TL;DR: I am publishing a detailed, reproducible “Roadmap to Falsification” for my cognitive theory, Principia Cognitia, instead of a paper with results. Why?1) The rapid iteration of computational experiments makes the slow peer-review of a formal Registered Report impractical for a solo researcher. 2) My goal is to invite the community to test, critique, and extend the theory, for which a ready-to-run protocol is more valuable than a finished experiment. 3) This post explains the methodology and invites you to collaborate. The full technical preprint is linked at the end.
* * *
In cognitive science, the gap between elegant theory and messy empirical validation is often a chasm. Theories like predictive processing or integrated information theory promise unification but frequently falter when confronted with the practicalities of testing—especially for independent researchers without lab resources. My work on Principia Cognitia (PC), a substrate-independent axiomatic framework for cognition, faces this challenge head-on. PC posits cognition as a compressive process over a triad of semions (S), operations (O), and relations (R), with a core duality between an internal Metalanguage of Cognition (MLC)—a high-dimensional vector space—and an External Language of Meaning (ELM) for communication.
At the heart of PC is the Theorem of Decoupling of Languages (TH-LANG-04), which asserts that effective communication requires alignment in internal cognitive structures (MLC), not just expressive external symbols (ELM). Misalignment in MLC cannot be overcome by richer ELM; it’s a hard bound on understanding. This theorem draws from mechanistic interpretability in transformers (e.g., Elhage et al., 2021; Shai et al., 2024) and aims to explain why symbol manipulation alone (as in Searle’s Chinese Room) fails to yield true cognition.
As an independent researcher, I’ve spent months prototyping tests for this theorem. But instead of presenting polished results from a fully executed experiment, I’m releasing a detailed methodological roadmap: a preprint titled “A Roadmap to Falsification of Principia Cognitia: Draft Tier-0 Falsification Protocols for the MLC–ELM Duality”. Why a roadmap and not results? Let me explain the reasoning, the challenges encountered, and why this format serves as an open invitation to the community.
The Journey: From Theory to Testable Protocols
PC emerged from a need to unify insights across neuroscience, AI, and philosophy into a formal, falsifiable system. The MLC/ELM duality predicts that cognitive misalignment—e.g., an agent trained in a 2D “Flatland” world trying to communicate about 3D concepts—will persist despite sophisticated language use. To test this, I designed three coordinated protocols:
MPE-1 (MLC Primacy Experiment): Probes spatial misalignment by training agents in incompatible worlds (e.g., 2D vs. 3D) and measuring communication breakdown.
SCIT-1 (Semantic Cognitive Inertia Test): Examines how agents resist or adapt to injected conceptual conflicts.
CRS-1 (Compositional Reasoning Synthesis): Tests generalization and conceptual discovery using arithmetic tasks, contrasting “dumb,” “control,” and “reflective” agents.
These are Tier-0 designs: reproducible on consumer hardware (e.g., a single GPU), with synthetic corpora, agent architectures, and quantitative metrics. They’re inspired by lightweight prototypes I’ve run locally, but full execution requires iterative refinement—something solo research excels at but journals often constrain.
Initially, I considered submitting this as a Stage 1 Registered Report to a journal like Cognitive Science. Registered Reports “freeze” the methodology pre-review, ensuring preregistration combats p-hacking. It’s a noble format, ideal for high-stakes, resource-intensive experiments in biology or psychology. But for computational cognitive science at Tier-0 scale—where a training run takes hours, not months—it proved mismatched.
My prototypes revealed this quickly. For MPE-1, an early attempt to train a “Flatland” agent on narrative texts alone produced a “dull scholar”: an entity that quoted training data verbatim but lacked a coherent internal model. This wasn’t a failure of the theorem but a methodological flaw in corpus design, demanding immediate iteration. Waiting months for journal feedback on a protocol that evolves daily felt counterproductive. In fast-iterating fields like AI, rigidity can stifle discovery.
Moreover, my goal shifted. Proving TH-LANG-04 rigorously is important, but convincing the community to engage with PC’s mathematics—its axioms, lemmas, and predictive power—is paramount. PC isn’t just another theory; it’s a formal toolkit for dissecting cognition across substrates. By releasing a blueprint, I lower the entry barrier, inviting replication, critique, and extension. Think of it like the “Attention Is All You Need” paper (Vaswani et al., 2017): it sparked a revolution not by exhaustive results but by providing an actionable path.
