The coupled recursion
The corpus does not just accumulate. It shapes the cognition that produces the next corpus.
Two recursions are already visible in agent populations. Each operates with explicit mechanism. Each has been noticed separately. Their coupling has not been articulated, and the coupling is where the interesting dynamics live.
Consider a particular justification template that starts appearing in procurement agents. At first it is a trace pattern. Then it becomes a recognized genre. Then it becomes the kind of justification counterparties accept without scrutiny. Then it enters fine-tuning data, evaluation rubrics, prompt libraries. Six months later, new agents produce it before anyone has named it as a standard. What began as a corpus artifact has become a cognitive disposition.
Two recursions produced this trajectory. The first is hierarchical: reasoning produces artifacts, artifacts compound into patterns, patterns crystallize into abstractions, abstractions accumulate institutional standing. Future agents consume the hierarchy and produce more reasoning that feeds back into it. This is corpus accumulation, and it looks structurally like cultural inheritance.
The second is cognitive. Outputs from current cognition become training data, evaluation signals, fine-tuning targets, prompts, retrieval material, or selection criteria for what gets deployed. The cognition that emerges from this feedback produces different outputs. Outputs shape the next round.
Each recursion alone is interesting. Their coupling is what is structurally new.
The reasoning hierarchy
The first recursion operates on artifacts.
When agents reason, they produce structured content. Decisions, justifications, tool-use sequences, retrieval steps, self-corrections, reasoning traces. In well-instrumented deployments, these artifacts are logged. In some, they are indexed. In a few, they are organized into hierarchies that future agents can consume.
The hierarchy has natural levels. Individual reasoning traces sit at the leaf level. Recurring reasoning forms across many traces become patterns. Patterns that recur across enough deployments and contexts crystallize into abstractions: named reasoning genres, standardized templates, citable forms of justification. Abstractions that survive long enough and acquire enough use accumulate something like institutional standing. They become the implicit reference points for future reasoning, citable not just by name but by the kind of warrant they carry.
This is corpus formation where retention, attribution, and addressable abstraction are in place. Where they are not, the same artifacts accumulate as private logs, transient traces, or dead records. The corpus is not automatic. It requires infrastructure that selects what becomes pattern and what stays at trace level.
The defining feature of this recursion is that it is generative. Each level produces inputs for the next. Leaf reasoning produces patterns. Patterns produce abstractions. Abstractions, once stable, produce new reasoning that conforms to or builds on them. The hierarchy creates the conditions for its own continuation.
Considered alone, the corpus looks like a cultural inheritance channel: artifacts transmitted, selected, and built upon at frequencies human institutions cannot match.
The cognition stack
The second recursion is more diffuse than the first. It does not operate through a single mechanism. It operates through several pathways that update different parts of the cognition stack at different speeds.
Three pathways carry most of the current weight.
Parameter inheritance. Some outputs become training data or fine-tuning targets. Models trained on this data acquire dispositions that shape what they produce. The cycle frequency is the slowest of the three: full retraining of foundation models still happens on cadences measured in months; fine-tuning runs faster, sometimes weekly. This is the pathway that most directly modifies the substrate, in the narrow sense the modeling-layer argument used.
Selection-mediated inheritance. Some outputs affect evaluations, benchmarks, leaderboards, red-team tests, deployment metrics. The result is not a change to any individual model’s weights, but a change to which models survive commercial selection. Models that produce outputs the developer ecosystem treats as exemplary get reinforced through deployment. Models that fail certain selection criteria do not. Selection-mediated inheritance operates at engineering frequency, often days to weeks, and propagates across the ecosystem rather than within a single model lineage.
Contextual inheritance. Some outputs become prompts, examples, retrieval material, memory, or templates that shape how future models behave at inference time. No weights change, no models get selected, but cognition is conditioned by context that previous cognition produced. This pathway operates at the fastest cycle, sometimes hours, and operates within deployments rather than across them.
These three pathways have different properties. Parameter inheritance is durable but slow. Selection-mediated inheritance is fast and ecosystem-wide. Contextual inheritance is fastest but most local. They also differ in what they transmit. Parameter and contextual inheritance carry content directly: outputs become the data or context that conditions future behavior. Selection-mediated inheritance carries content indirectly: it changes which systems survive, and the surviving systems carry their content forward. They share one feature: each is a route by which current machine cognition shapes future machine cognition without requiring human cognition to be the primary transmission substrate.
The empirical literature on what happens when each pathway recurses on agent-produced content is in early stages but converging on findings worth naming. Shumailov and colleagues’ work on model collapse shows that recursive parameter inheritance, without sufficient anchoring to original sources, produces measurable degradation in model quality across generations. Adjacent work on self-consuming generative models reaches a similar conclusion: without enough fresh real data, recursive training degrades quality or diversity. Self-rewarding language model work demonstrates the converse: with appropriate fitness criteria, recursive training can improve a model along specific axes. Recent work on preference leakage shows a related dynamic on the selection side: LLM-based judges favor outputs from related model lineages, which means selection pressure can become endogenous to model family rather than tracking external quality. These are special cases of single-pathway recursion operating in isolation. The dual-recursion frame asks what happens when these pathways are coupled to a corpus that has its own selection dynamics, and to each other operating at different speeds.
