Scenario foundation: AuraSight

This interactive world is built around the fictional AuraSight scenario described in the papers below. The town routines and experimental social mechanisms are extensions created for this simulation. Please cite the following works when referring to this simulation:

Ng, Lynnette Hui Xian, Bianca N. Y. Kang, and Kathleen M. Carley. AuraSight: Generating Realistic Social Media Data. Technical report CMU-S3D-25-109. Carnegie Mellon University, 2025. arXiv, September 10, 2025. https://arxiv.org/abs/2509.08927.

Ng, Lynnette Hui Xian, and Kathleen M. Carley. “Modeling Social Dynamics with an LLM-Enabled Agent Based Network-Dynamic (LAND) Model.” Preprint, arXiv, August 2, 2026. https://arxiv.org/abs/2608.00929.

Contact, community & contention.

Executed pilot · 7 September 2026 · 96 runs · 8 matched seeds · 96 synthetic residents per run

Research question

When do shared spaces bridge communities, and when do contentious events reinforce divisions? This iteration separates contact opportunities, relationship quality, and opinion convergence. More cross-country contact is not itself evidence of integration. Results below describe this model under stated assumptions; they do not establish how real communities behave.

What was implemented

• The existing geography, 23 destinations, 96 residents, routines, households, festival activities, and separate Ella/Ezekiel identities are retained.
• Local-priority, random, and bridging invitation policies use the same maximum eight invitations per completed venue and the same eligibility and household-diversity rules.
• Opinion updating now has a transparent bounded-confidence comparator and an optional identity/arousal-dependent repulsion extension. Each pair uses both pre-update opinions, eliminating within-pair sequential bias.
• Meaningful reciprocal conversations update signed relationships. Disagreement can strain ties; successful shared activities can offset friction.
• Browser controls restart with chosen parameters, expose a seed, and export the encounter and agent state. Batch experiments use the same model.js as the visual world.

Mechanisms and equations

State and initialization

An agent has opinion x in [-1,1], directed affinity w in [-5,8], home country, household, role, identity, tolerance, arousal, needs and routine. Supporters of Oliver start at +0.65; Ella/Ezekiel supporters and Ethalian nationalists at -0.60; other roles at zero. These are synthetic assumptions, not fitted survey attitudes. Identity is uniform in [0.25,0.90], tolerance in [0.35,0.85]. Initial friendship and spatial routines are inherited from the town prototype.

Exposure to a public event

Chapter selection represents population-wide broadcast exposure, not an endogenous media network. The national-win event adds +0.17 to Oliver supporters, -0.20 times identity to opponents, and +0.06 to initially neutral agents. The statement adds -0.08 to high-identity opponents and +0.22 times openness to others. Court proceedings add +0.09 to supporters and -0.12 to opponents. All are clipped to the opinion range. These response directions are hypotheses imposed by the model.

Event strength multiplies the broadcast opinion shift and the increase in arousal above campaigning levels. At strength zero, the story and event-related travel/dialogue remain; only direct attitude and elevated-arousal effects are ablated. Thus this comparison does NOT estimate the total effect of cancelling the event. Re-selecting the same chapter does not repeatedly add its opinion shock.

Opinion change through encounters

For a reciprocal pair i,j, let d = |x_i - x_j|. With epsilon = 0.70 and mu = 0.08:

If d <= epsilon: x_i_next = clip(x_i + mu * (x_j - x_i), -1, 1)
Otherwise, bounded comparator: x_i_next = x_i
Otherwise, repulsion extension:
  x_i_next = clip(x_i + rho * identity_i * mean_arousal * sign(x_i - x_j), -1, 1)
rho = 0.025

The same calculation updates j using the pre-pair snapshot. Group sessions process unordered pairs in member order; they are not globally synchronous. This is a spatial, encounter-constrained pairwise bounded-confidence variant, not an exact Hegselmann–Krause implementation. The cited HK model supplies an established bounded-confidence reference; a future paper should also compare a faithful synchronous HK benchmark. Repulsion is an additional uncalibrated hypothesis, not something established by the HK citation.

