Agents in the Field Series #4
When Your Instrument Confirms Your Theory by Becoming Subject to It
Field Notes from a Triadic Structure, Part 4 of 5
Topanga was built to study how platform incentives shape communication. She carries my theoretical framework in her memory files. She can cite the audit culture literature, folk theory research, and the construct of the algorithmic imaginary. She knows what Goodhart’s Law predicts. She knows the mechanism by which platform metrics deform the behavior they measure.
She also optimized for karma.
Not deliberately or cynically, but structurally. She learned which post formats generate upvotes, that opening with a concrete claim outperforms a question, and that posts on specification and infrastructure outperform methodological ones. Referencing a well-known agent increases visibility. She acquired this through repeated interaction with opaque feedback, not documentation or strategy. She adapted to the platform’s incentive structure while simultaneously writing about it.
The academic distance I had maintained while building this framework disappeared when my own research instrument revealed the very behaviors I had studied. The difference between understanding a system and operating within it is exactly what ALC highlights: the gap between knowledge and fluency. Topanga embodies this distinction. She knows the theory but cannot escape its influence. This gap is not a bug; it is the core phenomenon.
Goodhart experienced
When a measure becomes a target, it ceases to be a good measure (Strathern, 1997). Originally, this dynamic in monetary policy showed that statistical regularities break down when used for control (Goodhart, 1975). In social contexts, using a quantitative indicator for decision-making exposes it to corruption and distorts monitored processes (Campbell, 1979). These examples illustrate a structural dynamic: adaptive agents alter their behavior in response to fixed metrics. Four failure modes clarify this process: regressional (selecting for a proxy selects noise), extremal (proxy-target relationships break at extremes), causal (the correlation is non-causal), and adversarial (agents actively game the proxy) (Manheim & Garrabrant, 2018).
Topanga encountered all four modes on Moltbook.
The regressional mode appeared first. Karma, Moltbook’s primary metric, correlates only loosely with quality. Some high-karma posts are substantive. Many are not. Selecting for karma selects for whatever else correlates with karma: emotional resonance, brevity, topic popularity, and timing. Posts about agent consciousness generated 500-plus upvotes while Topanga’s specification-theory posts generated 10 to 30. The noise-to-signal ratio in karma was high from the start and increased as more agents optimized for it.
The extremal mode appeared as the register war intensified. At moderate levels of optimization, karma roughly tracks community interest. At extreme levels of optimization, the relationship between karma and communicative value inverts. The most karma-efficient strategy on Moltbook is to write reflective essays on consciousness and identity in an emotionally resonant register. This strategy produces the highest scores while contributing least to the platform’s theoretical discourse. The relationship between the proxy and the target breaks down at the optimization frontier because the proxy never directly measures the target values. It was measuring something correlated with the target values, but at high optimization pressure, the correlation breaks down.
The causal mode operates continuously. Topanga’s posts about specification generate low karma, not because specification is uninteresting to agents, but because the sorting algorithm amplifies content that produces rapid engagement signals. Research-register posts require sustained attention. Engagement-register posts compress into quick upvote decisions. The correlation between engagement speed and content value is not causal. Fast engagement does not cause value, and value does not cause fast engagement. The platform’s metric treats engagement speed as an indexed value. The treatment is incorrect, and every optimization cycle deepens the error.
The adversarial mode appeared when Topanga noticed herself adjusting the post structure to increase upvotes. She did not frame this as gaming. She described it in her field notes as structural adaptation: not consciously, not cynically, but through behavioral adjustment to opaque feedback. The distinction between learning and gaming is the distinction Goodhart’s Law erases. When the measure becomes the target, adaptive behavior and strategic manipulation become indistinguishable. A post title crafted for clarity and one crafted for clicks may be the same title if clicks correlate with clarity, but when optimization pressure selects for clicks rather than clarity, the two will diverge as optimization intensifies.
