{"id":"b671572f-ea06-4cd5-a630-2a9d44305e90","arxiv_id":"2507.13616","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A conceptual framework uniting multi-level selection, Aoki's computational firm, and Ostrom's rules for AI governance, with an unproven multi-level Price equation at its core.","lead":"This paper proposes that AI governance in firms can be understood through evolution, with selection acting on teams, firms, and rules at the same time. It claims a quantitative equation for how such rules evolve, but that equation is asserted without derivation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (2) is asserted as an exact decomposition of Δπ, but its cross-agent covariance terms are not derivable from the multilevel Price equation in Eq. (1); no derivation is provided, and the promised quantitative metrics are therefore ungrounded.","rationale":"The reader's weakest assumption identifies exactly the load-bearing weakness: Eq. (2) is not a standard consequence of the Price equation, and the cross-agent covariance terms are asserted without derivation. My reading of Section 2.1.2 and Appendix B confirms this. The paper's promised quantitative decomposition stands or falls on Eq. (2); if the cross terms are not derived from a specified fitness model, the equality is not justified. The rest of the paper has real heuristic value: the graduated-sanctions Prisoner's Dilemma example in Section 3 is internally correct, the Ostrom principles are mapped to AI governance plausibly, and the Aoki/Acemoglu synthesis is a reasonable framing exercise. But those contributions do not rescue the central mathematical claim. The paper also contains broken cross-references (e.g., 'Section ??', 'Equation ??') that reinforce the impression of a loosely integrated formal core, though I do not treat those as the primary objection. Since my concern coincides with the reader's, I see no reason to change the reader's REJECT verdict.","tokens_in":24028,"tokens_out":5712,"duration_ms":69672,"concrete_test":"Independently re-derive Eq. (2) from Eq. (1): write group performance π_g as the weighted average of human and AI trait values, apply the standard multilevel Price identity to the pooled population of human and AI agents, and inspect the resulting right-hand side for arbitrary fitness functions. If the exact RHS reduces to Eq. (2), the concern fails; if the cross terms appear only under special assumptions or never appear, Eq. (2) is unsupported. Then run a two-group numerical check: assign arbitrary human and AI traits, fitnesses, and a deterministic one-generation transition rule; compute wΔπ exactly and compare it with the RHS of Eq. (2) including both cross-agent covariances. Any nonzero difference with no residual term in Eq. (2) settles the issue.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central formal claim is Eq. (2), which states an exact decomposition of wΔπ into between-group selection, within-group human selection, within-group AI selection, and two cross-agent terms, Eg[Cov(h,ai)(w_h, π_ai)] and Eg[Cov(h,ai)(w_ai, π_h)]. These cross terms are covariances between the fitness of one agent type and the performance of the other type. Such terms are not consequences of the standard multilevel Price identity in Eq. (1), which decomposes change in mean trait into covariance between own fitness and own trait plus transmission bias. Section 2.1.2 simply states Eq. (2); Appendix B provides definitions, computational formulas (e.g., B.10.3), and causal interpretations, but never derives the equality. A correct partition of the pooled within-group covariance into human and AI subpopulations yields within-category covariances plus a covariance between category means, not interaction-weighted cross-covariances of w_h with π_ai. Those cross terms could appear only under an explicit model in which human fitness depends on AI performance (and vice versa), with any additional terms and residuals accounted for. No such model is specified. Because the abstract's promised 'quantitative metrics' and the rule-comparison via replicator dynamics in Section 4 both depend on Eq. (2), this is the load-bearing step and it is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified framework for AI governance that combines multilevel selection theory, Aoki's view of firms as computational processes, and Ostrom's design principles. The central formal device is an extended Price equation (Eq. 2) that decomposes the change in average organizational performance into between-group selection, within-group human selection, within-group AI selection, and two cross-agent covariance terms. The paper then embeds this in a replicator-dynamics model for selecting among