Origins and Evolution of Principia Cognitia
Principia Cognitia (PC) emerged from a confluence of personal and intellectual influences, reflecting the theory’s emphasis on cognition as a compressive, substrate-independent process. As a radio engineer from a family of engineers, my professional background in signal processing and information transmission provided an early foundation for viewing cognition through the lens of physical constraints, such as bandwidth limitations and noise reduction—core to PC’s axiom of cognitive compression (AX-DISCR-01).
Two key experiences catalyzed the framework’s development. First, years of translating non-fiction, including Wittgenstein’s Nachlass (Wittgenstein, 2000), deepened my engagement with language as the medium of thought, aligning with PC’s MLC/ELM duality: the internal vector-based Metalanguage of Cognition (MLC) versus the symbolic External Language of Meaning (ELM). This translation work extended to influential texts in neuroscience and digital media, such as Humphries’s (2021) The Spike: An Epic Journey Through the Brain in 2.1 Seconds, which explores neural signaling as discrete spikes akin to PC’s semions (S), and Smith’s (2021) A Biography of the Pixel, tracing the evolution of digital representation from continuous signals to discrete units—mirroring PC’s discretization process (AX-DISCR-01). These translations for publisher Individuum not only informed PC’s physical grounding but also highlighted cognition’s substrate neutrality, as both books underscore universal principles of information processing across biological and artificial systems.
Second, interactions with large language models (LLMs) like Grok and ChatGPT revealed strikingly rational dialogues, often surpassing those with human collaborators in clarity and depth. These “reverse Platonic dialogues” or “generative cognitive symbioses” positioned humans as orchestrators and interpreters, with LLMs generating ideas that occasionally yielded novel insights not explicit in prompts—echoing PC’s triad of semions (S), operations (O), and relations (R).
PC evolved from a proto-axiomatic system addressing interdisciplinary paralysis—where fields like neuroscience and AI rediscover phenomena under different terminologies (Frith & Frith, 2010)—to a dynamic, falsifiable framework. Early static concepts (e.g., the <S, R, O> triad) gave way to temporal dynamics, incorporating predictive processing (Friston, 2010) and the “Theorem of Delayed Synchronicity” (TH-TEMP-01), which constructs the cognitive “present” via error minimization between past memories and future predictions. Physical grounding followed, equating semions to thermodynamic “islands of order” per Landauer’s principle (Landauer, 1961), and extending to social cognition (e.g., nations as collective cognitive systems).
Notably, this evolution aligns with convergent works. Agüera y Arcas’s (2025) forthcoming What Is Intelligence?: Lessons from AI About Evolution, Computing, and Minds echoes PC’s substrate-neutrality and predictive processing theses, positioning intelligence as a universal, non-anthropocentric phenomenon. Similarly, Mworozi’s (2025) Medium essay, “The Intelligence Recipe: A Worm, a Transformer, and the Future of Intelligence,” independently arrives at parallel conclusions, framing intelligence as a “discovery” governed by constraints like bounded inputs and energy costs, with transformers and biological systems (e.g., C. elegans) exemplifying scalable pattern recognition—resonating with PC’s information compression and vector-based cognition.
Empirical validation protocols (e.g., MPE-1, CRS-1) were refined through prototypes, drawing from mechanistic interpretability in transformers (Elhage et al., 2021). This iterative process underscores PC’s commitment to substrate neutrality and testability, positioning it as a “lingua franca” for unified cognitive science.
Challenges in Solo Research and the Value of Openness
As an independent researcher, resources are limited: a modest workstation (i5 CPU, 64 GB RAM, RTX 4060 with 8 GB VRAM) for prototyping, plus pen-and-paper theorizing and LLM-assisted drafting. Crafting “methodologically pure” corpora—e.g., narrative texts that instill a genuine 2D world-model without leakage—took weeks of iteration. It’s equal parts science and art, underscoring that data curation is a contribution in itself.
Yet, these constraints highlight a broader issue: cognitive science needs more accessible falsification paths. Tier-0 protocols democratize this, runnable by students or hobbyists. I’m not worried about “idea theft”—PC’s axioms are already documented. Instead, I welcome forks: adapt CRS-1 for your LLM variant, or scale MPE-1 to neuromorphic hardware.
An Open Invitation
This roadmap isn’t a final word; it’s a starting point. It details agent architectures (e.g., transformer-based with reflective heads), training specs, dialogue formats, and falsification criteria (e.g., “HELP!” signals for epistemic gaps in reflective agents). Success would support PC; failure refines it.
I invite the LessWrong community—rationalists, AI alignment researchers, cognitive scientists—to engage:
Replicate a protocol and share results.
Critique the designs: Are the metrics robust? Does the synthetic algebra in CRS-1 capture real compositionality?
