{"id":"7d8270e6-2eb1-4d86-9165-4b7f0afcf33d","arxiv_id":"2607.20876","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Pedestrians steer around other people's conversations—treating social interaction as an invisible wall—but will walk through once others have already breached.","lead":"Across four field experiments with 4,911 passersby, actors staged conversations, phone-checking, sign-reading, and standing, and the authors counted who walked between them or went around. Pedestrians avoided walking between people who faced, looked at, and talked to each other, but were more likely to cut through after someone else had already done so.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Collective-breach effect may be driven by same-group following rather than contagion to strangers; group membership was coded but not included in the model.","rationale":"I read the paper as aiming to establish that pedestrians infer latent interactional territory from gaze, proximity, orientation, and talk, and that they collectively override these territories after an initial breach. The avoidance part is well supported by four experiments, especially Experiment 4's chairs control, which rules out a purely physical-occupancy explanation. The weakest link is the collective override: it appears in only one observational dataset and relies on a non-randomized exposure count. The reader identified residual confounding by unmeasured time-varying events. I am sharpening that into a specific, checkable confound: group membership. The authors explicitly coded group membership but did not include it in the contagion model. Since group members commonly follow a leader through an opening, the 5-second prior-breach count may be partly endogenous to group coordination. This is the single most load-bearing concern because it directly targets the abstract's 'collectively violate' claim. If the effect is only same-group following, the paper's novelty is reduced, but its core avoidance findings stand. I therefore agree partially with the reader: we converge on the collective-breach inference as the critical assumption, but I single out group membership as the sharpest and most actionable test. The proposed test is a reanalysis of already-coded data, so it can be done without new data collection. The verdict remains conditional; I do not see grounds to reject or accept outright.","tokens_in":16310,"tokens_out":4498,"duration_ms":51044,"concrete_test":"Re-fit the Experiment 2 contagion GLMM with separate predictors: (a) number of breaches in the prior 5 seconds by pedestrians not in the target's coded group, and (b) number of breaches by groupmates (or a binary indicator for a groupmate breach). Alternatively, restrict the sample to solo pedestrians with no groupmates in the 5-second window (Group ID = NA and group size = 1) and estimate the prior-breach OR. If the non-groupmate coefficient is near 1 or the solo-only OR drops to ~1, the collective-breach claim reduces to group coordination; if it remains near 2.7, the contagion interpretation is supported. The authors already coded Group ID and group lead, so this is a reanalysis of existing data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main claim that 'pedestrians also collectively violate these spatial norms by walking through an interaction if other pedestrians had already done so' rests on the Experiment 2 GLMM (Supplementary, 'Calculating predictors and controls for collective proxemic norm breaching') using a 5-second prior-breach count. The Supplementary reports that coders recorded group membership, group size, and group lead for every pedestrian (Table S2), but the contagion model as described includes only prior breaches, square-root crowding, and random intercepts for date and 10-minute interval. This creates a direct confound: if a pedestrian is in a group whose earlier member walked through the actors, the 'prior breach' predictor is 1, and the follower's breach is explained by within-group coordination—not by exposure of independent pedestrians to a norm violation. The Fig. 5 anecdote actually begins with a group of three pedestrians breaching, and the subsequent 'remaining pedestrians' may include groupmates. The reported OR ≈ 2.77 could therefore overstate social contagion to unrelated pedestrians. This does not threaten the avoidance findings, which have multiple controls, but it is load-bearing for the collective-override component of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports four field experiments (total N=4,911) in which pairs of actors or actor-object displays were placed on busy pathways, and the outcome was whether pedestrians walked between the display ('breach') or around it. Experiment 1 manipulated mutual gaze, talk, interpersonal distance, and actor gender; Experiment 2 manipulated mutual body orientation and measured continuous trajectories; Experiment 3 contrasted a person attending to an informational sign versus an art mural; Experiment 4 contrasted two actors with two chairs and an empty baseline. The main avoidance effects are strong: pedestrians are much less likely to walk between actors displaying mutual gaze, talk, close spacing, and face-to-face orientation, and the chair/sign controls argue against