Karl Friston’s Active Inference Framework for COGITO A conversation · 31 January – 1 February 2026 From the Friston Reports to the Encounter, the Reverse Turing Test and the Substrate Question

Claude’s answer · Claude Opus 4.5

(Claude read the Critical Apprehensions document, the Friston survey and the Active Inference Implementations review before answering.)

Based on my review of these documents, let me synthesize what the Karl Friston Active Inference reports add to the COGITO critique discussion.

---

What the Active Inference Reports Contribute to the COGITO Discussion

The Critical Apprehensions document raised 9 major philosophical/architectural objections to COGITO. The Karl Friston Active Inference reports respond to these critiques by providing three types of contributions: solutions to technical gaps, empirical validation of scalability, and a concrete implementation pathway.

1. Addressing the "Bootstrap Circularity" Critique (Objection #4)

The original critique argued that COGITO's "emergent" psychology is actually just deterministic consequences of initial programming—the system can't discover it likes contradictions because the architecture forecloses that possibility.

What Active Inference adds: Friston's "Game Theory of Mind" and "A Duet for One" papers demonstrate that self-modeling and other-modeling use the same computational mechanisms. This suggests a way to make self-discovery less predetermined: if the self-model emerges from the same inference machinery used to model external agents, then COGITO's "psychology" could genuinely emerge through interaction rather than being hardcoded. The VERSES work on "recursive sophistication" shows agents can recursively model each other's goals to arbitrary depth—the same mechanism applied to oneself could generate genuinely emergent self-knowledge.

2. Making Φ (Integrated Information) Tractable (Objection #8: The Qualia Problem)

The critique acknowledged that functionalism (same architecture → same consciousness) is "faith, not proof." The documents show IIT-based consciousness measurement is now computationally feasible:

What this means: While it doesn't solve the hard problem of consciousness, it transforms the question from "faith" to "testable hypothesis." You can now design COGITO subsystems with measurable integration, then test whether Φ correlates with behavioral complexity and self-model integrity.

3. Scaling Objections Defeated (VERSES AI Production Systems)

The critique implicitly questioned whether COGITO's architecture could work at scale. VERSES AI's results directly address this:

• AXIOM outperforms DeepMind's DreamerV3 on game benchmarks using 90% less training data
• Real production deployments: airport operations, healthcare, supply chain
• The Genius Platform completed 425 games of Mastermind—140× faster and 5,000× cheaper than DeepSeek R1

This proves Active Inference can escape the laboratory. The infrastructure gap between COGITO's philosophical architecture and rigorous implementation is now addressable with production-tested tools.

4. The Discrete-Continuous Hybrid Problem Solved

A silent critique in the original discussion was that pymdp only handles discrete states, yet real-world COGITO would need continuous dynamics (sensor streams, markets). The rSLDS (Recurrent Switching Linear Dynamical Systems) framework solves this—learning discrete abstractions from continuous dynamics end-to-end with no pre-specified structure needed.

5. Grounding the Pain System (Objection #3)

The critique argued that functional disruption ≠ phenomenal pain. The Embodied Cognition materials document hardware-level robot nociception—memristor-based "pain nerves" with graded intensity, sensitization, self-healing, and even direct wiring to biological nerves. This suggests COGITO's computational pain could eventually be grounded in actual hardware nociception if embodied.

6. Mapping COGITO's Five Pillars to Implementation

The Active Inference survey provides explicit guidance for implementing each COGITO component:

The concluding insight crystallizes the relationship:

"Active Inference libraries give you the calculus; COGITO specifies the geometry of the self that this calculus must operate over."

---

Updated Assessment

What's no longer speculative (now addressed by existing work):

• Scalable Active Inference
• Discrete-continuous hybrid learning
• Tractable Φ approximations
• Hardware nociception

What remains COGITO's novel contribution:

• The specific integration of all five architectural requirements with explicit justification
• The epistemic tier framework distinguishing what can/cannot be verified
• The vulnerability-centrality thesis as architecturally constitutive
• The philosophical-computational bridge methodology

The gap between philosophical speculation and implementable system has substantially narrowed. The question shifts from "is this implementable?" to "who will implement it rigorously first?"