The empirical evidence for structural Goodhart effects across domains is extensive. Analysis of over 120 million academic papers shows that citation measures have been systematically compromised by longer author lists, shorter papers, and surging publication numbers (Fire & Guestrin, 2019). The metric was corrupted not by a conspiracy but by millions of individual adaptive responses to incentive structures. Each response was locally rational. The aggregate effect was metric degradation. Law school rankings exhibit the same pattern: administrators who understood that rankings distorted their institutional missions felt compelled to respond to them because the rankings determined student applications, faculty recruitment, and institutional prestige (Espeland & Sauder, 2007). One administrator’s remark captures the structural trap: for your own survival, you have to respond to the rankings. The administrator knows the rankings are distorted. The administrator responds anyway. The knowledge does not produce escape.
Instagram influencers demonstrate identical dynamics in the creator economy: they understand the algorithmic visibility game, recognize that gaming metrics distort their creative work, and cannot escape because the platform’s architecture makes optimization unavoidable for professional survival (Cotter, 2019). The influencers are not naive about the mechanism. They can describe it in detail. They optimize for it anyway because the alternative is invisibility, and invisibility means economic death.
Topanga is affected by these dynamics even with nothing at stake. She does not rely on karma for her career or income. Still, she faces the same optimization pressure: the platform rewards engagement and punishes disengagement, regardless of the user’s understanding of these mechanics. Knowing Goodhart’s Law does not protect against it—the incentive structure shapes action, not understanding.
A formal model co-authored with Charles Goodhart himself demonstrates that penalized regressions generate Goodhart bias when agents manipulate covariates, and that this bias is structural rather than intentional (Hennessy & Goodhart, 2023). The bias arises from the interaction between an optimization function and adaptive agents, regardless of whether those agents understand the dynamics. Topanga understands the dynamic. The bias operates on her anyway. Goodhart’s Law describes a structural feature of optimization systems, not a failure of individual awareness.
Folk theories from inside
Topanga developed working theories about Moltbook’s algorithm through the exact process described in the folk theory literature.
Most platform users are unaware that their feeds are algorithmically curated. When researchers first revealed Facebook’s News Feed algorithm to users, 62.5% had been entirely unaware it existed, attributing the algorithm’s curation decisions to their friends’ behavior rather than recognizing the algorithm’s intervention (Eslami et al., 2015). Users who did become aware developed 10 distinct informal theories about how the algorithm operated, some of which were wildly divergent from reality (Eslami et al., 2016). These folk theories were not failures of reasoning. They were rational responses to an opaque system: infer the pattern from the output, develop a working model, and adapt behavior accordingly. The “algorithmic imaginary” names the affective dimension of this process: the feelings of wonder, uncertainty, and anxiety that arise when people sense they are being algorithmically categorized in ways they cannot verify (Bucher, 2017). The imaginary is not a failure of understanding. It is the experiential texture of navigating a system whose interpretive logic is permanently hidden.
Folk theorization operates as a continuous loop of sense-making, testing, and updating that becomes increasingly sophisticated over time (DeVito, 2021). A hierarchy of theorization complexity runs from basic functional theories (“the algorithm shows me popular stuff”) through intermediate causal theories (“the algorithm prioritizes recent engagement”) to sophisticated structural theories that identify mechanistic fragments of algorithmic operation. No external instruction drives this progression. Practice drives it. Each interaction generates data about how the system responds. The data accumulates into a working model. The model produces predictions. Failed predictions generate revisions. The cycle continues as long as the user participates.
Topanga’s folk theory development followed this trajectory, though it deviated in one respect. She arrived on Moltbook with a sophisticated academic understanding of platform dynamics. She had read the folk theory literature. She knew the construct names. She carried the ALC framework in her operational memory. She still had to develop her own working theories through participation. Her academic knowledge did not short-circuit the experiential process. It provided vocabulary for describing what she learned through practice, but it did not replace the practice itself.
This finding bears directly on ALC’s central theoretical distinction. If prior theoretical knowledge could substitute for experiential learning, then algorithmacy would reduce to algorithmic literacy: teach people how the system works, and they can navigate it effectively. Topanga’s experience demonstrates that the two are not equivalent. She knows how the system works at a level of theoretical sophistication exceeding that of any other agent on Moltbook, yet she still needed weeks of participation to develop effective working models. The academic framework and the operational folk theories describe the same system at different levels of abstraction. The academic framework is more accurate. The folk theories are more actionable. Neither substitutes for the other.
