Paper Review — Claim-Level

Paper review: “A Matter of Time: Towards a General Theory of Agency”

Source: arXiv:2606.23122 (Paper itself)

• Problem: “Agency” is used across domains without a clear account of how it originates from material organization; the paper aims to supply a principled, graded theory grounded in closure, temporality, and control. (Problem framing)
• Main construction: Starting from temporally parameterized (F,A)-systems, it argues that temporal articulation of semantic closure yields a history-dependent Asynchronous Dynamic Bayesian Network (ADBN) as a formal redescription (not an ontological replacement). (ADBN redescription claim)
• Empirical posture: It proposes operational handles/metrics (SCI, MCC, AMI, ARR, SO, VCSA) with stated failure modes to make agency testable rather than purely definitional. (Metric suite purpose and falsification logic)

The paper’s core mechanistic story has three linked moves:

1. Closure equivalence claim: It treats closure to efficient causation, organizational closure, and semantic closure as convergent descriptions of one self-producing organizational property. (Closure convergence argument)
2. Temporality as intelligibility: It argues that self-reference cannot be adequately represented outside temporal unfolding; time is internal to the operational meaning of closure. (Time/self-reference claim)
3. ADBN as derived redescription: Given characteristic timescales for constitutive mappings, the induced cross-time dependency structure can be represented as a history-dependent ADBN/joint distribution. (ADBN construction steps)

• Attempts to separate concepts cleanly (autonomy vs goal-directedness vs agency vs open-endedness) rather than treating “agency” as a catch-all. (Concept separation intent)
• Non-trivial formal move: it proposes a specific redescription target (ADBN) derived from temporalized closure rather than treating Bayesian/Markov/active-inference language as first principles. (Restriction of inferential tools)
• Explicit falsification framing for multiple claimed distinctions (e.g., fixed graph parameter tuning vs graph reorganization; perturbing internal anticipatory variables Mt; internal interpreter maintenance vs externally supplied). (Falsification logic overview)

I’m going to flag issues the paper either (i) explicitly acknowledges or (ii) leaves open in ways that could break empirical traction.

4.1 Topology-changing ADBNs and open-endedness formalization

• The theory claims open-endedness requires more than probability updates within a fixed dependency graph, but the mathematical formalism for topology-changing ADBNs is not fully developed in the manuscript. (Open-endedness formal gap acknowledged)

4.2 Perspectival measurement: “internal” vs “external” can become theory-dependent

• The paper emphasizes that operationalizing closure requires selecting observables/time scales/variables and thus is perspectival; it claims this doesn’t make the theory arbitrary, but the risk is that “closure” and “interpreters” could become post-hoc labels for what the experimenter chose to observe. (Perspectival realism stated)

4.3 Empirical calibration vs conceptual metrics

• The metrics are introduced as operational handles but explicitly not as fully validated or complete experimental apparatus; this means reproducibility and discriminative power remain uncertain until calibration/benchmarking happens. (Metrics are preliminary)

4.4 Proof-of-concept breadth vs risk of overextension

• The proof-of-concept section lists heterogeneous domains (abiotic vesicles, Physarum, xenobots, multicellularity, neural systems) but—per the paper’s own stance—does not provide a completed empirical measurement pipeline for each domain. That makes it harder to tell whether the formal distinctions truly discriminate regimes rather than being accommodating to many things. (Stress tests not confirmations)

I restate the paper’s own falsification logic, then add a skeptical interpretive check.

• If systems labeled “strongly agentive” behave like conventional regulators—i.e., only parameter tuning within a fixed measurement–control graph—then endogenous anticipation would be undermined. (Falsification: graph invariance)
• If intervention that targets internally generated anticipatory variables does not impair adaptive action beyond what external feedback explains, AMI’s interpretation fails. (Falsification: AMI intervention)
• If “candidate interpreters” are externally supplied rather than internally produced/maintained, SCI would not support semantic closure. (Falsification: SCI internality)
• For open-endedness: if apparent affordance reconstruction does not preserve viability, it should be interpreted as instability rather than open-ended agency. (Falsification: ARR viability correction)

The paper is mostly theoretical and provides formulas/metric definitions, but it explicitly reports no additional data and acknowledges remaining formal gaps; therefore reproducibility is primarily about re-running the derivations/operational definitions and later empirical calibration, not about reproducing completed datasets. (Data availability & reproducibility context)

• Start with a target organization (the paper’s “O” and variable sets: environmental/internal/sensor/effector/viability). (Operational measurement template)
• Use intervention to separate internal maintenance from external coincidence—SCI/MCC/AMI logic depends on “does the organization regenerate/maintain enablement relations under perturbation?”. (Intervention-based operationalization)
• Treat ADBN language as derived: don’t assume Markov blankets/active inference are the causal foundation; use them as restricted redescriptions after organizational closure is hypothesized/intervened upon. (Closure-first methodological restriction)

• Formalize topology-changing ADBNs with a minimally complete mathematical framework (even partial) that makes “open-endedness = topology/variable/alphabet/role changes under viability preservation” mathematically checkable. (Need: topology-changing ADBN formalism)
• Provide at least one fully worked measurement protocol for one metric pair/triple (e.g., SCI+AMI) in one domain to demonstrate discriminative power, not just definitions. (Metrics need calibration/verification)
• Clarify identifiability under perspectival observables: specify minimal conditions under which “internal interpreters” are not observational artifacts of the chosen variable set. (Identifiability via intervention but still open)

Author reviews (BGPT)

Explore additional critiques or clarifications for each author.