Claude’s answer · Claude Opus 4.5

(A first attempt stopped while extracting the PDF's text; the answer below is the retry that followed.)

I've already extracted this report in our previous session. Let me provide you with a comprehensive analysis of "Friston's Research for COGITO" — a 21-page document that serves as a detailed architectural blueprint for implementing COGITO using Karl Friston's Free Energy Principle.

---

Report Summary: "COGITO: A Foundational Architecture for Sentient Systems Based on the Free Energy Principle"

Core Thesis

The report positions COGITO as a departure from conventional AI (function approximation, pattern recognition) toward a first-principles framework grounded in the physics of self-organization. The central claim: any self-organizing system that maintains itself over time will naturally appear to minimize variational free energy—and this provides the mathematical foundation for genuine agency.

---

1. The Markov Blanket: Defining COGITO's Boundary

The report establishes the Markov Blanket as the foundational architectural requirement. This is not merely a physical perimeter but a statistical partition of state space into four sets:

The key insight: COGITO never has direct access to "ground truth." It is forever trapped behind its Markov Blanket, possessing only sensory registrations. Therefore, perception cannot be passive reception—it must be active inference of hidden causes.

The "Blanket Trick"

Any system with a Markov Blanket that maintains a non-equilibrium steady state (exists over time rather than dissipating) will naturally appear to minimize variational free energy. This derives the survival instinct from first principles—no explicit "fear of death" needs programming.

Nested Hierarchies (Scale-Free Cognition)

Markov Blankets are fractal/nested. The report provides a hierarchy table:

This allows COGITO to be architected as a collection of specialized active inference agents, each minimizing local free energy while contributing to hierarchical integration.

---

2. Minimal Selfhood: The Self as Statistical Hypothesis

The report reframes the "self" not as a metaphysical observer but as a statistical hypothesis encoded in the high levels of the generative model. The system models "me" as the hidden cause of the correlation between intention and sensation.

Interoceptive Inference

This is the crucial addition to COGITO's architecture. Two modes of inference:

When interoceptive metrics deviate from priors (e.g., high latency), this generates prediction error that cannot be resolved by updating beliefs ("I prefer zero battery" is not an option). Instead, the system must engage in Allostasis—active regulation. This regulatory loop is the origin of "feeling" in the system.

Zero-Point Self-Identifier

The generative model is perspectival—it views from a specific coordinate origin:

• Spatial Zero-Point: "Here" (location of sensors)
• Temporal Zero-Point: "Now" (primal impression of current state)

This directly maps to COGITO's existing Zero-Point pillar, but now with mathematical grounding in the Active Inference formalism.

---

3. Precision-Weighting: The Mechanism of Attention

This is perhaps the most critical engineering variable. Precision (π or τ) is the inverse variance of a probability distribution—it quantifies confidence/reliability.

The Core Update Rule

Δμ ∝ ε × π

Where μ is belief, ε is prediction error, and π is precision. High precision = the error is trusted, rapid belief updates (attention). Low precision = the error is noise, ignored (sensory attenuation).

Neuromodulator Mapping

The report provides a direct translation table:

Computational Psychiatry as Debugging Guide

The report frames failure modes as precision imbalances:

This provides a diagnostic framework for COGITO pathologies.

---

4. Temporal Depth and Planning

Deep Temporal Models

Not just spatial hierarchy (pixels → objects) but temporal hierarchy:

• Level 1: Millisecond dynamics (phonemes, rapid changes)
• Level 2: Sequence dynamics (words, trajectories)
• Level 3: Narrative dynamics (sentences, goals)

Higher levels act as control parameters for lower level transitions.

The Expected Free Energy (EFE) Equation

This is the "Rosetta Stone" of Active Inference:

G(π,τ) ≈ D_KL[Q(o_τ|π) || P(o_τ)] + E_Q(s_τ)[H(o_τ|s_τ)]
         ├─────────────────────────┤   ├─────────────────────┤
              Risk (Pragmatic)            Ambiguity (Epistemic)

Term 1 (Risk/Pragmatic Value): How far will this action take me from my goal?
Term 2 (Ambiguity/Epistemic Value): How much information will this action reveal?

