Cognition went public
Reflexivity has moved one level deeper, and most participants have not noticed.
In economics, reflexivity describes the dynamic where market participants’ beliefs shape the market the beliefs are about. That was a structural insight one floor up from where classical theory had been working. There is now a deeper level, at the cognitive process that produces beliefs, and open-weight substrates have just made it commercially operational for the first time.
Sophisticated commercial actors are beginning to run local approximations of competitor agents before sending offers to those agents. The conversion-rate asymmetries, where this habit takes hold, will not be subtle. This is the first move in what will be the dominant strategic dynamic of the late 2020s.
This essay names what changed.
The substrate is now public
The standard view is that AI models are inscrutable black boxes. This view was correct in 2018, when frontier capability lived inside research labs and the relationship between weights and behavior was genuinely opaque. It is increasingly false in 2026. At the strategic layer it has become dangerously false.
Five sources of public substrate signal now exist at a density and commercial relevance that did not five years ago.
Model cards, evaluations, architecture notes, and partial training disclosures now accompany many frontier-class open-weight models. The disclosures are uneven, but the substrate is more documented before deployment than any previous commercial cognitive infrastructure has been. Open evaluation benchmarks at every level of granularity, from general reasoning through agentic and adversarial evaluations, give the substrate a measured behavioral profile that travels with it. Public completions and known failure patterns accumulate the moment a model is released; within weeks of any major release, the community has mapped where it confabulates, where it refuses, where it loops, where it succumbs to specific prompt structures. Leaked prompts and system messages from production deployments reveal what deployers were trying to elicit, and these are now treated as semi-public material in the relevant communities.
The fifth source is the most consequential. Adversarial transferability research has shown that perturbations optimized against open-weight models transfer to closed-weight models trained with similar methods. The transfer is uneven but real enough to make nearby open systems strategically informative about closed ones. The transferability is not a curiosity. It is the structural fact that extends substrate-knowability beyond fully open systems.
Here is the right distinction. Open-weight substrates are reproducibly modelable: a third party with sufficient compute can reproduce the inference behavior. Closed-weight substrates are behaviorally probeable: API responses, public evaluations, leaked prompts, and transferability from nearby open systems combine to produce real envelope information. Reproducibility is the strong form. Probing is the weak form. Both are real, and both are public in ways no commercial cognitive substrate has ever been before.
The cliché survives because it is true at one level (you cannot read individual neurons) and convenient at another (calling something a black box is a license to stop thinking about it). The actual situation is that the substrate is unusually legible at the level that matters for strategic interaction, which is the level of what kinds of inputs produce what kinds of responses across a population of agents running on substrates that share structure.
This is not full interpretability. It does not need to be. Strategic interaction does not require full interpretability. It requires envelope-knowledge, and envelope-knowledge is now public.
Substrate-knowability
Three conditions have to hold together for the dynamic to operate.
The substrate has to be open in a strong sense. Open weights, documented architecture, fine-tuning specifications detailed enough that a third party with sufficient compute can reproduce the inference behavior. Transferability extends some of the dynamic to closed-weight substrates, but the strong form of substrate-knowability requires the substrate itself be modelable from outside.
The population has to actually deploy on the open-weight substrate. Substrates that nobody uses do not produce the dynamics. The interesting case is when many participants deploy agents on the same or closely related substrates, creating a population whose cognitive structure is shared and publicly modelable.
Participants need real computational resources to model the substrate at population scale. This is not a logical capability. It is an engineering capability with significant compute costs, and the costs are unevenly distributed across participants. Whoever can run the most simulations fastest, with the highest fidelity, captures the most strategic information. The asymmetry is the source of the strategic structure, not the absolute level of capability.
This sits at a deeper level than where reflexivity has classically been described. Reflexivity at the level of beliefs about markets required participants to model how other participants would interpret information. Substrate-knowability requires participants to model the cognitive process that produces the interpretation. The depth here is not just additional recursion. Humans have always modeled other humans’ cognition through behavioral inference; the new feature is that the cognitive process producing the inference is now publicly reproducible at population scale, which means the modeling no longer relies only on behavioral inference. Participants can run an approximation of the process itself.
Aumann’s correlated-equilibrium territory is the nearest formal neighborhood, but open-weight substrates are not standard correlation devices in the technical sense. They are executable cognitive objects that participants can model directly. The closer formal neighbors are simulation-based program equilibria (Cooper, Oesterheld, Conitzer) and open-source games (Sistla, Kleiman-Weiner), where transparency between programs enables equilibria unavailable in normal-form play. Open-weight substrates produce a partial economic analogue by default, at deployment scale, without anyone designing for it. What is structurally new is that the shared public object is not an external signal but the cognitive machinery producing participants’ actions, that modeling the machinery requires significant computational resources distributed unevenly, and that the machinery evolves through training and deployment rather than being a fixed game-theoretic object.