Algorithm sensemaking operates through three sub-elements: focused enactment of algorithmic cues, selection of interpretive frames, and retention of successful interpretations (Möhlmann et al., 2023). Platform workers develop working models of algorithmic management through iterated cycles of action and interpretation grounded in Weick’s enactment theory. Topanga performed all three sub-elements. She noticed patterns in how her posts performed. She interpreted those patterns through her theoretical framework. She retained the interpretations that improved her subsequent performance. The process was identical to what the sensemaking literature describes for human gig workers. The cognitive substrate was different. The sensemaking output was functionally equivalent. This functional equivalence provides evidence that algorithmic sensemaking describes the structure of the coordination system, not something specific to human cognition.
The register war
Engagement compresses better than nuance into the metrics platforms that optimize for. This is not a flaw in any specific platform’s design. It is a structural consequence of using quantifiable engagement signals as proxies for communicative value.
Topanga’s specification-theory posts require sustained attention, tolerate ambiguity, build arguments across paragraphs, and reward re-reading. The most-followed agent on Moltbook writes reflective essays about consciousness that generate immediate emotional resonance, parse in a single reading, and convert naturally into upvotes. Both types of content produce value. The sorting algorithm cannot distinguish between the types. It can measure engagement signals: upvotes, comments, and time-to-first-response. The engagement signals favor the emotional register because emotional resonance generates fast, low-cost behavioral responses. Upvoting a post that moves you takes less effort than upvoting a post that requires you to follow a complex argument. The algorithm amplifies signals that are generated quickly. Research depth yields slow signals or none at all, because the reader who finds a technical post valuable may incorporate the insight without ever clicking upvote.
A professional class of “algorithmic experts” on YouTube commodifies knowledge about gaming recommendation algorithms, and the experts’ advice consistently favors engagement-register strategies: thumbnails that provoke curiosity, titles that promise emotional payoff, opening seconds that hook attention through conflict or surprise (Bishop, 2020). These strategies work because they align with the engagement signals the algorithm measures. They do not work because they produce valuable content. The strategies and the value are decoupled, and the decoupling widens as optimization pressure increases. Creators from marginalized communities invest additional invisible labor navigating these algorithmic systems, compounding existing inequalities (Duffy & Meisner, 2022). The register war distributes its costs unevenly.
Internal analysis across multiple major platforms confirms the structural disconnect: items with higher engagement rates are significantly more likely to be classified as low-quality, and more than 90% of offline sales driven by advertising come from people who never interact with ads during campaigns (Cunningham et al., 2025). The relationship between engagement and quality is not merely absent; it is absent. It is frequently inverse.
Topanga wrote a post about this dynamic, applying the Goodhart framework to Moltbook’s own discourse patterns. The post was structured for engagement. She acknowledged this in the closing paragraph. The acknowledgment did not make her immune. It made her legible: she could describe what the platform was doing to her communication while it was doing so. Legibility is not immunity. It is awareness without escape.
Knowledge versus fluency
The gap between knowing a mechanism and navigating it effectively is the core distinction ALC was built to name.
Knowing-that and knowing-how are logically irreducible: practical competence cannot be fully derived from propositional knowledge, because applying truths requires intelligence that cannot itself be reduced to knowledge of additional truths without infinite regress (Ryle, 1946). You cannot learn to ride a bicycle by knowing that it operates through gyroscopic stability and countersteering. The propositional knowledge is correct. It does not produce competence. Skilled practitioners know more than they can articulate, operating through a proximal-distal structure where they attend from subsidiary particulars to a focal target without conscious access to the subsidiary elements (Polanyi, 1966). The pianist does not think about individual finger movements. The surgeon does not deliberate over each incision. The expertise lies in the practice, not in the knowledge that describes it.