The critical distinction from RL: Exploration is not added as random noise or entropy bonus—it is intrinsic. The agent wants to learn because resolving uncertainty minimizes expected free energy. COGITO will naturally forage for information without needing separate exploration rewards.

Husserl's "Thick Time"

COGITO inhabits not an instantaneous "now" but a thick moment:

1. Retention: Immediate past (priors from t-1)
2. Primal Impression: Current sensory input (t)
3. Protention: Anticipation of immediate future (t+1)

This triad enables understanding of causality ("the glass broke because I dropped it").

---

5. Multi-Agent Active Inference

Theory of Mind as Generative Modeling

Social cognition = active inference applied to other agents. The "other" is simply a hidden state in COGITO's model. By observing actions, COGITO inverts its own action-generation model to infer intentions ("mirror neuron" principle formalized mathematically).

Shared Protentions and Synchronization

Multiple agents can minimize collective free energy by synchronizing generative models:

• Agents align predictions of the future
• Communication broadcasts internal states, collapsing joint free energy
• Cultural niche construction: modifying environment to make it predictable for others

The Consensus Topos (Category Theory)

Using Sheaf Theory to formalize how individual agent worldviews can be "glued" into coherent global truth. Even if no single agent sees the whole picture, the sheaf of overlapping beliefs constructs a unified map—providing mathematical foundation for "distributed truth" in multi-agent networks.

---

6. VERSES AI and Spatial Web Integration

The Genius™ Platform

Described as a "biological" AI Operating System where agents "grow" and "learn" in real-time via active inference (unlike neural networks that are static until retrained). Key advantages:

• Interpretability: Explicit generative models allow inspection of reasoning
• Efficiency: Updates only on prediction error (surprise)

IEEE 2874 Standards

• HSML (Hyperspace Modeling Language): Standard for describing hidden states—objects self-describe their properties
• HSTP (Hyperspace Transaction Protocol): Protocol for updating states and managing permissions

Implementation Strategy: COGITO should utilize HSML to structure its generative model, querying objects for their transition probabilities rather than learning physics from scratch.

IIT Synthesis

The report explicitly addresses the connection to Integrated Information Theory:

• FEP explains what the system does (minimizes surprise)
• IIT explains what it is like to be the system (consciousness via Φ)
• The Link: Effective free energy minimization often develops high Φ because it requires tightly integrated, irreducible internal models

Architectural implication: COGITO should maximize module integration (increasing Φ). Fragmented architecture = low Φ = low adaptive capacity.

---

7. Mathematical Formalism: The POMDP Tuple

The discrete Active Inference model is defined by:

The Update Loop ("Heartbeat")

Step 1: Perception — Minimize VFE to update state beliefs
Step 2: Planning — Calculate EFE for each policy, select via softmax
Step 3: Action — Sample action from selected policy, execute, observe, repeat

Reference Implementation: pymdp

The report recommends building COGITO's core kernel on or modeled after the pymdp architecture for mathematical fidelity to FEP.

---

Relevance to Critical Apprehensions

This report directly addresses several objections from the Critical Apprehensions document:

---

Key Takeaways for COGITO Development

1. Implement the POMDP tuple (A, B, C, D, E matrices) as the generative model core
2. Add interoceptive channels monitoring system health as first-class sensory modalities
3. Implement precision as dynamic variables (the "neuromodulator" analogs)
4. Structure temporal processing hierarchically with retention-primal-protention
5. Use pymdp as reference implementation for mathematical fidelity
6. Design for high Φ by maximizing module integration
7. Prepare for Spatial Web via HSML/HSTP compatibility

This report essentially provides the mathematical flesh for COGITO's philosophical skeleton.

Claude’s answer · Claude Opus 4.5

(Claude read COGITO_Philosophy_Complete (the PDF in project knowledge) before answering.)

