Critical appraisal — claims, evidence, blindspots

1. Claim: Bayesian models are often unfalsifiable in practice. Evidence: full-text documents iterative post-hoc adjustments and offers Box 1 rebuttals to common defenses, arguing these moves immunize the theory Unfalsifiability evidence (Box1 & Fig2).
2. Claim: Biological & neurophysiological implausibility (metabolic cost, representational format, dendritic nonlinearity). Evidence and analysis are presented with citations and mechanistic arguments in the text Neurophysiological constraints. Comment: these are compelling concerns — they frame specific implementational gaps rather than merely rhetorical objections.
3. Claim: Bayesian frameworks retain descriptive utility but should not be conflated with mechanistic implementation. The author explicitly separates 'as-if' descriptions from implementational claims and warns against reification of mathematical constructs (priors, posteriors) As-if vs implementation.
4. Alternatives recommended: dynamical systems, ecological psychology, embodiment, and constraint-based models are proposed as biologically grounded frameworks; the paper argues they better match non-Gaussian dynamics, multistability, and real-time coordination demands Alternatives to Bayesianism.
5. Institutional & sociological critique: the manuscript persuasively argues that institutional incentives, cross-disciplinary charisma, and mathematical prestige have amplified Bayesian frameworks beyond their empirical support, making pluralism an epistemic need Sociology of Bayesianism.

Limits, counterpoints, and blindspots

• The paper is primarily conceptual and normative; it intentionally avoids generating new datasets ("no new data created") so empirical claims depend on interpretation of existing literature rather than fresh experiments No new data (limitation).
• While the paper persuasively enumerates mechanistic gaps, it occasionally treats diverse Bayesian models as a monolith (predictive coding, probabilistic population codes, FEP) — some sub-variants already contain explicit, testable mechanism proposals (sampling-based approximations, PPCs, etc.). The author acknowledges these but argues they remain underconstrained Acknowledged Bayesian approximations.
• The recommendation to shift to dynamical/embodied frameworks is well-motivated but risks underestimating the explanatory power Bayesian models have offered in constrained tasks; a pluralist bridge (using Bayesian descriptions at the computational level while requiring mechanistic mapping) is a pragmatic path forward — the paper does recommend pluralism but is strongly normative about abandoning Bayesian centrality Final recommendations (pluralism).

Concrete experiments the paper prescribes (operational)

1. Pre-register an animal or human experiment with explicit priors (e.g., cue with 80/20 probability) and abrupt likelihood changes; record multi-site spiking/LFP or high-density arrays in putative prior-encoding and error-encoding regions; specify exact parametric neural shifts predicted by the Bayesian model a priori; failure to observe those shifts falsifies the strong claim Crucial experiment (Box 2).
2. Test precision-weighting hypothesis by pharmacologically or optogenetically manipulating neuromodulators (e.g., noradrenaline, acetylcholine) while monitoring putative precision-encoding gain changes in sensory cortices under controlled expectation violations; compare predicted monotonic gain changes to observed dynamics Precision-weighting tests.

How to use this paper (practical advice)

• Modelers: convert current Bayesian/PP models into pre-registered, mechanistic predictions at the algorithmic and circuit levels (identify neurons/populations, timecourse, expected effect sizes).
• Experimentalists: design parametric manipulations of priors/likelihoods, preregister analyses, and measure neural signals with sufficient temporal & spatial resolution (multi-electrode, Neuropixels, 2P imaging + optogenetic control where possible).
• Theorists: produce constrained approximations (explicit sampling algorithms mapped to spiking circuits) and produce risky predictions with quantitative effect sizes and timescales.