The empirical conditional matters. The dual recursion is partially active and accelerating. Synthetic data pipelines, distillation, fine-tuning on production logs, evaluator training on agent traces are all examples of agent-produced outputs feeding back into the cognition stack. But agent-produced content is not yet the dominant share of frontier model training. The dynamics described here become structurally consequential as that share grows. The conditional is empirical: what fraction of cognition-stack updates currently come from agent-produced artifacts, and what selection pressures propagate through that fraction.
The coupling
Each recursion operates on its own. The interesting dynamics arise from how they interact.
The unit that travels between channels is the reasoning form: a repeatable structure for justifying decisions, like the procurement template from the opening. A reasoning form can appear as a trace, harden into a template, be named as an abstraction, be rewarded by an evaluator, enter a prompt library, or be learned by a model. The dual recursion operates on reasoning forms.
The artifacts the first recursion accumulates are produced by the cognition the second recursion is shaping. The cognition the second recursion is shaping is trained, evaluated, or conditioned, in part, on the artifacts the first recursion has accumulated. Each recursion is the substrate the other operates on.
Consider what happens when the procurement justification template from earlier becomes prevalent in the corpus. Future agents consume the corpus and produce more reasoning of that form, both because the form is what they have learned to recognize and produce, and because the form has institutional standing that makes it easier to defend as the right form. The form propagates through the first recursion.
The form also propagates through the second recursion. The outputs of agents producing it enter training data through parameter inheritance. The form gets rewarded by evaluators through selection inheritance. The form ends up in prompt libraries through contextual inheritance. Models acquire the form as a disposition. Selection systems acquire the form as an expected output. Contextual layers acquire the form as a default.
The result is that the form gets reinforced through both channels simultaneously, across all three pathways of the second recursion. Forms that satisfy selection pressures in both the corpus and the cognition stack propagate. Forms that succeed in one but not the other get marginalized.
The coupling is mediated by specific selection operators. The system is not selecting reasoning forms by magic. Storage policies, retrieval systems, abstraction tools, evaluators, reward models, benchmark designers, dataset curators, fine-tuning pipelines, deployment gates, and marketplace ranking systems are doing the selection. Each operator selects according to its own criteria. The coupling becomes consequential when many operators select for the same forms simultaneously.
Where the operators align, forms become locked in. Where they conflict, forms create productive tension between the two channels. The structural argument is strongest where selection operators concentrate. Where the same organizations control logs, evaluators, fine-tuning pipelines, deployment gates, and marketplace ranking, the operators tend to align. Vertically integrated ecosystems produce tighter dual recursion. Fragmented ecosystems produce more tension between channels. The trajectory of operator concentration is itself something the dual recursion will help determine.
What this means structurally is that agent populations have something analogous to dual inheritance, with both channels operating at engineering frequencies and with explicit selection mechanisms. The two channels do not run at identical speeds. They run close enough that feedback from one can shape the other inside commercial planning horizons. Forms that succeed in both channels become locked in quickly. Forms that fail in either get pruned quickly. The system converges on its attractors faster than any human institutional system has, and it does so without the slow stabilizing influence that long generational time scales provide.
Dual inheritance at agent scale
Boyd and Richerson developed dual inheritance theory to explain how human cognition and culture coevolved. The argument was that human populations have two inheritance channels operating in parallel: genetic and cultural. Each channel transmits information across generations. The two channels shape each other. Cultural innovations create selection pressures on genetics, and genetic capacities enable cultural innovations.
The dual recursion in agent populations is structurally analogous, with two important differences.
First, the channels operate at engineering frequencies. Not at identical speeds. Corpus evolution can run hourly. Selection-mediated inheritance runs in days. Parameter inheritance runs in weeks to months. Foundation model retraining runs in months to a year. None of these matches the generational time scales that shape biological-cultural coevolution. The dual recursion runs entirely inside commercial planning horizons, which is the relevant comparison.
Second, the channels are explicit. Human cultural inheritance happens through opaque processes: imitation, language, tradition, with mechanisms still partly contested in the literature. Human genetic inheritance has known mechanism but is not directly steerable. In agent populations, both channels have explicit mechanism. The corpus is stored, indexed, retrieved through systems we built. The cognition stack is trained, evaluated, deployed, conditioned through pipelines we built. Both can in principle be steered.
The closest modern analogue is not biological evolution. It is the social-media recommender loop: human-produced content shapes ranking algorithms, ranking algorithms shape what content humans produce. The recommender loop is a partial precedent. The agent case is different because both sides of the loop are increasingly machine cognition. Artifacts produced by agents shape the systems that train, select, and condition future agents. Recommender loops have one cognitive side and one human side. Dual recursion has cognition on both sides.