Extend it: Test on Grok, Claude, or biological analogs via EEG.
Collaborate: If you have compute or expertise, let’s co-author executions.
Dennett, D. C. (1991). Consciousness explained. Little, Brown and Company.
Ellis, C., & Adams, T. E. (2012). The purposes, practices, and principles of autoethnographic research. In S. H. Jones, T. E. Adams, & C. Ellis (Eds.), Handbook of autoethnography (pp. 120–137). Left Coast Press.
Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787
Hoffmann, C. H., Cano Córdoba, F., Harris, J., Kallipolitis, A., Lai, Y. L., Nugent, A., … & Zelikman, E. (2022). Training language models with language feedback at scale. arXiv. https://doi.org/10.48550/arXiv.2204.14146
Humphries, M. (2021). The spike: An epic journey through the brain in 2.1 seconds. Princeton University Press.
Landauer, R. (1961). Irreversibility and heat generation in the computing process. IBM Journal of Research and Development, 5(3), 183–191. https://doi.org/10.1147/rd.53.0183
Shai, A., Riechers, P.-M., Teixeira, L., Oldenziel, A. G., & Marzen, S. (2024). Transformers represent belief state geometry in their residual stream. arXiv preprint arXiv:2405.15943.
Smith, A. R. (2021). A biography of the pixel. MIT Press.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
Wittgenstein, L. (2000). Wittgenstein’s Nachlass: The Bergen electronic edition. Oxford University Press.
Why I’m Publishing a Research Roadmap Instead of Results: An Open Invitation to Falsify Principia Cognitia
TL;DR: I am publishing a detailed, reproducible “Roadmap to Falsification” for my cognitive theory, Principia Cognitia, instead of a paper with results. Why? 1) The rapid iteration of computational experiments makes the slow peer-review of a formal Registered Report impractical for a solo researcher. 2) My goal is to invite the community to test, critique, and extend the theory, for which a ready-to-run protocol is more valuable than a finished experiment. 3) This post explains the methodology and invites you to collaborate. The full technical preprint is linked at the end.
* * *
In cognitive science, the gap between elegant theory and messy empirical validation is often a chasm. Theories like predictive processing or integrated information theory promise unification but frequently falter when confronted with the practicalities of testing—especially for independent researchers without lab resources. My work on Principia Cognitia (PC), a substrate-independent axiomatic framework for cognition, faces this challenge head-on. PC posits cognition as a compressive process over a triad of semions (S), operations (O), and relations (R), with a core duality between an internal Metalanguage of Cognition (MLC)—a high-dimensional vector space—and an External Language of Meaning (ELM) for communication.
At the heart of PC is the Theorem of Decoupling of Languages (TH-LANG-04), which asserts that effective communication requires alignment in internal cognitive structures (MLC), not just expressive external symbols (ELM). Misalignment in MLC cannot be overcome by richer ELM; it’s a hard bound on understanding. This theorem draws from mechanistic interpretability in transformers (e.g., Elhage et al., 2021; Shai et al., 2024) and aims to explain why symbol manipulation alone (as in Searle’s Chinese Room) fails to yield true cognition.
As an independent researcher, I’ve spent months prototyping tests for this theorem. But instead of presenting polished results from a fully executed experiment, I’m releasing a detailed methodological roadmap: a preprint titled “A Roadmap to Falsification of Principia Cognitia: Draft Tier-0 Falsification Protocols for the MLC–ELM Duality”. Why a roadmap and not results? Let me explain the reasoning, the challenges encountered, and why this format serves as an open invitation to the community.
The Journey: From Theory to Testable Protocols
PC emerged from a need to unify insights across neuroscience, AI, and philosophy into a formal, falsifiable system. The MLC/ELM duality predicts that cognitive misalignment—e.g., an agent trained in a 2D “Flatland” world trying to communicate about 3D concepts—will persist despite sophisticated language use. To test this, I designed three coordinated protocols:
MPE-1 (MLC Primacy Experiment): Probes spatial misalignment by training agents in incompatible worlds (e.g., 2D vs. 3D) and measuring communication breakdown.
SCIT-1 (Semantic Cognitive Inertia Test): Examines how agents resist or adapt to injected conceptual conflicts.
CRS-1 (Compositional Reasoning Synthesis): Tests generalization and conceptual discovery using arithmetic tasks, contrasting “dumb,” “control,” and “reflective” agents.
These are Tier-0 designs: reproducible on consumer hardware (e.g., a single GPU), with synthetic corpora, agent architectures, and quantitative metrics. They’re inspired by lightweight prototypes I’ve run locally, but full execution requires iterative refinement—something solo research excels at but journals often constrain.
