Modeling AI-TPACK in Practice Insights from Teachers Multi-Agent Workflow Design

Referee: [Methods] Methods section: the cluster and Markov analyses are presented without specifying the algorithm (k-means, hierarchical, etc.), distance metric, cluster-number selection procedure (elbow, silhouette, etc.), or validation steps such as cross-validation or stability checks. These omissions are load-bearing because the three archetypes are the foundation for all subsequent claims about cognitive underpinnings and dynamic emergence.

Authors: We agree that the Methods section requires explicit specification for reproducibility. In the revised manuscript we will state that k-means clustering with Euclidean distance was used, that the number of clusters (k=3) was selected via the elbow method supplemented by silhouette scores, and that cluster stability was validated through 100 bootstrap resamples and multiple random initializations. For the Markov analysis we will describe the maximum-likelihood estimation of transition probabilities from the sequence data and the use of additive smoothing. These details were part of the original analysis pipeline but were omitted to meet length constraints; we will expand the section accordingly. revision: yes

Referee: [Results] Results, interview subsection: the inference of a 'dynamic interplay of systems thinking, pedagogical beliefs, and self-efficacy' rests on n=12 interviews. No details are given on coding reliability (e.g., inter-rater kappa), theme saturation, or explicit triangulation between logs and interview transcripts. This small, likely convenience subsample limits support for generalizable cognitive mechanisms.

Authors: We acknowledge the modest size of the interview subsample as a genuine limitation. In the revision we will report that two coders independently analyzed the transcripts, achieving Cohen’s kappa = 0.83, and that thematic saturation was reached after the tenth interview. We will also describe the explicit triangulation procedure in which log-derived behavioral metrics (e.g., iteration frequency, workflow complexity) were mapped onto corresponding interview excerpts. While we cannot enlarge the sample retrospectively, we will add an explicit discussion of the convenience-sampling constraint and its implications for generalizability. revision: partial

Referee: [Discussion] Discussion: the central claim that integration 'emerges from a dynamic interplay ... not merely from the possession of discrete knowledge' is interpretive. A clearer contrast—e.g., a baseline measure or direct comparison against a discrete-knowledge TPACK model—would be needed to substantiate the 'not merely' distinction.

Authors: We agree that the contrast can be made more concrete. In the revised Discussion we will add a paragraph that reports participants’ pre-study self-assessed TPACK component scores (collected via a validated instrument) and shows that these static scores did not reliably predict archetype membership, whereas the behavioral and interview data revealed integration patterns that diverge from what discrete-component models would predict. This comparison will be framed against prior TPACK literature that emphasizes additive rather than emergent knowledge structures. revision: yes