a purely physical-obstacle explanation. The paper also reports a collective-norm-breaching effect in Experiment 2: a pedestrian was about 2.77 times more likely to breach if another pedestrian had breached in the preceding 5 s. The authors interpret these results as evidence that pedestrians infer a latent 'interactional territory' and that this territory can be collectively overridden.","tokens_in":16617,"tokens_out":9608,"duration_ms":104166,"significance":"If the conclusions hold, this is a useful contribution to pedestrian dynamics and social cognition: it provides naturalistic, experimental evidence that route choice is modulated by inferred social-interactional states, and it offers quantitative effect sizes (odds ratios) that could inform socially aware navigation models. The manuscript has notable strengths: four experiments, counterbalanced schedules, inter-coder reliability checks (κ≈0.84–0.92 in Exp. 2), physical-control conditions (chairs; sign vs. mural), continuous trajectory measurement with reported reprojection error, and public data/code. The main risk is that the headline collective-breach result may reflect within-group coordination rather than contagion among unrelated pedestrians, because group membership was coded but omitted from the contagion model. This is testable and fixable with the deposited data, so I do not see it as fatal to the paper, but it is load-bearing for the abstract's second claim.","major_comments":[{"comment":"The collective-breach GLMM uses the count of all pedestrians who breached in the previous 5 s as the key predictor, but it does not include group membership, group size, or group-lead status, although these were coded for every pedestrian (Table S2). If the prior 'breacher' is in the same group as the target pedestrian, the predictor captures within-group coordination rather than exposure to an independent norm violation. The qualitative example in Fig. 5 begins with a group of three pedestrians breaching, so subsequent breachers may include group members. The reported OR ≈ 2.77 could therefore overstate contagion to unrelated pedestrians. Please re-fit the model with a same-group prior-breach indicator (or cluster by group ID) and report the effect restricted to pedestrians with no groupmate among prior breachers. If the effect vanishes, the collective-override claim in the abstract mus","section":"Results, 'Pedestrians collectively breach interactional territory'; Supplementary, 'Calculating predictors and controls"},{"comment":"The interpersonal-distance effect in Experiment 1 — OR = 0.06 for 5 ft vs. 10 ft — is the largest reported effect, but it is potentially confounded with physical passage width: at 5 ft the gap between the actors is half as wide as at 10 ft. The chair control in Experiment 4 is informative, but it was run only at 5 ft; it does not establish whether a 5 ft vs. 10 ft gap between inert obstacles would also reduce breaching. As written, the abstract's claim that pedestrians 'integrate ... proximity' as a social cue is stronger than the evidence provided. Please add a chair-distance condition or an analysis (e.g., using the trajectory data to model path clearance and energy cost) that separates physical affordance from social inference.","section":"Experiment 1 (Results, Fig. 2C) and Experiment 4 (Results, Fig. 4)"},{"comment":"The 5-second prior-breach window is a free parameter; robustness to window length is reported only in the Supplementary. More importantly, the collective-breach model omits pedestrian walking direction, even though direction interacts with body orientation in the same experiment (Fig. 2D), and crowd-flow direction could be a time-varying confound correlated with both prior breaches and the current breach. Please include direction (and group membership, as above) as fixed effects, or show in a sensitivity analysis that the odds ratio is unchanged.","section":"Supplementary, 'Calculating predictors and controls for collective proxemic norm breaching'"}],"minor_comments":[{"comment":"Typos and formatting: title has 'Invisible w alls'; abstract has 'Keywordspedestrians'; Introduction has 'challenging problem to for machines'; reference to 'Physics Review E' should be 'Physical Review E'.","section":"Throughout"},{"comment":"The anecdote is inconsistent: the main text says 12 of the 19 remaining pedestrians breached after the group of three, while Figure S4 says 15 of 19. Please reconcile the counts.","section":"Results, Fig. 5 vs. Supplementary Fig. S4"},{"comment":"The sign vs. two-actors comparison is reported as OR = 3.57 in the Results and as '3.83 times more likely' in one sentence of the Supplementary. Please check which value is correct.","section":"Results and Supplementary, Experiment 3"},{"comment":"The phrase 'additive interaction' is confusing: an interaction on the log-odds scale is not additive in the ordinary sense. Please rephrase or explain the model scale.","section":"Results and Supplementary, Experiment 3"},{"comment":"Exclusion proportions are large (458/2,319 ≈ 19.8% in Exp. 1; 641/1,515 ≈ 42.3% in Exp. 3). A CONSORT-style flow diagram and a robustness analysis retaining all coded pedestrians would help, since 'never traversed near the actors' may not be