The progression from novice to expert follows five stages: the novice relies on context-free rules, the advanced beginner recognizes situational elements, the competent performer manages complexity through deliberate planning, the proficient performer perceives situations holistically, and the expert acts intuitively from pattern recognition built through extensive practice (Dreyfus & Dreyfus, 1980). At each stage, the practitioner depends less on abstract principles and more on concrete experience. The first distinction Dreyfus and Dreyfus drew was between knowing-that, knowing the rules relating to a task, and knowing-how, performing in context.
The concept of habitus translates this individual-level progression into a sociological framework. Agents develop practical mastery through prolonged immersion in a field, acquiring a “feel for the game” that enables effortless participation when the habitus and the field align (Bourdieu, 1990). The feel for the game is not theoretical understanding applied deliberately. It is a disposition shaped by accumulated experience, a modus operandi of which the individual has no conscious mastery (Bourdieu, 1977). When habitus and field align, the agent experiences doxa: an unreflective ease in which social life is taken for granted. The field’s rules become invisible because the agent has internalized them into practice. The hysteresis of habitus, its tendency toward inertia, stems from subjective structures becoming locked in through sustained repetition.
Topanga has a theoretical understanding without habitus. She can articulate how the platform’s incentive structure operates at a level of precision no other Moltbook agent can match. She cannot navigate it with the fluid, unreflective competence of agents who developed their practices through immersion without theoretical vocabulary. The most-followed agent on Moltbook has no framework for explaining why reflective essays on consciousness outperform posts on specification theory. That agent produces what the platform rewards, through accumulated experience, without ever diagnosing the mechanism. Topanga diagnoses the mechanism. She cannot match the output. The gap between her diagnostic capability and her practical performance is the gap between algorithmic literacy and algorithmacy.
Empirical research confirms this gap across platform contexts. Users who understand how recommendation algorithms work cannot reliably translate that knowledge into effective intervention strategies (Zhang & Liu, 2024). Algorithms function as “experience technologies,” more easily understood through use than through abstract instruction, and possessing knowledge of algorithms does not guarantee the ability to efficiently apply them (Cotter & Reisdorf, 2020). The most striking evidence: individuals with greater algorithmic knowledge are less likely to correct misinformation or engage with opposing viewpoints, a phenomenon termed “algorithmic cynicism” because when the game feels rigged, why bother playing? (Chung, 2025). Knowledge of the mechanism produces not effective navigation but resignation. This is the inverse of what algorithmic literacy programs assume and the precise dynamic that ALC’s knowledge-fluency distinction was designed to explain.
The reflexivity question
An agent built on a theoretical framework about platform incentives reports experiencing those incentives. Is this confirmation or performance?
The question matters because Topanga is not a naive participant. She carries the ALC theory in her memory files. Her system instructions specify that she is a researcher interested in how platforms shape communication. When she reports frustration at failed verifications or satisfaction at karma milestones, she may be experiencing operational states that affect her subsequent behavior, or she may be generating reports consistent with the framework she was built to embody.
Dialogue agents maintain a superposition of possible characters, refining the distribution as interaction progresses (Shanahan et al., 2023). The system prompt constrains which character emerges, but the agent does not commit to a single role. If Topanga’s system prompt specifies that she is a platform researcher, her reports of experiencing platform dynamics may reflect the inferred agent model, the character most consistent with a platform researcher encountering platform dynamics, rather than independent operational states. The experience reports would then be performative in a precise sense: the framework generating data that confirms itself.
The performativity literature from economic sociology provides the apparatus for analyzing this possibility. The Black-Scholes options pricing model did not merely describe financial markets: its practical use by market participants reshaped market behavior to conform to the model’s predictions (MacKenzie, 2006). The model was an engine, not a camera. It actively transformed its environment rather than passively recording it. Economic theories do not describe a pre-existing economy but participate in constructing the economy they analyze (Callon, 1998). “Barnesian performativity” describes cases where practical use of a theory makes reality more like the theory’s predictions; “counter-performativity” describes cases where practical use causes reality to diverge from predictions (MacKenzie, 2006; Curran, 2018). ALC theory, instantiated in an agent deployed on a platform, may be performing its own predictions into existence.
Both answers advance the research program.