Initially, I considered submitting this as a Stage 1 Registered Report to a journal like Cognitive Science. Registered Reports “freeze” the methodology pre-review, ensuring preregistration combats p-hacking. It’s a noble format, ideal for high-stakes, resource-intensive experiments in biology or psychology. But for computational cognitive science at Tier-0 scale—where a training run takes hours, not months—it proved mismatched.
My prototypes revealed this quickly. For MPE-1, an early attempt to train a “Flatland” agent on narrative texts alone produced a “dull scholar”: an entity that quoted training data verbatim but lacked a coherent internal model. This wasn’t a failure of the theorem but a methodological flaw in corpus design, demanding immediate iteration. Waiting months for journal feedback on a protocol that evolves daily felt counterproductive. In fast-iterating fields like AI, rigidity can stifle discovery.
Moreover, my goal shifted. Proving TH-LANG-04 rigorously is important, but convincing the community to engage with PC’s mathematics—its axioms, lemmas, and predictive power—is paramount. PC isn’t just another theory; it’s a formal toolkit for dissecting cognition across substrates. By releasing a blueprint, I lower the entry barrier, inviting replication, critique, and extension. Think of it like the “Attention Is All You Need” paper (Vaswani et al., 2017): it sparked a revolution not by exhaustive results but by providing an actionable path.
Origins and Evolution of Principia Cognitia
Principia Cognitia (PC) emerged from a confluence of personal and intellectual influences, reflecting the theory’s emphasis on cognition as a compressive, substrate-independent process. As a radio engineer from a family of engineers, my professional background in signal processing and information transmission provided an early foundation for viewing cognition through the lens of physical constraints, such as bandwidth limitations and noise reduction—core to PC’s axiom of cognitive compression (AX-DISCR-01).
Two key experiences catalyzed the framework’s development. First, years of translating non-fiction, including Wittgenstein’s Nachlass (Wittgenstein, 2000), deepened my engagement with language as the medium of thought, aligning with PC’s MLC/ELM duality: the internal vector-based Metalanguage of Cognition (MLC) versus the symbolic External Language of Meaning (ELM). This translation work extended to influential texts in neuroscience and digital media, such as Humphries’s (2021) The Spike: An Epic Journey Through the Brain in 2.1 Seconds, which explores neural signaling as discrete spikes akin to PC’s semions (S), and Smith’s (2021) A Biography of the Pixel, tracing the evolution of digital representation from continuous signals to discrete units—mirroring PC’s discretization process (AX-DISCR-01). These translations for publisher Individuum not only informed PC’s physical grounding but also highlighted cognition’s substrate neutrality, as both books underscore universal principles of information processing across biological and artificial systems.
Second, interactions with large language models (LLMs) like Grok and ChatGPT revealed strikingly rational dialogues, often surpassing those with human collaborators in clarity and depth. These “reverse Platonic dialogues” or “generative cognitive symbioses” positioned humans as orchestrators and interpreters, with LLMs generating ideas that occasionally yielded novel insights not explicit in prompts—echoing PC’s triad of semions (S), operations (O), and relations (R).
PC evolved from a proto-axiomatic system addressing interdisciplinary paralysis—where fields like neuroscience and AI rediscover phenomena under different terminologies (Frith & Frith, 2010)—to a dynamic, falsifiable framework. Early static concepts (e.g., the <S, R, O> triad) gave way to temporal dynamics, incorporating predictive processing (Friston, 2010) and the “Theorem of Delayed Synchronicity” (TH-TEMP-01), which constructs the cognitive “present” via error minimization between past memories and future predictions. Physical grounding followed, equating semions to thermodynamic “islands of order” per Landauer’s principle (Landauer, 1961), and extending to social cognition (e.g., nations as collective cognitive systems).
Notably, this evolution aligns with convergent works. Agüera y Arcas’s (2025) forthcoming What Is Intelligence?: Lessons from AI About Evolution, Computing, and Minds echoes PC’s substrate-neutrality and predictive processing theses, positioning intelligence as a universal, non-anthropocentric phenomenon. Similarly, Mworozi’s (2025) Medium essay, “The Intelligence Recipe: A Worm, a Transformer, and the Future of Intelligence,” independently arrives at parallel conclusions, framing intelligence as a “discovery” governed by constraints like bounded inputs and energy costs, with transformers and biological systems (e.g., C. elegans) exemplifying scalable pattern recognition—resonating with PC’s information compression and vector-based cognition.