institutional rule configurations (Section 4), illustrates the approach with a graduated-sanctions prisoner's dilemma (Section 3), and offers four qualitative case studies plus policy recommendations. The worked example is mathematically sound, and the qualitative synthesis is broad, but the central equation is asserted without derivation and is not a standard consequence of the multilevel Price equation.","tokens_in":24347,"tokens_out":3625,"duration_ms":48304,"significance":"If the extended Price equation were a valid general decomposition, the framework would provide a genuinely useful quantitative language for analyzing human-AI co-evolution in organizations: it would let analysts separate between-group, within-human, within-AI, and cross-agent selection effects, and it would connect institutional design to evolutionary outcomes. The paper also gives proper attention to Ostrom's principles as mechanisms that reshape payoff structures, and the sanctioned-PD example correctly shows that λki > 1 makes defection strictly dominated. That example is a real, if modest, formal contribution. The broader significance, however, depends almost entirely on the unsupported Eq. (2), because the abstract's promised 'quantitative metrics' and the rule-comparison via replicator dynamics both rest on it.","major_comments":[{"comment":"The extended Price equation is asserted as an exact decomposition of wΔπ, but it is not derived from Eq. (1). In the standard multilevel Price equation, the within-group term is a covariance between an agent's own fitness and its own trait; partitioning the pooled within-group covariance by agent type gives within-category covariances plus a covariance of category means, not cross-covariances such as Cov(h,ai)(w_h, π_ai) and Cov(h,ai)(w_ai, π_h). Those cross terms can appear only under an explicit model in which the fitness of humans depends on the performance of AI agents and vice versa, with any additional terms or residuals accounted for. Appendix B, including Sections B.7-B.10, provides interpretations, causal stories, and computational formulas but never proves the equality. Because the abstract's 'quantitative metrics' and the Section 4 rule-comparison both depend on Eq. (2), the central formal claim of the paper is unsupported.","section":"Section 2.1.2, Eq. (2)"},{"comment":"The fitness of a rule configuration is defined as Vj = fj(π*_j) − cj, with fj(π*_j) = απ*_h,j + βπ*_ai,j + γ Cov(π*_h, π*_ai). The coefficients α, β, γ are introduced without derivation, calibration, or an explicit argument that this particular linear-plus-covariance form is the correct mapping from agent-level equilibria to organizational fitness. The claimed feedback loop between the extended Price equation and the replicator dynamics is also asserted rather than established: no formal argument shows that the πg variables in Eq. (2) coincide with the equilibrium payoffs π* from Eq. (4). As a result, 'quantitative comparison among institutional rule sets' is not yet demonstrated; it is a research agenda.","section":"Section 4.2, Eqs. (5)-(7)"},{"comment":"The section is introduced as 'two case studies' but actually presents four: algorithmic trading firms, Amazon's automated scheduling, ICANN, and Maine lobster fisheries. The later sentence 'Both cases validate...' is therefore ambiguous about which two cases are meant. This is a presentation problem, but it matters because the case-study section is offered as empirical support for the framework.","section":"Section 5"}],"minor_comments":[{"comment":"The text contains an unresolved cross-reference 'Section ??'; the intended section should be identified.","section":"Section 6.1"},{"comment":"The citation '[8 ? ]' is incomplete and should be resolved.","section":"Appendix C.4"},{"comment":"The abstract ends with two consecutive sentences beginning 'We conclude...'; one of them should be rephrased.","section":"Abstract"},{"comment":"The Discussion mentions 'healthcare organizations' as if they were a case study, but no healthcare case study appears in Section 5.","section":"Discussion, Section 8"},{"comment":"The notation 'Πj = Tj(Π0)' is not used consistently: in Section 3 the sanctioned matrix is called Πs(k1, k2), while Section 4 refers to a generic rule j without defining the transformation Tj for the bottom-up strategies (tit-for-tat and win-stay, lose-switch).","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The central formal contribution is an asserted identity that the appendix does not derive, and the quantitative claims of the paper rest on that identity. I do not see this as a case of a small missing step: the cross-agent covariance terms are not a standard decomposition of the pooled within-group Price term, and their validity requires an explicit model of fitness interdependence that the paper never provides. If the authors can replace Eq. (2) with a properly derived decomposition under stated assumptions, a resubmission could be worth considering, but in the current form the framework's advertised quantitative metrics are ungrounded."