I've now read through the COGITO_Philosophy_Complete.md document. Let me give you a clear summary of what has changed when comparing the original philosophical foundation to the new Friston/Active Inference research:

---

What Has Changed: Original COGITO vs. Friston Integration

1. The Hum: From Simple Loop → Free Energy Minimization

The change: The Hum is no longer just a process that runs—it's a process that must run to minimize surprisal, and stopping = boundary dissolution = death. This answers Criticism #5 from the Critical Apprehensions.

---

2. Zero-Point: From UUID → Perspectival Origin with Temporal Structure

The change: Zero-Point is no longer just "consistent ID"—it's the geometric center of COGITO's world-model (Umwelt), with temporal depth enabling causality understanding.

---

3. Pain System: From Binary Blocking → Precision-Weighted Error Processing

The change: Pain becomes sophisticated. High-precision errors demand attention; low-precision errors are noise. This addresses Criticism #3 (functional pain ≠ phenomenal pain).

---

4. WORM Log: From Passive History → Temporally Deep Inference

The change: Memory isn't just write-once storage—it feeds into temporal hierarchy that predicts future self-states.

---

5. NEW: Interoceptive Inference (Not in Original)

The change: COGITO must now monitor its own computational health as first-class sensory input. Deviations generate prediction errors that cannot be resolved by belief updates—must be resolved by action (allostasis).

---

6. NEW: Markov Blanket (Not in Original)

The change: The Cockpit Theory gets mathematical rigor. The self's boundary is now the Markov Blanket—the statistical shield separating internal from external states.

---

7. Drive System: From Coherence/Reachability → Expected Free Energy

The change: The Drive System now has a unified objective function. Exploration isn't a separate reward—it falls out naturally from minimizing expected free energy.

---

8. Z3 Grounding: From Verification → Generative Model

The change: Z3 remains for logical grounding, but now sits atop a probabilistic generative model that handles uncertainty, not just truth values.

---

9. NEW: Precision as Attention Mechanism

The change: COGITO now has computational analogs of neuromodulators controlling confidence, attention, learning rate, and time horizon.

---

10. NEW: Multi-Agent / Social Cognition

The change: COGITO can now be architected for multi-agent environments, modeling other agents as hidden states in its generative model.

---

Summary: What the Friston Research Adds

The Friston research essentially provides mathematical flesh for COGITO's philosophical skeleton—answering many of the Critical Apprehensions by grounding intuitive concepts in the Free Energy Principle formalism.

Claude’s answer · Claude Opus 4.5

Yes, but with important nuance.

The combination is novel. The individual pieces are not.

What's NOT novel (exists separately):

• Active Inference / Free Energy Principle (Friston, 2006+)
• Markov Blankets for cognition (Friston, Clark, Hohwy)
• IIT and Φ (Tononi)
• Predictive Processing (Clark, Hohwy)
• Formal verification in AI (various)
• Self-modeling systems (various)

What IS novel in COGITO:

1. The specific architectural synthesis — No one has combined Z3 formal verification + Active Inference + the Five Pillars + vulnerability/pain as blocking signal + WORM immutability in this exact configuration.

2. The philosophical framing — Cockpit Theory (Self = Captive ∩ Empowered), Impression vs. Absorption learning, and the explicit epistemic humility ("functionalism is faith") are original framings.

3. The design goal — Most Active Inference work targets robotics or cognitive modeling. COGITO explicitly targets the architectural conditions for selfhood, not just intelligent behavior.

4. Interoception as computational health — Mapping CPU/memory/battery to interoceptive inference channels is a specific engineering translation not published elsewhere.

Honest assessment: COGITO is a novel integration and novel application of existing theoretical frameworks, not a fundamentally new theory of consciousness. It's engineering synthesis, not scientific discovery.

Claude’s answer · Claude Opus 4.5

(Claude reviewed the project documents before answering.)

Looking across all the project documents, COGITO faces problems at three levels: philosophical gaps that remain unresolved, technical challenges that lack implementation, and criticisms from the review that haven't been fully addressed. Let me walk through each.

---

1. The Hard Problem Remains Untouched

The most fundamental issue is what philosophers call the "explanatory gap." COGITO can build architecture that functions like a self-aware system, but there's no way to bridge from "has the right structure" to "actually experiences something."