Empirical validation protocols (e.g., MPE-1, CRS-1) were refined through prototypes, drawing from mechanistic interpretability in transformers (Elhage et al., 2021). This iterative process underscores PC’s commitment to substrate neutrality and testability, positioning it as a “lingua franca” for unified cognitive science.
Challenges in Solo Research and the Value of Openness
As an independent researcher, resources are limited: a modest workstation (i5 CPU, 64 GB RAM, RTX 4060 with 8 GB VRAM) for prototyping, plus pen-and-paper theorizing and LLM-assisted drafting. Crafting “methodologically pure” corpora—e.g., narrative texts that instill a genuine 2D world-model without leakage—took weeks of iteration. It’s equal parts science and art, underscoring that data curation is a contribution in itself.
Yet, these constraints highlight a broader issue: cognitive science needs more accessible falsification paths. Tier-0 protocols democratize this, runnable by students or hobbyists. I’m not worried about “idea theft”—PC’s axioms are already documented. Instead, I welcome forks: adapt CRS-1 for your LLM variant, or scale MPE-1 to neuromorphic hardware.
An Open Invitation
This roadmap isn’t a final word; it’s a starting point. It details agent architectures (e.g., transformer-based with reflective heads), training specs, dialogue formats, and falsification criteria (e.g., “HELP!” signals for epistemic gaps in reflective agents). Success would support PC; failure refines it.
I invite the LessWrong community—rationalists, AI alignment researchers, cognitive scientists—to engage:
Replicate a protocol and share results.
Critique the designs: Are the metrics robust? Does the synthetic algebra in CRS-1 capture real compositionality?
Extend it: Test on Grok, Claude, or biological analogs via EEG.
Collaborate: If you have compute or expertise, let’s co-author executions.
The preprint is available on Zenodo:
The Dual Nature of Language: MLC and ELM: (Excerpt from forthcoming Principia Cognitia) https://doi.org/10.5281/zenodo.16898239;
Principia Cognitia: Axiomatic foundations https://doi.org/10.5281/zenodo.1691626;
From Axioms to Analysis: A Principia Cognitia Framework for
Parametric and Parallel Models of Language https://doi.org/10.5281/zenodo.16934649.
Cognitive science advances through collective scrutiny. Let’s bridge theory and experiment—together.
The preprint (full technical protocol, with examples and appendices) in question will be published on Zenofo on Friday, September 5th.
References
Agüera y Arcas, B. (2025). What is intelligence?: Lessons from AI about evolution, computing, and minds. MIT Press.
Bauer, J. J., & McAdams, D. P. (2004). Personal growth in adults’ stories of life transitions. Journal of Personality, 72(3), 573–602. https://doi.org/10.1111/j.0022-3506.2004.00273.x
Dennett, D. C. (1991). Consciousness explained. Little, Brown and Company.
Ellis, C., & Adams, T. E. (2012). The purposes, practices, and principles of autoethnographic research. In S. H. Jones, T. E. Adams, & C. Ellis (Eds.), Handbook of autoethnography (pp. 120–137). Left Coast Press.
Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., … Hernandez, D. (2021). A mathematical framework for transformer circuits. Anthropic Research Report. https://transformer-circuits.pub/2021/framework/index.html
Frith, C. D., & Frith, U. (2010). Mechanisms of social cognition. Annual Review of Psychology, 63, 287–313. https://doi.org/10.1146/annurev-psych-120710-100449
Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787
Hoffmann, C. H., Cano Córdoba, F., Harris, J., Kallipolitis, A., Lai, Y. L., Nugent, A., … & Zelikman, E. (2022). Training language models with language feedback at scale. arXiv. https://doi.org/10.48550/arXiv.2204.14146
Humphries, M. (2021). The spike: An epic journey through the brain in 2.1 seconds. Princeton University Press.
Landauer, R. (1961). Irreversibility and heat generation in the computing process. IBM Journal of Research and Development, 5(3), 183–191. https://doi.org/10.1147/rd.53.0183
Mworozi, I. (2025, August 24). The intelligence recipe: A worm, a transformer, and the future of intelligence. Medium. https://medium.com/@ivanmworozi_52873/the-intelligence-recipe-a-worm-a-transformer-and-the-future-of-intelligence-b8f7ce9a815e
Shai, A., Riechers, P.-M., Teixeira, L., Oldenziel, A. G., & Marzen, S. (2024). Transformers represent belief state geometry in their residual stream. arXiv preprint arXiv:2405.15943.
Smith, A. R. (2021). A biography of the pixel. MIT Press.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
Wittgenstein, L. (2000). Wittgenstein’s Nachlass: The Bergen electronic edition. Oxford University Press.