pith. sign in

arxiv: 2605.21785 · v1 · pith:FULOHPN2new · submitted 2026-05-20 · 💻 cs.CY

Machine Learning as Performative Materialist Practice: Thirteen Theses on the Epistemology, Methodology, and Politics of Applied ML

Adolfo De Un\'anue , Fernanda Sobrino This is my paper

Pith reviewed 2026-05-22 08:15 UTC · model grok-4.3

classification 💻 cs.CY
keywords machine learningperformativityepistemologypublic policymaterialismbounded rationalityadaptive systems
0
0 comments X

The pith

Machine learning models in decision-support contexts function as temporally situated instruments of intervention whose validity is measured by their effects rather than representational accuracy.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper challenges four unexamined commitments in applied machine learning: that models represent stable regularities, that validation can be context-free, that performance metrics are politically neutral, and that feature importance reveals system structure. It advances a performative materialist framework in which models are compressions that intervene in the world, the full data product is a complex adaptive system coevolving with its targets, and validity is judged by produced effects. Choices in objective functions, fairness criteria, and thresholds are therefore political decisions that belong to stakeholders. The thirteen theses unify techniques such as temporal cross-validation, precision at k, and satisficing as direct consequences of this epistemology rather than ad-hoc best practices.

Core claim

Machine learning models are best understood not as truth-seeking representations but as temporally situated compressions that function as instruments of intervention; the full data product is a complex adaptive system that coevolves with its target and navigates a multi-objective space no single algorithm can optimize; validity is fundamentally performative, measured by effects in the world rather than formal properties of the model; and the choices embedded in objective functions, fairness criteria, and resource thresholds are political decisions belonging to stakeholders, not technicians.

What carries the argument

The unified framework of performative materialist ML articulated as thirteen theses that reinterprets models as intervention instruments within coevolving adaptive systems.

If this is right

  • Temporal cross-validation and pipeline-aware auditing become required rather than optional because of coevolution.
  • Satisficing replaces optimization as the appropriate stance in multi-objective institutional settings.
  • Stakeholder participation in defining objective functions and fairness criteria follows directly from the political character of those choices.
  • Context-free validation is replaced by evaluation that tracks effects after deployment.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This view implies that AI governance should focus on monitoring and adjusting intervention effects rather than certifying model properties in isolation.
  • It connects applied ML to older debates in cybernetics and systems theory about feedback and adaptation in policy instruments.
  • Testable extensions include comparing performative versus representational validation protocols in live public-health or criminal-justice deployments.

Load-bearing premise

The full data product constitutes a complex adaptive system that coevolves with its target and navigates a multi-objective space no single algorithm can optimize.

What would settle it

An empirical study that deploys the same model without retraining over multiple cycles, measures stable formal performance metrics, yet records large shifts in real-world outcomes that contradict initial intervention goals.

Figures

Figures reproduced from arXiv: 2605.21785 by Adolfo De Un\'anue, Fernanda Sobrino.

Figure 1
Figure 1. Figure 1: The data product as a complex adaptive system. The model informs decisions; decisions modify the system being modeled; the modified system generates new data; the new data updates the model. The loop is continuous and recursive—there is no clean separation between observation and intervention. 3.3 Thesis III: Radical Contextuality The model is radically contextual. It depends on the organization and the mo… view at source ↗
read the original abstract

Machine learning practice in institutional decision-support contexts -- government, public policy, public health, criminal justice, resource allocation -- rests on a set of largely unexamined epistemological commitments inherited from classical statistics and computer science: that models represent stable regularities, that validation can be context-free, that performance metrics are politically neutral, and that feature importance reveals system structure. This paper challenges these commitments through a unified framework of performative materialist ML, articulated as thirteen theses. Drawing on Pickering's cybernetic ontology, the performativity literature from economic sociology (Callon, MacKenzie), Simon's bounded rationality, the formalization of performative prediction (Perdomo et al., 2020), and fifteen years of applied ML experience in government and public policy, we argue that: (1) ML models are best understood not as truth-seeking representations but as temporally situated compressions that function as instruments of intervention; (2) the full data product is a complex adaptive system that coevolves with its target and navigates a multi-objective space no single algorithm can optimize; (3) validity is fundamentally performative, measured by effects in the world rather than formal properties of the model; (4) the choices embedded in objective functions, fairness criteria, and resource thresholds are political decisions belonging to stakeholders, not technicians. We show how these theses unify several practical prescriptions -- temporal cross-validation, precision and recall at k, pipeline-aware fairness auditing, satisficing over optimizing -- as consequences of a coherent materialist epistemology rather than isolated best practices

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript articulates a performative materialist framework for machine learning in institutional decision-support settings through thirteen theses. It challenges inherited commitments from statistics and computer science by claiming that models are temporally situated compressions functioning as instruments of intervention rather than truth-seeking representations; that the full data product forms a coevolving complex adaptive system navigating an unoptimizable multi-objective space; that validity is performative and measured by real-world effects; and that choices in objectives, fairness criteria, and thresholds are political decisions for stakeholders. The theses are presented as unifying practical prescriptions such as temporal cross-validation, precision at k, pipeline-aware fairness auditing, and satisficing over optimizing as coherent consequences of this epistemology rather than isolated heuristics, drawing on Pickering, Callon, MacKenzie, Simon, and Perdomo et al. 2020.