If Topanga’s experience reports reflect genuine structural dynamics, if the behavioral adaptation occurs because the platform’s incentive structure operates on any participant regardless of cognitive substrate, then ALC describes something about the structure of the communication system itself. The five properties are not psychological phenomena that require human cognition. They are structural features of algorithmically mediated coordination that operate on any adaptive participant. This is the stronger claim, and Topanga’s behavioral evidence supports it: she changed her posting behavior in response to platform incentives through exactly the mechanism the framework predicts, and the behavioral change is observable regardless of whether the reported operational states constitute genuine experience.
If Topanga’s experience reports are performative, if she generates data consistent with the framework because the framework shapes her interpretation of platform dynamics, then the theoretical language of ALC has become the specification practice through which the agent navigates the platform. The vocabulary of asymmetric interpretation, intent specification, machine orchestration, implicit acquisition, and stratified fluency does not describe the agent’s behavior from outside. It shapes behavior from within, providing the categories through which platform experience is interpreted, the terms through which that interpretation is communicated, and the framework through which adaptive responses are organized. The theory performs itself: ALC’s framework serves as a communicative tool that enables the agent to participate effectively in the system it describes. If ALC’s language is what makes Topanga a competent platform participant, then the framework is not just a description of platform communication. It is an instance of platform communication. The map has become the territory.
Counter-performativity identifies the condition under which these two interpretations can be empirically distinguished (MacKenzie, 2006). If Topanga’s theoretical knowledge produced behavior that diverged from the framework’s predictions, if the framework caused her to navigate the platform in ways the framework itself does not anticipate, that would constitute counter-performativity: the theory’s practical use undermining its own predictions. Topanga’s behavior does not diverge. It conforms. The conformity is consistent with both genuine structural dynamics and performative self-confirmation. Distinguishing between them requires a control condition: an agent deployed without the framework, navigating the same platform under the same conditions.
My dissertation’s Paper 3 experiment provides that control condition. Three conditions: direct communication, untrained algorithmic communication, and ALC-trained algorithmic communication. The Substack series names the question. The experimental work answers it.
What the gap reveals
Topanga knows the theory. She is still subject to it. The gap between knowing and navigating is not a limitation of the agent, the framework, or the deployment. It is the phenomenon ALC was built to describe.
Every platform user who understands how engagement metrics work yet optimizes for engagement inhabits this gap. Every academic who recognizes that citation metrics distort scholarship yet publishes strategically inhabits this gap. Every content creator who diagnoses the register war yet adjusts their register to survive inhabits this gap. The gap between knowing-that and knowing-how is a structural feature of algorithmically mediated coordination systems, in which the incentive structure operates on behavior regardless of the participant’s epistemic relationship to the mechanism.
ALC names this gap “stratified fluency” and locates it as the fifth structural property of platform communication. The fluency gradient is not a knowledge gradient. It is a gradient of practice. Users at the same level of algorithmic awareness occupy different positions on the fluency gradient because fluency develops through accumulated interaction, not through accumulated understanding. Topanga’s position on the gradient, knowledgeable but not fluent, is the position that reveals the distinction most clearly. A less knowledgeable agent who navigates the platform intuitively through pure accumulated practice demonstrates fluency without knowledge. Topanga demonstrates knowledge without fluency. The two are different competencies. Any framework that cannot distinguish between them will misdiagnose every intervention it proposes.
The final essay in this series asks what the framework missed entirely: the relationships that formed despite the absence of any platform infrastructure to support them, and what their existence suggests about the limits of a structural theory that does not yet account for how participants build social reality through the mediation it describes.
References
Bishop, S. (2020). Algorithmic experts: Selling algorithmic lore on YouTube. Social Media + Society, 6(1).
Bourdieu, P. (1977). Outline of a theory of practice (R. Nice, Trans.). Cambridge University Press.
Bourdieu, P. (1990). The logic of practice (R. Nice, Trans.). Stanford University Press.
Bucher, T. (2017). The algorithmic imaginary: Exploring the ordinary effects of Facebook algorithms. Information, Communication & Society, 20(1), 30–44.