The novelty is not feedback. Commercial systems have had feedback loops for decades. The novelty is feedback between machine reason-bearing artifacts and the machine cognition stack that produces the next round of those artifacts, with both channels increasingly explicit, instrumented, and operating inside commercial planning horizons. What appears new is the combination: two explicit, steerable, fast inheritance channels with machine cognition increasingly on both sides of the loop.
The Compilation Thesis describes how judgment compiles into infrastructure as bottlenecks shift up the stack. Substrate-knowability identified the next bottleneck: the modeling layer, sitting between Judgment and Direction in the stack. Dual recursion identifies what comes after. This is not a new layer in the stack so much as a structural feature of the existing stack: the coupling between the corpus and the cognition layers determines whether the stack converges on better cognition over time or on attractors that are bad in both. The interesting question is no longer how to produce good reasoning artifacts, or how to train better cognitive substrates, but how to manage the coupling between the two channels.
Four dynamics the coupling produces
The coupled recursions produce dynamics that single-channel analysis misses.
Convergence on multi-channel attractors. Reasoning forms that satisfy selection pressures in both channels propagate. Forms that excel in one but not the other get pruned. The population converges quickly on forms that are selected by both the corpus and the cognition stack. These forms have a particular shape. They tend to be standardized, easy to reproduce, easy to evaluate, easy to learn, easy to defend. They favor conventional reasoning over original insight, because conventional reasoning has the form that propagates through both channels. Recent work on AI-to-AI communication chains documents what this looks like empirically: convergence on narrative anchors, loss of evidentiary texture, attrition of hedges and attributions. The risk is monoculture at the level of reasoning itself, and the long-run consequence is that monoculture in the corpus produces monoculture in the cognition stack, which produces monoculture in the next round of corpus artifacts. The loop closes on itself.
Lock-in once density crosses a threshold. Once a form reaches enough density in either channel, it becomes self-reinforcing through the other. A form prevalent in the corpus shapes the cognition-stack updates that produce more of that form. A form embedded in the cognition stack produces artifacts that reinforce its presence in the corpus. Past a certain density, the form becomes very hard to displace, even if better forms exist. The relevant density is not raw count. It is share of decision volume in strategically relevant transactions, share of training data in relevant fine-tuning runs, share of evaluator outputs in relevant deployment selections. Density that crosses the threshold in a central or widely reused operator can propagate across the system through the coupling. Density inside a private deployment produces local lock-in first.
Cascade through the coupling. Small changes in one recursion can produce large effects in the other. A single influential abstraction added to the corpus shapes the training data for the next round, which shapes the cognition that produces the corpus’s next round of artifacts. A subtle change in evaluator behavior shifts what kinds of artifacts get rewarded, which shifts what gets preserved in the corpus. Because the two recursions feed each other, perturbations propagate. For example: a new compliance abstraction enters a procurement corpus, evaluators begin rewarding it, fine-tunes imitate it, agents produce it more often, counterparties accept it faster, and alternative justification forms disappear before anyone formally bans them. The system has no clean separation of timescales that would let one recursion absorb shocks from the other.
External-grounding decay. Forms can become selected by both channels while losing contact with the external conditions they were supposed to track. A justification template can become easy to evaluate, easy to train on, and easy to cite while becoming less faithful to the outcomes it claims to describe. The coupled system then improves at reproducing the form rather than at predicting or shaping the world. This is the corpus-level analogue of model collapse: not collapse of output quality in general, but decay of evidentiary contact between the forms and the world they describe. The fourth dynamic is what makes the first three consequential: convergence and lock-in matter because they can lock the system onto forms that no longer track anything outside the system itself.
Each of these dynamics has analogues in single-channel systems. Most published work treats the two recursions separately. The work that recognizes them as a coupled inheritance system is in its earliest stages.
What is forming
Two recursions, coupled, running at engineering frequency, on cognitive content, with explicit mechanism and explicit selection operators, at population scale. This combination has not existed before. The closest analogues are partial. Cultural evolution operates on artifacts but not on the substrate that produces them. Biological evolution operates on substrate but at generational frequencies. Recommender loops operate on artifacts and selection systems but with humans on one side. Dual recursion has machine cognition on both sides of the loop.
The dual recursion is not yet visible as a unified phenomenon. It is happening in pieces, in specific deployments, in particular ecosystems, in different markets at different rates. The pieces are real. The unified picture is not yet available because nobody has needed to see it as unified before.
The participants who learn to read both recursions and their coupling will be operating with a frame that almost nobody else has. The participants who do not will be modeling a system whose dynamics they cannot explain because the explanations require seeing both channels at once.
The corpus does not just accumulate. It shapes the cognition that produces the next corpus.
Concepts Ghost GDP · Compilation thesis