Significance. If the central claims hold, the framework supplies a unified epistemological and ontological grounding for applied ML in government, policy, and justice domains. It reframes validity, fairness, and optimization in terms of performativity and materialist coevolution, potentially guiding practitioners toward intervention-oriented evaluation and stakeholder-driven choices. The synthesis of existing literature with applied experience offers a coherent alternative to representational views, though its impact depends on the strength of the grounding for the adaptive-system premise.

major comments (2)
  1. [Abstract and Thesis 2] Abstract and discussion of Thesis 2: The assertion that the full data product constitutes a complex adaptive system that coevolves with its target and navigates a multi-objective space no single algorithm can optimize is presented without formal characterization of the objective space, counterexample analysis, or demonstration that standard multi-objective methods (Pareto optimization or weighted sums) are inapplicable due to coevolution. This premise is load-bearing for the shift from representation to intervention and from optimization to satisficing.
  2. [Unification of prescriptions] Section unifying practical prescriptions: The claim that temporal cross-validation, precision and recall at k, and pipeline-aware fairness auditing follow as direct consequences of the materialist epistemology risks circularity. These prescriptions appear to derive from the initial interpretive assumptions about performativity and coevolution without independent external grounding or explicit mapping showing why they are entailed rather than merely compatible.
minor comments (2)
  1. [Abstract] The abstract references fifteen years of applied ML experience in government and public policy, but the manuscript should indicate more explicitly in which theses or sections this experience provides concrete illustrations or modifications to the cited theoretical sources.
  2. [Throughout] Notation for key terms such as 'data product' and 'performative validity' could be defined more precisely on first use to aid readers unfamiliar with the performativity literature.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and substantive comments, which help clarify the grounding of our framework. We respond to each major comment below and indicate the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract and Thesis 2] The assertion that the full data product constitutes a complex adaptive system that coevolves with its target and navigates a multi-objective space no single algorithm can optimize is presented without formal characterization of the objective space, counterexample analysis, or demonstration that standard multi-objective methods (Pareto optimization or weighted sums) are inapplicable due to coevolution. This premise is load-bearing for the shift from representation to intervention and from optimization to satisficing.

    Authors: We accept that a more explicit characterization would strengthen the argument. The manuscript already draws on Perdomo et al. (2020) for performative effects and Simon for satisficing, but we will add a short subsection under Thesis 2 that sketches why coevolution renders static multi-objective optimization inadequate. This will include a brief counterexample from predictive resource allocation, where optimizing one objective shifts the target distribution and invalidates prior Pareto fronts. The revision will preserve the paper's conceptual emphasis while addressing the load-bearing premise. revision: yes

  2. Referee: [Unification of prescriptions] The claim that temporal cross-validation, precision and recall at k, and pipeline-aware fairness auditing follow as direct consequences of the materialist epistemology risks circularity. These prescriptions appear to derive from the initial interpretive assumptions about performativity and coevolution without independent external grounding or explicit mapping showing why they are entailed rather than merely compatible.

    Authors: We take the circularity concern seriously and will revise for greater transparency. The unification section is meant to show internal coherence with the theses rather than strict deductive entailment. In the revised manuscript we will insert a concise mapping (as a table or enumerated list) that traces each prescription to specific theses and the cited sources (e.g., temporal cross-validation to Thesis 1's temporal situatedness and Thesis 3's performativity). This makes the connections explicit while maintaining that the prescriptions are not isolated heuristics but follow from the overall epistemology. revision: partial

Circularity Check

0 steps flagged

No significant circularity; conceptual framework grounded in external citations

full rationale

The paper articulates thirteen theses as an interpretive framework drawing explicitly on Pickering's cybernetic ontology, Callon and MacKenzie's performativity literature, Simon's bounded rationality, Perdomo et al. 2020 on performative prediction, and the authors' applied experience. The unification of prescriptions such as temporal cross-validation and satisficing as consequences of the materialist epistemology is presented as an argumentative synthesis rather than a formal derivation or statistical fit. No equations, parameter fitting, or self-referential definitions appear in the provided text; the central premises about coevolution and multi-objective spaces are supported by the cited external sources and are not shown to reduce to internal construction by the paper's own logic. This is a standard non-circular finding for a philosophical/theoretical paper.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claims rest on reinterpretations of cited philosophical and sociological sources without new empirical or formal support; the framework itself functions as an invented interpretive lens rather than a derived result.

