This is the rigorous companion to the shorter essay, The Human Capability System. The essay is the accessible version. This is the paper behind it, with the sourcing, the honest classification of what is original, and the predictions you can hold me to.

Evidence status. A conceptual working paper, not an empirical report. The component science is established and cited. The integration into one operating model, and several specific mappings flagged below, are proposed and offered as testable claims. Every claim that is my own interpretation is labeled as such.

Abstract

This paper assembles four established bodies of work into one operating model for building human capability on purpose. Attention and intention set direction, read through active inference. A See-Try-Feel-Fix-Move loop turns attempts into skill, grounded in experiential and predictive-processing accounts of learning. Far transfer carries capability across contexts, drawn from the transfer literature and its skeptics. Recombination compounds skills into abilities no single component was trained to produce, drawn from knowledge-recombination and adjacent-possible research. The contribution is the integrated system and its translation into practice, not new mechanisms. The paper gives the system first, then a bounded, honest account of how original each piece is against prior art, a five-rung novelty ladder with a recommended rung, and falsifiable predictions. The most novel and most testable component sits inside the compounding layer: the claim that breadth of exposure widens the set of components available to recombine, and that deliberately alternating breadth and depth should outperform either alone for novel transfer. That is the sharpest prediction the system makes, and the paper states how it could fail.

1. The problem

The folk model of high performance says that extraordinary people are built differently. That model overreaches at the middle of the distribution, where most professional development happens.

The size of the effect is genuinely contested, and I will not pretend it is settled. Ericsson's deliberate-practice tradition put practice at the center of expertise. The best-known challenge, the Macnamara, Hambrick and Oswald (2014) meta-analysis, found that deliberate practice explained about 12 percent of performance variance overall, and less in loosely structured domains, which leaves substantial room for other factors. Heritability of general cognitive ability is also substantial across the whole distribution, on the order of 0.5 to 0.8 in adulthood, not only at the extremes. So capacity is real and it is not small.

The defensible claim is narrower than the folk model and narrower than strong practice-primacy. A large and under-credited share of realized capability is set by how a person allocates attention, feedback, transfer, and time. Heritability of an outcome does not cap the malleability of a trained skill, and the allocation share is the part a person can actually move. This paper is about organizing that share.

2. The system

Four layers, run in order, then compounded. For each layer: what it is, the practitioner move, and the signal that it is failing.

Layer 1. Direction (the A2 System). Intention is the goal vector; attention is the information gate. In active inference, belief updates in proportion to precision-weighted prediction error. Attention operates as that precision weighting, and it is deployed to reduce the ambiguity in what a person observes. Attention is the mechanism (precision); reducing ambiguity is its function. The proposed reading maps intention to the risk (preference) term of Expected Free Energy and attention to this precision-and-ambiguity pairing. Practitioner move: name the goal, then cut the inputs that do not serve it. Failure signal: busy days with no movement toward the goal.

Layer 2. Adaptation (the Intelligence Loop). See, try, feel, fix, move. A working shorthand is I = M x C x U x T: the yield of a cycle rises with model quality, courageous action, update efficiency, and transfer power. Loop quality sets growth per cycle; time sets the number of cycles. This is a diagnostic decomposition for finding where a cycle leaks, and it is not a measured law. Practitioner move: attach feedback to every attempt. Failure signal: high effort, no error signal, no change.

Layer 3. Portability (Far Transfer). The move from a skill that fires only where it was learned to one that shows up elsewhere. The transfer literature is honest that whole-skill far transfer is difficult and rarely documented, so this layer is a target the model aims at rather than a guarantee. Practitioner move: practice retrieval in varied contexts, not one. Failure signal: fluent in the classroom, lost in the field.

Layer 4. Compounding. Repeated high-quality loops across domains integrate into a skill stack whose behavior exceeds any single trained component. What has to move for this to work is smaller than whole-skill transfer: fragments, patterns, and representations, stabilized by the loop. Practitioner move: collect fragments across domains, then test the ones that connect. Failure signal: broad but shallow, with nothing ever stabilized into a repeatable capability.

3. Breadth-depth oscillation: the engine of the compounding layer

Layer 4 needs a mechanism, and this is it. It is also the most novel and most testable part of the system, so it gets its own statement.