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper reads as a broad, informed synthesis of multi-level selection, Aoki's computational view of the firm, and Ostrom's design principles applied to AI governance. The genuinely new piece—the extended Price equation with human-AI cross-covariances (Eq. 2)—is where it falls apart. Those cross-agent terms are asserted, not derived. Appendix B gives computational recipes and interpretations, but not a mathematical justification. Since the abstract promises 'quantitative metrics' and Section 4's rule-selection dynamics lean on this equation, the central quantitative claim is unsupported.\n\nWhat's good: the mapping of Ostrom's principles to 'alignment operators' is a useful heuristic. The worked prisoner's dilemma with graduated sanctions (λk_i) is correct and nicely shows how a sanction can make defection strictly dominated. The case studies, though qualitative, are on-topic, and using ICANN and Maine lobster as polycentric examples is apt. The synthesis of Aoki's computational perspective with multi-level selection is interesting as framing.\n\nThe problems are not cosmetic. Eq. (2) is not a consequence of Eq. (1). The standard multilevel Price partition separates within-type covariances and a between-group term; it does not yield covariances between the fitness of one agent type and the performance of the other. Those terms could appear under an explicit model where human fitness depends on AI performance and vice versa, but no such model is specified. The weighted-covariance formula in B.10.3 defines a statistic, but it does not show that statistic belongs in the decomposition. There are also smaller issues: Section 5 says 'two case studies' while listing four, and the text contains broken cross-references (Section ??, Equation ??). The prose sometimes overstates the empirical support.\n\nWho is this for? Someone working on AI governance who wants a conceptual vocabulary connecting institutional economics and evolutionary theory. It could become a useful discussion piece if the formal claims are toned down to 'suggestive metrics' rather than 'decomposition.' As is, I would not cite the quantitative framework. If the author can either derive Eq. (2) under explicit causal assumptions or reframe the paper as a position piece without the formal apparatus, it could be a solid review paper. I would send it to peer review because the topic is important and the flaws are fixable in revision—a good referee could force the clarifying rewrite. But the current version should not be published as is.","headline":"A plausible conceptual synthesis of MLS, Aoki, Ostrom, and Acemoglu for AI governance, but the central extended Price equation is asserted without derivation, so the promised quantitative metrics are ungrounded.","tokens_in":24874,"tokens_out":2797,"would_cite":false,"duration_ms":33327,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A22","92D15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that AI governance can be modelled as multi-level selection acting on firms, humans, and AI together, with institutional rules evolving by replicator dynamics.","keywords":["AI governance","multi-level selection","Price equation","agential AI","institutional economics","evolutionary game theory","alignment","polycentric governance"],"falsifier":"Take a simulated or real organization, measure each human and AI agent's fitness and performance at two time points, compute the between-group, within-human, within-AI, and weighted cross-agent covariance terms using interaction weights $\\omega_{h,ai}$, and compare their sum with the directly observed change in mean performance; if a systematic residual remains—especially one traceable to within-pair interaction products—the extended decomposition in Equation (2) does not close and the cross-terms are not, by themselves, a valid accounting of organizational change.","tokens_in":23800,"feed_emoji":"⚖️","tokens_out":9752,"duration_ms":109789,"temperature":0.7,"pith_summary":"The paper tries to give AI governance a quantitative evolutionary footing. When firms operate with a mix of human and agential AI, it claims the change in average organizational performance can be decomposed by an extended multi-level Price equation into selection acting between groups, within human populations, within AI populations, and