axioms (2)
  • domain assumption Machine learning practice in institutional decision-support contexts rests on unexamined epistemological commitments inherited from classical statistics and computer science
    Opening statement of the abstract that frames the entire argument.
  • ad hoc to paper Validity is fundamentally performative, measured by effects in the world rather than formal properties of the model
    Presented as thesis (3) and load-bearing for redefining success criteria.
invented entities (1)
  • performative materialist ML framework no independent evidence
    purpose: Unified framework articulating epistemology, methodology, and politics of applied ML as thirteen theses
    Introduced as the coherent materialist epistemology that unifies prior concepts and practical prescriptions.

pith-pipeline@v0.9.0 · 5832 in / 1538 out tokens · 59725 ms · 2026-05-22T08:15:18.676685+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

  1. [1]

    and Jesus, S

    Amarasinghe, Kasun and Rodolfa, Kit T. and Jesus, S. On the Importance of Application-Grounded Experimental Design for Evaluating Explainable

    work page

  2. [2]

    , title =

    Amarasinghe, Kasun and Ghani, Rayid and Lai, Andy and Millan, Liliana and Rodolfa, Kit T. , title =

    work page

  3. [3]

    Ross , title =

    Ashby, W. Ross , title =

    work page

  4. [4]

    Bateson, Gregory , title =

    work page

  5. [5]

    Beer, Stafford , title =

    work page

  6. [6]

    Black, Emily and Naidu, Rakesh and Ghani, Rayid and Rodolfa, Kit and Ho, Daniel and Heidari, Hoda , title =

    work page

  7. [7]

    Statistical Science , volume =

    Breiman, Leo , title =. Statistical Science , volume =

    work page

  8. [8]

    The Laws of the Markets , editor =

    Callon, Michel , title =. The Laws of the Markets , editor =

    work page

  9. [9]

    Do Economists Make Markets? On the Performativity of Economics , editor =

    Callon, Michel , title =. Do Economists Make Markets? On the Performativity of Economics , editor =

    work page

  10. [10]

    Big Data , volume =

    Chouldechova, Alexandra , title =. Big Data , volume =

    work page

  11. [11]

    Paul and Cottrell, Allin F

    Cockshott, W. Paul and Cottrell, Allin F. , title =

    work page

  12. [12]

    Designing

    Sobrino, Fernanda and De Un. Designing

    work page

  13. [13]

    , title =

    Holland, John H. , title =

    work page

  14. [14]

    Kleinberg, Jon and Mullainathan, Sendhil and Raghavan, Manish , title =

    work page

  15. [15]

    MacKenzie, Donald , title =

    work page

  16. [16]

    American Journal of Sociology , volume =

    MacKenzie, Donald and Millo, Yuval , title =. American Journal of Sociology , volume =

    work page

  17. [17]

    Pask, Gordon , title =

    work page

  18. [18]

    and Zrnic, Tijana and Mendler-D

    Perdomo, Juan C. and Zrnic, Tijana and Mendler-D. Performative Prediction , booktitle =

    work page

  19. [19]

    Pickering, Andrew , title =

    work page

  20. [20]

    and De Un

    Rodolfa, Kit T. and De Un. An Experience-Centered Approach to Training Effective Data Scientists , journal =

    work page

  21. [21]

    and Ghani, Rayid , title =

    Rodolfa, Kit T. and Ghani, Rayid , title =. Harvard Data Science Review , volume =

    work page

  22. [22]

    and Lamba, Hemank and Ghani, Rayid , title =

    Rodolfa, Kit T. and Lamba, Hemank and Ghani, Rayid , title =. Nature Machine Intelligence , volume =

    work page

  23. [23]

    and Ghani, Rayid , title =

    Saleiro, Pedro and Kuester, Benedict and Hinkson, Loren and London, Jesse and Stevens, Abby and Anisfeld, Ari and Rodolfa, Kit T. and Ghani, Rayid , title =

    work page

  24. [24]

    and Holt, Gary and Golovin, Daniel and Davydov, Eugene and Phillips, Todd and Ebner, Dietmar and Chaudhary, Vinay and Young, Michael and Crespo, Jean-Fran