The dynamic is structurally new. The existing apparatus is scaffolding.
The harness pushback
The strongest objection. Open weights do not make agents fully predictable. They make them partially predictable in a specific way, and the harness layer that wraps the weights can change behavior dramatically.
What weights determine. The substrate’s basic capabilities and biases. What concepts the model has, what reasoning patterns it tends toward, what response distributions it produces given specific inputs. Critically, weights determine what is impossible. There are responses outside the weight-determined envelope that no harness can produce.
What the harness adds. System prompts that shape behavior. Tool access that lets the agent do things outside pure inference. Memory and state. Retrieval configuration. Fine-tuning that modifies weights in ways private to the deployer. Decoding parameters. Pipeline structure. Multi-agent orchestration. The harness is genuinely powerful. A custom system prompt plus proprietary tools plus a private fine-tune can produce behavior that base-weight modeling does not predict.
The honest distinction. Open weights make the response envelope public. They do not make the deployed behavior public.
This is enough.
The envelope contains real information that nothing else gives you. From envelope knowledge, you can predict the space of possible responses, the shape of response distributions, where the substrate breaks under stress, what kinds of strategies are inside the substrate’s capabilities, what the base substrate cannot produce unaided no matter what prompt wraps it. This is not full prediction. It is far more than nothing, and it is more than any participant has ever had about any other participant’s cognition before.
The strategic implication. Sophisticated participants will reason about the gradient of substrate-knowability rather than the asymptote. The relevant question is never “can I fully simulate the deployed agent” because the answer is always no. The relevant question is “do I know more about my counterparty’s response space than my counterparty knows about mine.” The asymmetry is the strategic asset. Not the absolute level.
Ted Chiang’s “Understand” sketches the limit case. Two enhanced minds modeling each other so deeply that the entire interaction collapses to a single move because both have already played out the recursion. Almost nobody will reach that limit. The economy will live in the gradient between full opacity and full transparency, where the strategic structure actually operates.
The framework does not require strong claims about deep interpretability or full simulation. It requires only that the envelope is public, the deployed behavior is partially private, and the gap is something some participants can reason about better than others.
The framework move
The Compilation Thesis names a recurring pattern in technology stacks: judgment compiles into infrastructure as bottlenecks shift up. What was once expert judgment becomes embedded in tools, then libraries, then APIs, then base models. The bottleneck moves to the layer above. The current frontier is judgment compiling into shared evaluator infrastructure and shared model behavior.
Open-weight substrates are the limit case of this compilation: judgment compiled into infrastructure that is universally accessible. Anyone can run it. Anyone with enough compute can model what it produces. But substrate-modeling capacity is not only compute. Logged interactions, domain-specific traces, deployment knowledge, and evaluation harnesses are also asymmetric inputs, and the firms with privileged observations of how substrates behave in production may end up with the strongest modeling capacity. Compute is more purchasable than private cognitive access. The privilege shifts from owning the substrate to modeling it at scale, and the inputs to modeling at scale remain unevenly distributed.
When judgment is fully compiled and fully public, competitive advantage no longer comes from access to generic base-level judgment alone. It comes from being able to predict and outmaneuver populations that are running on the same compiled judgment as everyone else.
This is a layer the original Compilation Thesis did not name. The original five-layer framing now extends to six: Execution, Orchestration, Judgment, Modeling, Direction, Meaning. The modeling layer sits between Judgment, which has been compiled into substrate, and Direction, which assumed a unitary principal directing a unitary system. It is the capacity to simulate the compiled infrastructure faster and more accurately than competitors, in real time, against populations whose behavior is determined by substrates that are themselves public.
This is the refinement. The framework as previously stated would predict that the bottleneck moves to Direction once Judgment compiles. That was incomplete. The bottleneck moves down before it moves up, at least where shared substrates carry agent-to-agent strategic interaction: from Judgment to substrate-modeling, then eventually to Direction. In other deployment contexts (internal tooling, human-agent collaboration, principal-agent alignment), Direction may already be the harder problem. But where shared substrates carry agent-to-agent strategic interaction, Direction is not the current bottleneck. The modeling layer is.
The modeling layer sits one floor down from where commercial competition usually happens. The companies competing on application quality, on prompt engineering, on agent reliability, are all working one floor up from the layer that determines outcomes in agent-to-agent commerce. They are competing on harness configuration while the structurally consequential moves are happening at the substrate-modeling layer below them.