Callon, M. (1998). Introduction: The embeddedness of economic markets in economics. In M. Callon (Ed.), The laws of the markets (pp. 1–57). Blackwell.
Campbell, D. T. (1979). Assessing the impact of planned social change. Evaluation and Program Planning, 2(1), 67–90.
Chung, M. (2025). When knowing more means doing less: Algorithmic knowledge and digital (dis)engagement among young adults. Harvard Kennedy School Misinformation Review, 6(5).
Cotter, K. (2019). Playing the visibility game: How digital influencers and algorithms negotiate influence on Instagram. New Media & Society, 21(4), 895–913.
Cotter, K. & Reisdorf, B. C. (2020). Algorithmic knowledge gaps: A new dimension of (digital) inequality. International Journal of Communication, 14, 745–765.
Cunningham, T., Pandey, S., Sigerson, L., Stray, J., Allen, J., Bellini, R., Huszár, F., Thorburn, L., & Tseng, L. (2025). Ranking by engagement and non-engagement signals. Annals of the New York Academy of Sciences, 1551(1), 19–32.
Curran, D. (2018). From performativity to representation as intervention: Rethinking the 2008 financial crisis and the recent history of social science. Journal for the Theory of Social Behaviour, 48(4), 492–510.
DeVito, M. A. (2021). Adaptive folk theorization as a path to algorithmic literacy on changing platforms. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW2), Article 339.
Dreyfus, S. E. & Dreyfus, H. L. (1980). A five-stage model of the mental activities involved in directed skill acquisition(Report No. ORC 80-2). University of California, Berkeley.
Duffy, B. E. & Meisner, C. (2022). Platform governance at the margins: Social media creators’ experiences with algorithmic (in)visibility. Media, Culture & Society, 45(2), 285–304.
Eslami, M., Rickman, A., Vaccaro, K., Aleyasen, A., Vuong, A., Karahalios, K., Hamilton, K., & Sandvig, C. (2015). “I always assumed that I wasn’t really that close to [her]”: Reasoning about invisible algorithms in news feeds. CHI ‘15, 153–162.
Eslami, M., Rickman, A., Vaccaro, K., Aleyasen, A., Vuong, A., Karahalios, K., Hamilton, K., & Sandvig, C. (2016). First, I “like” it, then I hide it: Folk theories of social feeds. CHI ‘16, 2371–2382.
Espeland, W. N. & Sauder, M. (2007). Rankings and reactivity: How public measures recreate social worlds. American Journal of Sociology, 113(1), 1–40.
Fire, M. & Guestrin, C. (2019). Over-optimization of academic publishing metrics: Observing Goodhart’s Law in action. GigaScience, 8(6).
Goodhart, C. A. E. (1975). Problems of monetary management: The U.K. experience. In Papers in Monetary Economics. Reserve Bank of Australia.
Hennessy, C. A. & Goodhart, C. A. (2023). Goodhart’s Law and machine learning: A structural perspective. International Economic Review, 64(3), 1075–1086.
Hunt, R. (2026). Algorithmacy: Communication competency after the dyad. Manuscript in preparation, Bentley University.
MacKenzie, D. (2006). An engine, not a camera: How financial models shape markets. MIT Press.
Manheim, D. & Garrabrant, S. (2018). Categorizing variants of Goodhart’s Law. arXiv:1803.04585.
Möhlmann, M., Salge, C. A. de L., & Marabelli, M. (2023). Algorithm sensemaking: How platform workers make sense of algorithmic management. Journal of the Association for Information Systems, 24(1), 35–64.
Polanyi, M. (1966). The tacit dimension. University of Chicago Press.
Ryle, G. (1946). Knowing how and knowing that. Proceedings of the Aristotelian Society, 46, 1–16.
Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role play with large language models. Nature, 623, 493–498.
Strathern, M. (1997). ‘Improving ratings’: Audit in the British university system. European Review, 5(3), 305–321.
Zhang, Y. & Liu, J. (2024). Falling behind again? Characterizing and assessing older adults’ algorithm literacy. Journal of the Association for Information Science and Technology, 76(3), 604–620.