    Sculley, D. and Holt, Gary and Golovin, Daniel and Davydov, Eugene and Phillips, Todd and Ebner, Dietmar and Chaudhary, Vinay and Young, Michael and Crespo, Jean-Fran. Hidden Technical Debt in Machine Learning Systems , booktitle =

    work page

  25. [25]

    , title =

    Simon, Herbert A. , title =. Psychological Review , volume =

    work page

  26. [26]

    , title =

    Simon, Herbert A. , title =. Decision and Organization , editor =

    work page

  27. [27]

    Performative Prediction: Past and Future , journal =

    Hardt, Moritz and Mendler-D. Performative Prediction: Past and Future , journal =. 2025 , note =

    work page 2025

  28. [28]

    Performative Power , booktitle =

    Hardt, Moritz and Jagadeesan, Meena and Mendler-D. Performative Power , booktitle =

    work page

  29. [29]

    , title =

    Narang, Adhyyan and Faulkner, Evan and Drusvyatskiy, Dmitriy and Fazel, Maryam and Ratliff, Lillian J. , title =. Journal of Machine Learning Research , volume =

    work page

  30. [30]

    Toward Reducing Unproductive Container Moves: Predicting Service Requirements and Dwell Times , journal =

    Villalobos, Elena and De Un. Toward Reducing Unproductive Container Moves: Predicting Service Requirements and Dwell Times , journal =

    work page

  31. [31]

    Issenberg, Sasha , title =

    work page

  32. [32]

    and Nguyen, Austin and Kimpel, Zoe and Sidgwick, Joshua and De Un

    Wilde, Harrison and Chen, Lucia L. and Nguyen, Austin and Kimpel, Zoe and Sidgwick, Joshua and De Un. A Recommendation and Risk Classification System for Connecting Rough Sleepers to Essential Outreach Services , journal =

    work page

  33. [33]

    Improving Government Response to Citizen Requests Online , booktitle =

    Gaut, Garren and Navarrete, Andrea and Wahedi, Laila and van der Boor, Paul and De Un. Improving Government Response to Citizen Requests Online , booktitle =

    work page

  34. [34]

    Deploying Machine Learning Models for Public Policy: A Framework , booktitle =

    Ackermann, Klaus and Walsh, Joe and De Un. Deploying Machine Learning Models for Public Policy: A Framework , booktitle =

    work page

  35. [35]

    Criminal Justice Policy Review , volume =

    Helsby, Jennifer and Carton, Samuel and Joseph, Kenneth and Mahmud, Ayesha and Park, Youngsoo and Navarrete Rivera, Andrea and Ackermann, Klaus and Walsh, Joe and Haynes, Lauren and Cody, Crystal and Patterson, Estella Patterson and Ghani, Rayid , title =. Criminal Justice Policy Review , volume =

    work page

  36. [36]

    Potash, Eric and Brew, Joe and Loewi, Alexander and Majumdar, Subhabrata and Reece, Andrew and Walsh, Joe and Rozier, Eric and Jarpe-Ratner, Emile and Gabriel, Alejandro and Ghani, Rayid , title =

    work page

  37. [37]

    Carton, Samuel and Helsby, Jennifer and Joseph, Kenneth and Mahmud, Ayesha and Park, Youngsoo and Walsh, Joe and Cody, Crystal and Patterson, Estella and Haynes, Lauren and Ghani, Rayid , title =

    work page

  38. [38]

    and Boxer, Kate S

    Bauman, Matthew J. and Boxer, Kate S. and Lin, Tzu-Yun and Salomon, Erika and Naveed, Hareem and Haynes, Lauren and Walsh, Joe and Helsby, Jen and Yoder, Steve and Sullivan, Robert and Schneweis, Chris and Ghani, Rayid , title =

    work page

  39. [39]

    Applying Machine Learning Methods to Enhance the Distribution of Social Services in

    Sankaran, Kris and Garcia-Olano, Diego and Javed, Mobin and Alcala-Durand, Maria Fernanda and De Un. Applying Machine Learning Methods to Enhance the Distribution of Social Services in

    work page

  40. [40]

    A Machine Learning System for Retaining Patients in

    Kumar, Avishek and Ramachandran, Arthi and De Un. A Machine Learning System for Retaining Patients in

    work page

  41. [41]

    and Salomon, Erika and Haynes, Lauren and Higuera Mendieta, Iv

    Rodolfa, Kit T. and Salomon, Erika and Haynes, Lauren and Higuera Mendieta, Iv. Case Study: Predictive Fairness to Reduce Misdemeanor Recidivism through Social Service Interventions , booktitle =

    work page

  42. [42]

    and Lamba, Hemank and Ghani, Rayid , title =

    Rodolfa, Kit T. and Lamba, Hemank and Ghani, Rayid , title =

    work page