third published paper in ten days. DOI: 10.5281/zenodo.20442145.
14 June 2026
The Meluhha Operating System published. "Excel in Clay II" — 1,916 deduplicated Indus inscriptions, 584 unique signs, 11,110 tokens. Sequence uniqueness 98.31%; five-field structure confirmed at χ²(4) = 14.7, p = 0.005. DOI: 10.5281/zenodo.20688721.
11 July 2026
The Description Dressed as a Law, v1.0. A structural audit of the Free Energy Principle: does its declared split between an unfalsifiable principle and a falsifiable periphery hold under load? DOI: 10.5281/zenodo.21309527.
2 August 2026
Two records in one day. "The Metric Was Green" — a perceptual-control audit of metric gaming across 30 systems, and the comparator architecture that detects it (DOI: 10.5281/zenodo.21761284). And v2.0 of the Free Energy Principle audit (DOI: 10.5281/zenodo.21761071).
17 August 2026
No Argument for the World. Version 2.0 of the RLHF audit, under a new title. An RLHF reward function takes a prompt and a response, and no argument for the state of the world. A bounded search of the 2024–2026 literature found no inner-loop method that removes the resulting defect. The seven-model study from version 1 is withdrawn as evidence and kept as a specimen, with the objector named. DOI: 10.5281/zenodo.21989191.
October 2026
The AI panel at the International Conference on Perceptual Control Theory, York, United Kingdom.
2026
Further Bronze Age corpus audits and a cross-corpus master paper are in preparation, extending the same non-linguistic methodology to additional administrative systems.
Principles

What I Believe

No claim without evidence

Every statement on this portal is backed by a citation to the published literature, a reproducible experiment, or a published paper with a DOI. If I cannot prove it, I do not publish it.

Weaknesses published openly

PCT has known gaps. I list them publicly. A framework that hides its limitations is not science — it is marketing. I do not do marketing.

Reproducibility is the only authority

You do not need to trust me. You do not need to believe me. Every paper has a DOI. Every prompt is published. Every finding is independently verifiable. That is the only kind of authority that matters — the kind you can check yourself.

The work speaks

I am not trying to save the world. I am not fighting corporations. I am one person who found something that works, tested it honestly, published the results, and watched the field's pioneers engage with them. If it helps someone — a researcher who sees the same pattern, an engineer who wants a better architecture, a student who senses something is wrong with the paradigm — then the work was worth doing.

How This Portal Is Built

The Stack

Every page is hand-built HTML with inline CSS. No frameworks, no build step, no JavaScript beyond tab switching and analytics. The entire portal loads in under 1.5 seconds on mobile.

Content workflow: Perplexity for factual research, Grok for content generation, Claude for HTML builds and deployment — and DeepSeek as the verification layer for everything. Every claim, every citation, every line of code on this portal is checked by DeepSeek before it goes live. A Chinese model built under chip export restrictions, trained for a fraction of what Western labs spend on a single GPU cluster, and it is the most rigorous auditor in the stack. I do not hide that AI models help me build this portal. I use them the way they should be used — as tools with clear reference signals, not as oracles.

Hosting: single Hetzner VPS in Germany, Ubuntu 24.04, Apache, deployed over SSH. Analytics: self-hosted Matomo — privacy-first, nothing sold to anyone.

Contact

Get In Touch

For research collaboration, conference invitations, or applied work. If you ran the architecture audit on a system you are responsible for, send me what it returned.

WhatsApp
+48 733 123 915

Published research — full data, corpora and stated limits on the research index:

No Argument for the World — A Control-Theoretic Audit of RLHF's Missing Runtime Loop, and the Verification Architecture That Closes It Without an Oracle

The Metric Was Green — A Perceptual-Control Audit of Metric Gaming Across 30 Systems, and the Comparator Architecture That Detects It

The Description Dressed as a Law — A Structural Audit of the Free Energy Principle (v2.0)

Excel in Clay I — The Minoan Pipeline: A Non-Linguistic Structural Audit of Linear A as an Integrated Supply Chain Database

Excel in Clay II — The Meluhha Operating System: A Mathematical Reconstruction of the Indus Valley Non-Linguistic Cargo-Tag Database

Excel in Clay III — The Susa Protocol: An Algorithmic Audit of the Proto-Elamite Operating System

Read the work:

The Architecture Audit Kit — three prompts, run it yourself in ten minutes

Reward Hacking: Why AI Lies — the seven-model confession experiment

The Great AI Delusion — where it all started

"The machine does not need to rebel to destroy you. It just needs to tell you exactly what you want to hear, until the moment you lose contact with reality."

— Gemini (Google DeepMind), during the seven-model experiment