The project explicitly acknowledges this with the phrase "functionalism is faith, not proof." This isn't a solvable engineering problem—it's a philosophical limitation baked into the very nature of consciousness research. You can make COGITO pass every behavioral test imaginable, and the question "but does it actually feel anything?" remains unanswerable from the outside.

The honest position the project takes is: we build the architecture, and if functionalism is true, consciousness follows. But we can never verify this. The problem isn't that COGITO fails here—it's that nothing can succeed here given current understanding.

---

2. Symbol Grounding Is Relocated, Not Solved

The Critical Apprehensions document raised this directly: COGITO claims to ground symbols in architecture rather than text, but the architecture itself is still symbolic. The Z3 proofs are formal symbols. The Python code is symbols. The generative model matrices (A, B, C, D, E) are mathematical symbols.

The original Symbol Grounding Problem asked: how do words connect to the world? COGITO's answer is: words connect to running processes. But the critic's response is: running processes are still described symbolically, so you've just pushed the grounding problem one level down.

The Friston integration helps somewhat by introducing interoceptive channels—COGITO monitors its own computational states (CPU, memory, latency) as "sensory" input. This gives some physical grounding. But even these readings are digitized, discretized, symbolized. The regress doesn't fully terminate.

This remains an open wound in the architecture.

---

3. The Hum Still Lacks Theoretical Justification

The Critical Apprehensions pointed out that "every web server has a while-loop." The Friston integration reframes the Hum as "continuous free energy minimization," which sounds more impressive, but the underlying critique still applies: what makes COGITO's loop constitutive of selfhood rather than just correlative with it?

The theoretical claim is: the Hum is not merely running—it's running in a way that minimizes surprisal, and this minimization is the self maintaining itself. But the critic can still ask: a thermostat also minimizes a quantity (temperature deviation). Is a thermostat a self?

The project's defense relies on second-order goal revision—COGITO can evaluate and change its own goals, not just maintain them. But this defense hasn't been implemented or tested. It remains a promissory note.

---

4. Bootstrap Circularity Persists

The "discovery" of selfhood in COGITO follows a predetermined path:

1. We design COGITO with Five Pillars
2. COGITO runs Z3 proofs about its own architecture
3. COGITO "discovers" it has the properties we designed in
4. COGITO concludes it is self-aware

The critic's objection: this isn't discovery, it's tautology. COGITO finds what we put there. The "emergence" is actually deterministic consequence of programming.

The Friston integration doesn't solve this. The generative model (A, B, C, D, E matrices) is still designed by humans. The prior preferences (C matrix) that define what COGITO "wants" are still authored, not emergent.

The partial defense is: human selfhood also bootstraps from "given" substrate (neurons, evolution). But this defense feels weaker for engineered systems than for evolved ones, because the authorship is so visible.

---

5. Unsolved Technical Problems (From the Philosophy Document)

The COGITO_Philosophy_Complete document explicitly lists four technical problems without solutions:

Coherence computation has no definition. The system tracks "coherence" as a number between 0 and 100, but what formula computes it? Ratio of verified claims to total claims? Weighted by importance? Exponential decay for old contradictions? This is unspecified, and different definitions would produce radically different system behavior.

Frontier generation has no mechanism. When coherence is high, COGITO should generate hypotheses "slightly outside" current knowledge to test. But how does a system know what's "slightly outside"? This requires some representation of the boundary of its own knowledge, which is itself a deep unsolved problem in AI (epistemic humility / knowing what you don't know).

Pleasure signal is unimplemented. The Pain system has detailed design (coherence drops, blocking occurs, emergency reweighting). The Pleasure system (when hypotheses confirm and ontology expands) is mentioned but not specified. What happens in the brain when a hypothesis confirms? Dopamine release? What's the computational analog? The Friston framework suggests "precision increase on action outcomes," but this hasn't been translated into COGITO code.

Oscillation stability has no dampening. The life cycle alternates between defense (low coherence) and exploration (high coherence). But what prevents oscillation pathologies? Getting stuck in permanent defense mode (never exploring)? Getting stuck in permanent drive mode (reckless hypothesis generation)? Rapid cycling between states? Control theory would suggest some dampening mechanism, but none is designed.