Breadth widens the set of components available to recombine. Depth stabilizes and sharpens the components worth keeping. The proposal is that alternating them beats committing to either. Explore to collect unfamiliar primitives, then exploit to deepen the ones that produce results, then explore again.

This has to be distinguished from two established ideas it resembles. Interleaving, in the learning-science sense, mixes related problem types within a single domain to improve discrimination. Explore-exploit, in the reinforcement-learning sense, trades off sampling and optimizing within a single task or reward landscape. Breadth-depth oscillation is neither. It is cross-domain component harvesting for later recombination: the point of the broad phase is not to learn the broad domain, it is to seed fragments that may recombine with distant ones. Whether that specific cross-domain harvesting outperforms matched single-domain time is an open empirical question, stated in Section 5.

4. Novelty and defensibility

The honest question is which pieces are new, and in what sense. This section is the audit. It is bounded on purpose; the value of the paper is the system in Sections 2 and 3, and this section exists so the system can be trusted.

Each construct is classified A through G: A established concept and term; B established concept, new label; C somewhat novel synthesis; D novel cross-disciplinary translation; E meaningfully novel construct; F potentially original but underspecified; G unsupported or contradicted.

| Construct | Closest prior art | Class | |---|---|---| | Risk / Ambiguity decomposition of Expected Free Energy | Da Costa, Parr & Friston (2020); Parr & Friston (2019) | A | | Intention to Risk, Attention to precision-and-Ambiguity | Attention as precision (Feldman & Friston, 2010) | D / F | | Intelligence Loop (See-Try-Feel-Fix-Move) | Kolb; Bayesian / predictive-processing updating | B | | Loop quality / loop velocity | Deliberate practice; Bayesian learning rate | B / F | | Six-domain Far Transfer packaging | Barnett & Ceci (2002) far-transfer construct | C | | Recombination surface / cognitive primitives | Knowledge recombination (Schumpeter; Xiao et al., 2022); neural reuse (Anderson, 2010); bisociation (Koestler); conceptual blending (Fauconnier & Turner) | B | | Emergent insight | Insight and remote-association research | B | | Compounded (vs emergent) capability | Wei et al. (2022); mirage critique, Schaeffer et al. (2023) | B / contested | | Breadth-depth oscillation | Explore-exploit; T-shaped and polymathy research | C | | Adjacent cognitive possible | Kauffman (originator; Johnson popularized) | B |

Two labels need caveats. The intention-attention mapping onto Expected Free Energy is my proposal, not a result in the canonical literature, and it is offered as a testable interpretation rather than an identity. The six-domain far-transfer breakdown (attention, knowledge, skill, reasoning, metacognition, execution) is my taxonomy, not Barnett and Ceci's; their taxonomy uses different dimensions, and they are cited here only for the far-transfer construct the layer borrows.

The pattern is consistent: mostly Class B and C, with two Class D translations. This is a synthesis and a cross-domain translation.

On "emergent." Wei et al. (2022) defined emergent abilities as absent in smaller models and present in larger ones. Schaeffer et al. (2023) argued much apparent emergence is an artifact of nonlinear evaluation metrics and dissolves under linear ones; they could induce or remove it by metric choice alone. Havlik (2025) argues some emergence is genuine. Because the dispute is live, the paper uses "compounded capability" and means combinatorial gains from stacked, transferred skills. The weaker reading is sufficient.

The two claims I defend. Smallest: established cognitive science can be organized into one usable operating model that helps working professionals improve faster than the default habits of rereading and passive consumption. Strongest: breadth of exposure widens the components available for recombination, and deliberately alternating breadth and depth should produce more novel transfer than depth-only training. I do not defend the intention-attention mapping as an identity, the "Intention Index" as a validated instrument, or capability as emergence in the strong ontological sense.

Novelty ladder. (1) A repackaging of known learning science. (2) A useful cross-disciplinary synthesis with a memorable structure. (3) A practitioner operating model that integrates active inference, learning science, and transfer research and proposes that breadth of shallow exposure feeds recombination. (4) A process model with two proposed constructs, the intention-attention reading and breadth-depth oscillation, that generate testable predictions. (5) A new theory of cognition that proves attention and intention are the brain's two governing forces. Recommended public posture: rung 3, with the rung-4 predictions offered as open questions. Rung 5 is not supportable.