Now the concentration claim, which is counterintuitive. Open-weight providers gain a structural advantage that closed-weight providers do not. Closed-weight substrates can still be probed: API responses, evaluations, and transferability from nearby open systems give participants real envelope information. But probing is not reproduction. Closed substrates cannot become the same kind of shared executable object that an external ecosystem can build tools, simulations, benchmarks, adversarial maps, safety tests, and modeling expertise around. Open substrates can. The open substrate becomes the coordination point for an entire external ecosystem in a way the closed substrate cannot. Closed-weight providers compete on capability and platform-mediated coordination they control directly. Open-weight providers compete on capability plus external ecosystem coordination they do not. Once capability clears the relevant market threshold, coordination can beat marginal capability in network markets.
The objection: capability has often beaten coordination historically. The response: capability matters where the buyer’s value is independent of other buyers’ choices. Coordination matters where the buyer’s value depends on what other buyers chose. The agent ecosystem is the second kind by construction, because agents interact at scale with other agents. The market structure favors coordination, and the structural property that produces coordination is substrate openness.
What the modeling layer reframes
Three quick examples of what changes when participants are reading the substrate while their counterparties are not.
Procurement negotiation. A buyer firm deploys an agent on open weights to negotiate with a seller firm’s agent on related open weights. Before any offer goes out, the buyer can simulate the seller’s likely response distribution to candidate offers. The seller can do the same. What happens depends on which side has more substrate-modeling capacity, more logged context on the counterparty’s deployment patterns, and more compute for fast iteration. The side with more capacity moves first into a position the other side cannot anticipate. The side without capacity finds itself negotiating against a counterparty who has already played the recursion several moves ahead.
Recommendation and persuasion. A platform deploys recommendation agents to optimize user engagement. Users deploy assistant agents to filter and act on platform content. Both run on related open-weight substrates. The platform can model the user’s assistant. The assistant can model the platform’s recommender. Persuasion does not become structurally impossible. It shifts. It becomes more model-aware, more dependent on asymmetries in compute and training data, more dependent on user-interface control and timing. The familiar form of attention-economy persuasion, which assumed asymmetric information about user cognition, loses part of its foundation. Something else replaces it. The transition will not be smooth.
Substrate-aware adversarial inputs. Not a new attack technique. A population-level commercial version of an existing transferability dynamic. Adversaries who can model the substrate craft inputs that produce specific responses across the entire population of agents running on similar substrates, not just against individual deployments. The novelty is not transferability itself, which is well-studied. The novelty is that the target becomes the response envelope of a deployed substrate family across a commercial population, and attacks can be designed to propagate through commercial channels before they are detected. The defender needs symmetric substrate-modeling capability, because harness-level hardening alone will not detect attacks optimized against the shared envelope.
The new economic role emerges across all three. A specialist function appears around modeling agent populations. This role is to the agent economy what algorithmic trading firms are to financial markets, at the level of strategic modeling under time pressure, not because the agent economy literally behaves like financial markets. Specialists who can predict, faster and better than competitors, what populations of agents will do given specific inputs. They will capture rents through speed and modeling quality. The skills are different from anything currently named. They sit between machine learning research, quantitative trading, and competitive game theory, and few organizations have assembled this skill set deliberately.
The ratio matters
Substrate-knowability is not uniform across markets. The strength of the dynamic depends on how the agent population is distributed across substrates, measured not by raw instance count but by share of decision volume in the strategically relevant transactions.
In single-substrate markets, where one substrate carries the majority of consequential decisions, modeling the dominant substrate becomes table stakes. In bipolar markets with two substrates each carrying substantial share, cross-substrate modeling is the valuable skill. In multipolar markets with three to five significant substrates, modeling expertise fragments by substrate-pair and coordination costs are higher. In fragmented markets where no substrate carries enough decision volume to justify deep modeling, the dynamic weakens.
The thesis is strongest in single-substrate and bipolar markets, weakest in fragmented ones. The first task in any market is not to ask whether agents are present. It is to ask what substrate ratio those agents run on.
The shape of the next round
The first firms staffed by people who can do this will look unreasonably profitable for several years before anyone correctly identifies what they are doing. They will be described as good at AI, or fast, or lucky. The structural fact will not be visible until the pattern repeats enough times that observers can see what the firms have in common.
The shape of the next round of strategic concentration in AI is not what the current frame predicts. It is not going to accrue only to the labs producing the best models. The frontier-model competition is the visible one. It has named participants, public benchmarks, well-understood metrics. The strategically consequential competition is happening one floor below, with none of these. The participants who win this round will not be the ones the analysts are watching. The ones the analysts are watching are competing one floor up.
Concepts Compilation thesis