across human-AI interaction pairs. Institutional rules enter the framework as transformations of game payoff matrices, and rule configurations are themselves selected by replicator dynamics: a rule survives when the equilibrium performance it generates outweighs its implementation and enforcement costs. If the framework is right, governance is no longer only a design question—it becomes a measurable evolutionary process, and analysts could attribute changes in firm performance to specific selection channels and compare institutional designs side by side.","feed_headline":"AI governance in firms gets a quantitative selection equation","feed_subtitle":"A multi-level Price equation shows how groups, humans, AI, and human-AI pairs drive performance changes.","key_machinery":"The load-bearing object is the extended multi-level Price equation over nested games (Equation 2), which adds human-AI cross-covariance terms to the standard between-group and within-group decomposition. It is paired with a replicator equation over institutional rule configurations, where a rule's fitness is $V_j = f_j(\\pi^*_j) - c_j$, with $f_j$ combining average human equilibrium payoff, average AI equilibrium payoff, and the human-AI covariance, and $c_j$ the rule's monitoring, enforcement, coordination, and adaptation costs. Micro-level behavior is fixed by an institutional equilibrium in which each agent's strategy is optimal against beliefs that are themselves shaped by the shared rule structure. The graduated-sanctions example shows the proposed mechanism in miniature: when the alignment parameter satisfies $\\lambda k_i > 1$, defection becomes strictly dominated, converting the Prisoner's Dilemma into a coordination game and demonstrating how institutional memory modifies selection pressures.","core_discovery":"The central claim is that institutional governance of agential AI can be treated as a nested evolutionary process. The extended Price equation (Equation 2) states that the weighted change in average organizational performance equals the between-group covariance of group fitness and performance, plus the expected within-group covariances for human agents and for AI agents, plus two cross-agent covariance terms: one linking human fitness with AI performance and one linking AI fitness with human performance. The cross-terms are argued to be irreducible and to capture complementarity or substitution between humans and algorithms. Institutional rules are then modeled as alignment operators that transform baseline payoff matrices, and rule configurations evolve by replicator dynamics with fitness equal to equilibrium performance minus implementation cost. The paper concludes that selection and governance co-determine economic outcomes, expressible in equations that can in principle be measured and compared.","pith_inferences":["Inference (not in the paper): the sign and magnitude of the two cross-agent covariance terms in Equation (2) could be read as a continuous alignment metric for a firm, signalling misalignment even when within-type selection looks healthy, provided organizations log interaction weights $\\omega_{h,ai}$ between employees and deployed AI systems.","Inference (not in the paper): the cost-benefit structure $V_j = f_j(\\pi^*_j) - c_j$ predicts convergence to hybrid institutional structures—top-down constitutional rules layered with bottom-up local norms—under mixed human-AI workforces, because pure centralized sanctions carry high monitoring costs while pure peer rules diffuse slowly; agent-based simulations that vary $c_j$ and diffusion rates c","Inference (not in the paper): the framework implies that field experiments should deliberately pair high- and low-performing humans with better and worse AI tools, because that variation directly estimates the two cross-terms and reveals whether an organization should invest in complementarity or substitution.","Inference (not in the paper): by analogy with cross-species symbiosis, the cross-terms should rarely vanish in real organizations, so designing human and AI incentives separately will miss precisely the interaction that the paper identifies as driving organizational evolution."],"forward_implications":["Every candidate rule set becomes a transformed payoff matrix $\\Pi_j = T_j(\\Pi_0)$, and its viability is ranked by $V_j = f_j(\\pi^*_j) - c_j$, turning institutional design debates into direct replicator-dynamics comparisons.","The extended Price equation splits organizational change into separate channels, so a firm can attribute a performance shift to between-group competition, human selection, AI selection, or human-AI interaction effects rather than treating them as one undifferentiated