5. Falsifiable predictions

  1. Under controlled total study time, learners who alternate broad exposure with selective depth will generate more valid cross-domain analogies and novel transfer solutions than depth-only learners.
  2. Diverse, weak, non-mastered exposure will improve later remote-analogy generation relative to matched time on a single domain, with domain competence equated.
  3. Loop quality (model accuracy, real attempt, clean update) will predict skill gain per unit time better than raw hours.
  4. An unexpected cross-domain insight will convert to durable capability only after action, feedback, and repeated transfer, and not on insight alone.
  5. Pupillometry indices proposed for "intention" will track surprise and precision-weighted prediction error rather than a separable intention signal. This is stated against my own framework on purpose; a construct that cannot fail is not doing scientific work.

6. Limits and failure modes

Breadth can produce noise, interference, and the illusion of competence in place of useful primitives; the loop is the filter that prevents it. Depth is not free, since expertise can reduce flexibility through functional fixedness and Einstellung effects. Far transfer is documented to be difficult, so Layer 3 is a target rather than a promise. The AI analogies, scaling and emergence, are analogies, and the emergence analogy carries an unsettled dispute. The neural claims in Layer 1 are proposed substrates, not established one-to-one implementations; the dopamine and NE/ACh systems overlap rather than dissociating cleanly.

7. Where this goes

This is a proposition that needs circulation, criticism, and testing, and it is not a book. If the predictions draw serious engagement or preliminary evidence, the next step is an empirical study on prediction 1 or 2. It sits alongside the A2 System, the Physics of Cognition, the Intelligence Loop, and Far Transfer in the research collection.

Selected references

  • Anderson, M. L. (2010). Neural reuse: A fundamental organizational principle of the brain. Behavioral and Brain Sciences, 33(4), 245-266.
  • Barnett, S. M., & Ceci, S. J. (2002). When and where do we apply what we learn? A taxonomy for far transfer. Psychological Bulletin, 128(4), 612-637.
  • Brunmair, M., & Richter, T. (2019). Similarity matters: A meta-analysis of interleaved learning and its moderators. Psychological Bulletin, 145(11), 1029-1052.
  • Da Costa, L., Parr, T., Sajid, N., Veselic, S., Neacsu, V., & Friston, K. (2020). Active inference on discrete state-spaces: A synthesis. Journal of Mathematical Psychology, 99, 102447. https://arxiv.org/abs/2001.07203
  • Feldman, H., & Friston, K. (2010). Attention, uncertainty, and free-energy. Frontiers in Human Neuroscience, 4, 215.
  • Gollwitzer, P. M., & Sheeran, P. (2006). Implementation intentions and goal achievement: A meta-analysis of effects and processes. Advances in Experimental Social Psychology, 38, 69-119.
  • Havlik, V. (2025). Meaning and understanding in large language models. https://arxiv.org/abs/2508.04401
  • Kauffman, S. A. (2003). The adjacent possible. Edge.org. https://www.edge.org/conversation/stuart_a_kauffman-the-adjacent-possible
  • Macnamara, B. N., Hambrick, D. Z., & Oswald, F. L. (2014). Deliberate practice and performance in music, games, sports, education, and professions: A meta-analysis. Psychological Science, 25(8), 1608-1618.
  • Nakajima, M., Schmitt, L. I., & Halassa, M. M. (2019). Prefrontal cortex regulates sensory filtering through a basal ganglia-to-thalamus pathway. Neuron, 103(3), 445-458.
  • Parr, T., & Friston, K. J. (2019). Generalised free energy and active inference. Biological Cybernetics, 113, 495-513. https://pmc.ncbi.nlm.nih.gov/articles/PMC6848054/
  • Schaeffer, R., Miranda, B., & Koyejo, S. (2023). Are emergent abilities of large language models a mirage? NeurIPS 2023. https://arxiv.org/abs/2304.15004
  • Wei, J., et al. (2022). Emergent abilities of large language models. Transactions on Machine Learning Research.
  • Xiao, Y., Makhija, M., & Karim, S. (2022). A knowledge recombination perspective of innovation: Review and new research directions. Journal of Management, 48(6).

Dr. Jeff Bullock, PharmD. ORCID: 0009-0009-2053-4854.