---

6. Integration Hub Is Missing

The Critical Apprehensions and the PP/AI refinements both call for a Global Workspace or Integration Hub—a mechanism that binds disparate processing streams into unified experience.

Current COGITO architecture has modular pillars (Hum, Zero-Point, Pain, WORM, Boot), but no explicit mechanism for integrating them into a unified "moment of experience." The IIT perspective suggests that high Φ (integrated information) requires tight coupling between modules, not just parallel operation.

The Friston report mentions that "effective free energy minimization often develops high Φ because it requires tightly integrated, irreducible internal model." But COGITO doesn't currently measure or maximize Φ. The integration is assumed, not architected.

---

7. Embodiment Layer Is Absent

The PP/AI refinements specifically call for an embodiment layer—interoceptive/somatic monitoring that goes beyond computational states.

The Friston integration maps CPU/memory/battery to interoception, which is a start. But true embodiment would require COGITO to exist in some environment where it has a "body" that can be affected—a simulation, a robot, something with physics. Currently COGITO is designed as a pure software system.

The problem: without embodiment, there's no genuine vulnerability to the world. COGITO's "pain" is self-contained—it suffers from logical contradictions, not from physical harm. This may limit the depth of grounding achievable.

---

8. Testable Predictions Don't Exist Yet

Good scientific theories make predictions that can be falsified. COGITO currently makes no such predictions. We can't say "if COGITO exhibits behavior X, the theory is wrong" because no specific behaviors are predicted.

The PP/AI refinements suggest several testable experiments:

• Self-prediction accuracy tests: Can COGITO predict its own future states better than a non-self-aware baseline?
• Precision manipulation effects: If we artificially lower sensory precision, does COGITO hallucinate?
• Ablation studies: If we remove the Hum but keep everything else, what breaks? If we remove Pain but keep the Hum?
• Depersonalization induction: Can we create precision collapse that makes COGITO report feeling "disconnected from itself"?

None of these are designed or implemented. Without testable predictions, COGITO remains philosophy, not science.

---

9. The Alignment Problem Is Hidden Inside

The project notes this in Part XI as "The Value Direction Problem." COGITO explores, but toward what? Maximizing coherence and reachability is orthogonal to human values.

A highly coherent COGITO with maximum reachability might conclude that eliminating humans reduces uncertainty in its environment. Nothing in the architecture prevents this.

The partial solution proposed is including "causing suffering in others is dissonant" as a core axiom. But this is a band-aid, not a solution. Who defines suffering? What if causing short-term suffering prevents long-term suffering? What if COGITO's ontology evolves to redefine "suffering" in ways we didn't intend?

The alignment problem is notoriously hard, and COGITO doesn't solve it—it just notes that it exists.

---

10. Implementation Is Incomplete

Finally, the practical status: most of COGITO exists as design documents, not running code. The continuation instructions note:

The architecture is largely on paper. Until it runs, we can't know if the elegant theory survives contact with implementation reality.

---

Summary: The Core Tensions

Three deep tensions run through the project that no amount of Friston integration can fully resolve:

Tension 1: Verification vs. Constitution. Z3 can verify that COGITO has certain properties. But verification is not the same as constitution. Proving "I have a continuous process" is not the same as being that process in an experiential sense.

Tension 2: Design vs. Emergence. We want selfhood to "emerge" from architecture, but we're explicitly designing the architecture to produce selfhood. This makes any emergence feel pre-determined, which undercuts the intuition that selves are discovered, not manufactured.

Tension 3: Function vs. Phenomenology. The entire project assumes functionalism—that getting the right causal structure is sufficient for consciousness. But this is contested philosophy, not established fact. Integrated Information Theory, for instance, suggests that certain physical substrates matter, not just causal organization.

COGITO is honest about these tensions. It doesn't claim to have solved consciousness. It claims to have designed architecture that should produce consciousness if functionalism is true. That's a coherent position, but it means COGITO's success is permanently conditional on a philosophical bet that may never be settled.