How mild can a preference for like neighbors be and still split the timeline into conversational and broadcast blocks? The salt-and-pepper start coarsens into solid conversational and broadcast regions within about 15 ticks.
Try this Raise similarity-threshold from 40 to 60 (it takes effect live). Sorting takes longer, about 45 ticks instead of 15, and goes much further: the blocks become nearly pure.
About this model Where it ends up No stop condition, but moves stop once satisfied-count equals occupied-count and segregation-index reads 100: after about 15 ticks at the defaults.
Thomas Schelling's 1971 segregation model, reframed for posting styles: each occupied patch is a conversational or a broadcast slot, and a slot moves when too few of its neighbors are like it. Schelling's point carries over: mild individual preferences can still sort the whole map into blocks. Connecting X sets conversational-share from your reply, quote and repost mix and the pickiness from your interactions per post; no post text is used.
The grid is a 61 × 45 torus; setup fills density percent of patches, conversational-share percent of them conversational. Each tick every slot checks its eight neighbors: it is satisfied if it has none, or if the like share is at least similarity-threshold + engagement-power/10 percent (42.2% at the defaults). Then each unsatisfied slot, one at a time, moves to a randomly chosen empty patch. density and conversational-share apply at setup; the other two sliders are read every tick.
segregation-index is the percent of slots that are satisfied, not a measure of clustering. The average share of like neighbors, which does measure clustering but is not shown, rises from about 51% at the start to about 82% at the defaults.