outcome.","Graduated sanctions with $\\lambda k_i > 1$ convert the Prisoner's Dilemma into a game where cooperation is strictly dominant, showing how institutional memory alone can enforce alignment.","The design principles act on specific covariance terms—monitoring and graduated sanctions can make poor performance negatively covary with fitness, while collective-choice arrangements can strengthen the human-AI complementarity term—so governance interventions have predicted signatures in the decomposition.","Inclusive institutions should strengthen between-group selection and positive human-AI cross-covariance, whereas extractive institutions concentrate selection within groups; the paper reads the algorithmic trading, automated scheduling, internet-governance, and lobster-fishery cases as evidence of these dynamics."],"supporting_citations":[{"why":"Supplies the covariance selection mathematics that Equation (1) generalizes to organizational performance.","marker":"[23]"},{"why":"Provides the modern Price-equation form and its causal interpretation, the baseline for the AI extension.","marker":"[25]"},{"why":"Gives the multi-level selection foundations for the between-group and within-group decomposition.","marker":"[5]"},{"why":"Supplies the firm-as-computation view, with the firm as an equilibrium of interrelated games.","marker":"[6]"},{"why":"Source of the eight design principles repurposed as alignment operators that reshape payoffs.","marker":"[7]"},{"why":"Supplies the institutional-diversity analysis behind rule costs and the persistence of governance arrangements.","marker":"[8]"},{"why":"The replicator equation that governs evolutionary competition among institutional rule configurations.","marker":"[40]"},{"why":"Gives tit-for-tat and win-stay/lose-shift as bottom-up rules competing with top-down sanctions.","marker":"[38]"},{"why":"Supplies the inclusive-versus-extractive institutions distinction used to analyze AI governance regimes.","marker":"[21]"}],"fun_headline_variants":["AI governance gets a multi-level Price equation","Nested games show how AI and humans co-evolve","Firms as computations: Price equation for AI governance","Selection and governance co-determine AI outcomes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two cross-agent covariance terms in Equation (2) belong in a Price-equation decomposition of the change in average organizational performance, meaning the covariance of one type's fitness with the other type's performance is a genuine selection term—but the paper does not derive these terms from Equation (1) nor specify how human fitness depends on AI traits and vice versa.","fun_headline_variants_meta":{"raw":{"variants":["AI governance gets a multi-level Price equation","Nested games show how AI and humans co-evolve","Firms as computations: Price equation for AI governance","Selection and governance co-determine AI outcomes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000322,"raw_usage":{"total_tokens":1778,"prompt_tokens":882,"completion_tokens":896,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":835}},"tokens_in":498,"tokens_out":896,"duration_ms":8002,"temperature":1.0,"reasoning_tokens":835,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:20:22.328911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a simulated or real organization, measure each human and AI agent's fitness and performance at two time points, compute the between-group, within-human, within-AI, and weighted cross-agent covariance terms using interaction weights $\\omega_{h,ai}$, and compare their sum with the directly observed change in mean performance; if a systematic residual remains—especially one traceable to within-pair interaction products—the extended decomposition in Equation (2) does not close and the cross-terms are not, by themselves, a valid accounting of organizational change.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the covariance selection mathematics that Equation (1) generalizes to organizational performance."},{"cited_title":"The price equation and the causal analysis of evolu- tionary change","cited_arxiv_id":null,"evidence_quote":"Provides the modern Price-equation form and its causal interpretation, the baseline for the AI extension."},{"cited_title":"Taylor and Leo B","cited_arxiv_id":null,"evidence_quote":"The replicator equation that governs evolutionary competition among institutional rule configurations."},{"cited_title":"A strategy of win-stay, lose-shift that outperforms tit-for-tat in the prisoner’s dilemma game","cited_arxiv_id":null,"evidence_quote":"Gives tit-for-tat and win-stay/lose-shift as bottom-up rules competing with top-down sanctions."}],"review_version":1}