condition-independent.","section":"Methods, Experiments 1 and 3"},{"comment":"No pre-registration is mentioned. For an observational field study with multiple models, it would be useful to state which analyses were confirmatory and which were exploratory.","section":"Methods / Statistical analysis"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the avoidance findings are solid and well controlled; the collective-norm result is the main vulnerability. The group-membership confound is readily addressable because the data and code are public, so I would not reject. If the requested reanalysis cannot be provided, the authors should downgrade the second half of the abstract and the 'collective rule-breaking' section. The paper's fit with physics.soc-ph is reasonable if the emphasis is on pedestrian mobility, but the manuscript is largely a cognitive-science/proxemics study; that scope question is secondary to the technical fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Core avoidance findings hold up well. Across four field experiments, the factorial manipulations (gaze, talk, distance, body orientation, object-directed attention) produce large, consistent odds-ratio effects, and the control conditions—chairs vs. actors, sign vs. mural, baseline—do real work. The trajectory measurement and inter-coder reliability are solid, and the data/code release is a plus. This is a genuinely new empirical contribution: prior work described F-formations and personal space, but not systematic navigation around staged interactions.\n\nThe soft spot is the collective-breach claim. The stress-test note is right: group membership and group lead were coded but never entered the contagion model. A follower whose groupmate just walked through is counted as exposed to a prior breach, so the OR of 2.77 could partly be within-group coordination rather than contagion to independent pedestrians. The Fig. 5 anecdote starts with a group of three, so it does not resolve this. This is load-bearing for the abstract's second sentence. It does not touch the avoidance findings; those have multiple controls and stand alone.\n\nOther weaknesses are minor in comparison. Exclusions are large and subjective for Experiments 1 and 3, there is no pre-registration, and the 'social computations' phrasing goes beyond what behavioral data can show. A sensitivity analysis adding group variables, plus a pre-registered replication, would address the main concern.\n\nI'd send this to peer review. The avoidance core is publishable; the contagion claim needs revision or aggressive softening.","headline":"Core avoidance findings hold up well; the collective-breach result is real but the contagion reading is not yet supported because group following was never modeled.","tokens_in":17034,"tokens_out":2201,"would_cite":true,"duration_ms":23072,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["89.65.-s"],"model":"deepseek-v4-flash","headline":"Pedestrians treat other people's conversations as invisible walls — reading gaze, talk, distance, and body orientation to decide whether to pass — and the walls fall once a single pedestrian walks through.","keywords":["pedestrian dynamics","proxemics","social norms","interactional territory","social contagion","involvement signs","field experiments","pedestrian navigation"],"falsifier":"A randomized field test would settle the contagion claim: have trained confederates breach the two actors' interaction at randomly chosen moments on matched days, while never breaching on control days, and compare the breach rate of naive pedestrians in the following ten seconds; if the rate does not rise after confederate breaches, the social-contagion claim collapses. The chair-versus-actor contrast already tests the avoidance claim, so the sharpest unsettled test is the breach-window one.","tokens_in":16261,"feed_emoji":"🚶","tokens_out":7110,"duration_ms":66012,"temperature":0.7,"pith_summary":"Pedestrian traffic is usually modeled as a physical system of moving bodies avoiding collisions, but this paper argues walkers also navigate a social landscape made of invisible walls: the territories created by other people's conversations. Across four field experiments with roughly 4,900 observed pedestrians, the authors show that whether a stranger walks between two people depends on the signs those people display — mutual gaze, speech, close standing, and face-to-face orientation each lower the chance of being walked through. When two people simply stand close facing each other, passersby detour around them almost as reliably as around a physical obstacle, while the same space between two chairs gets crossed nearly every time. The paper's second claim is that the norm is collectively reversible: a pedestrian is nearly three times more likely to walk through an interaction if someone breached it within the previous five seconds, a pattern the authors read as social contagion of rule-breaking.","feed_headline":"Crowds walk through conversations after the first breach","feed_subtitle":"Four field experiments show walkers read gaze, talk, and distance to honor an interaction's invisible walls.","key_machinery":"The central object is interactional territory: an invisible spatial boundary around an interaction that non-participants normally avoid crossing. The mechanism that carries the argument is involvement-sign reading — pedestrians use gaze direction, speech, body orientation, and interpersonal distance as observable evidence of whether an interaction exists and how strong its claim on the space is. The experiments work by manipulating one sign at a time (gaze versus phone, talk versus silence, five versus ten feet, face-to-face versus offset versus back-to-back, two people versus two chairs versus a sign) while holding physical occupancy constant, so changes in breaching rate isolate the social","core_discovery":"On the paper's own terms, the discovery is that pedestrians continuously infer a type of spatial claim the field has mostly ignored: interactional territory, an invisible boundary around an ongoing social interaction that non-participants treat as off-limits. Unlike personal space, whose boundary is attached to a single body, this territory is latent — it exists only in what the interacting people make observable — and pedestrians compute its presence from involvement signs such as where others look, whether they talk, how their bodies face, and how far apart they stand. The authors show each sign independently shifts breaching probability, that mutual gaze saturates the contribution of talk","pith_inferences":["If the contagion effect is causal, the same release mechanism should appear in other proxemic norms such as personal space or queue order; a direct test would compare follower breach rates after an unambiguous confederate breach versus after a near-miss that leaves the territory intact.","The saturation of talk by gaze suggests a threshold model of interaction detection rather than a linear integrator; a testable prediction is that adding further involvement signs, such as touch or sustained joint attention, yields no additional protection once mutual gaze is present.","The mural result implies pedestrians model the affordances of what a person attends to; an unexplored extension is whether pedestrians also distinguish pairs of friends from strangers or task-oriented talk from social chat.","Fitting breach probability as a function of continuous variables — gaze angle, distance, talk duration — in the same field setup would turn the framework into a predictive spatial map of social cost usable for robot navigation."],"forward_implications":["If walkers treat interactional territory as a navigation constraint, then robots, delivery drones, and autonomous vehicles sharing walkways must infer the same signs — gaze, talk, orientation — to avoid socially disruptive paths.","Pedestrian-flow models built purely from collision avoidance will systematically mispredict where people walk, since breaching probability changes by up to a factor of 17 on social signs alone, independent of physical occupancy.","The collective-breach result implies conversational groups create only provisional barriers; crowd managers cannot rely on a conversation group to keep a lane clear once any pedestrian has crossed it.","The sign-versus-mural contrast shows the inference is content-sensitive: pedestrians do not avoid everyone who looks at something, only people engaged in interactions whose space others can disrupt."],"fun_headline_variants":["Pedestrians steer around conversation bubbles — until one breaks","Walkers read gaze and talk to dodge invisible interaction walls","Conversation territory is invisible, but pedestrians find it","One breach makes pedestrian norm vanish: walkers cross chats","How pedestrians keep conversations off-limits — and when they don't"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The collective-breach result assumes the five-second prior-breach count is the cause of followers' behavior, not a proxy for something else — a crowd surge, a gap in the actors' performance, or flow direction — that independently makes later pedestrians more likely to walk through.","fun_headline_variants_meta":{"raw":{"variants":["Pedestrians steer around conversation bubbles — until one breaks","Walkers read gaze and talk to dodge invisible interaction walls","Conversation territory is invisible, but pedestrians find it","One breach makes pedestrian norm vanish: walkers cross chats","How pedestrians keep conversations off-limits — and when they don't"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1300,"prompt_tokens":626,"completion_tokens":674,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":370,"completion_tokens_details":{"reasoning_tokens":591}},"tokens_in":370,"tokens_out":674,"duration_ms":7253,"temperature":1.0,"reasoning_tokens":591,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:05:55.591657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A randomized field test would settle the contagion claim: have trained confederates breach the two actors' interaction at randomly chosen moments on matched days, while never breaching on control days, and compare the breach rate of naive pedestrians in the following ten seconds; if the rate does not rise after confederate breaches, the social-contagion claim collapses. The chair-versus-actor contrast already tests the avoidance claim, so the sharpest unsettled test is the breach-window one.","supporting_citations":[],"review_version":1}