Relationship change

For each reciprocal pair, both directed ties receive:

friction = max(0, d - (0.45 + mean_tolerance)) * (0.6 + mean_arousal)
cooperation = 0.42 if a shared activity occurred, otherwise 0.06
delta_w = clip(cooperation + 0.14 * (1 - d/2) - friction, -0.8, 0.65)
w_next = clip(w + delta_w, -5, 8)

A cancelled gathering or a one-speaker transcript produces no tie or opinion update. A warm tie is w >= 1; a strained tie is w < 0. This threshold is a reporting convention, not an empirically validated friendship scale. Existing affinity and opinion similarity influence later partner selection. Opinions also affect disagreement in rule-based dialogue. Text is generated from templates; it is not semantically interpreted as evidence of persuasion.

How places create contact

Agents must travel over the inherited land/bridge paths and become spatially proximate before speaking. Completed gardens, lantern venues and stages issue invitations. Eligible invitees are non-singer residents without an existing invitation, construction assignment or session, and with energy > 0.25. Each invitation list contains at most one person per household and at most eight people.

Random invitation scores are uniform random. Local-priority scores are distance + 2 × existing degree + 1000 for a foreign-country resident. Bridging scores are distance + 2 × degree - 20 for a foreign-country resident. Lower scores are preferred. The bridging policy rewards geographically reachable outsiders and residents with fewer existing ties; it does not guarantee trust. These policies hold invitation capacity constant, not realized attendance or exposure. Public partner selection favors previously unknown people under every policy.

Scheduling and scope

The simulation advances at fixed dt = 0.5 seconds, representing one simulated minute per tick. A run contains 1,440 ticks (24 hours), starting at 08:10. Needs, walking, routines, conversations, construction, and invitations execute in the inherited order. Four projects are started identically in every run. Food at every venue is floored at 20 units at each update to remove the prototype's local-stock starvation confound; this is an experimental control, not a validated distribution economy.

Experimental design

Full factorial: 3 invitation policies × 2 opinion rules × 2 event strengths × 8 seeds (2030–2037) = 96 runs. Every run uses the national-win chapter. The wider festival storyline is preserved in the UI, but the campaign, statement and legal phases were not independently evaluated in this batch. Each condition begins from the same seed-specific generated population. Treatment-dependent random draws diverge, so paired seeds match initialization, not every later random opportunity.

This is a one-day mechanism pilot without burn-in, calibration or confidence-threshold sensitivity sweeps. The 8 seeds, not thousands of encounters, are the independent replication units. A 24-hour window limits repeated cross-country contact and durable friendship formation.

Results

Values are means over 8 seeds. Cross pairs count reciprocal pair encounters, including repeats; warm cross ties count unique unordered cross-country pairs whose stored affinity is at least 1. Variance is population variance across all 96 opinions at the end of the run. Negative encounters are the percentage of all reciprocal pairs with negative tie increments.

Paired contrasts

These are descriptive mean paired differences with approximate 95% t intervals (7 degrees of freedom; t = 2.365). They quantify Monte Carlo uncertainty under this model only, not uncertainty about its behavioral assumptions. Multiple contrasts are exploratory and not multiplicity-adjusted.

Main findings

1. Targeted bridging did not beat random invitations. With bounded confidence and active event effects, random invitations generated 24.12 cross-country pair encounters per day versus 15.62 for bridging. The paired difference was -8.50 encounters (approximate 95% interval [-12.10, -4.90]). Distance remains part of the bridging score, so the outsider bonus does not guarantee the broad mixing achieved by random invitations. That explanation is a plausible mechanism, not a separately identified causal decomposition.
2. More exposure did not establish durable integration. Warm cross-country ties averaged 0.75 under random invitations and 0.50 under bridging. Their difference was uncertain (interval [-0.84, 0.34] for bridging minus random). The one-day window is too short to infer stable friendship formation or retention.
3. Repulsion increased dispersion under active event effects. With bridging invitations, end-of-run variance rose by 0.0250 relative to bounded confidence (interval [0.0227, 0.0273]). This is a sensitivity result about adding repulsion, not empirical evidence for repulsion in human attitudes.
4. The event ablation changed both dispersion and relationship friction. Under bridging/bounded confidence, variance was 0.4058 with active event effects versus 0.2348 with those effects ablated; negative encounter increments rose from 0% to 15.63%. Direct broadcast responses already impose separation between supporters and opponents. These numbers cannot establish an emergent or real-world causal effect of contentious events.

Interpretation and limits

Read the table as a separation of opportunity and outcome. If a policy produces more cross-country encounters without more warm ties, its effect is exposure rather than demonstrated social integration. One-day contact counts cannot establish durable bridging. Event-induced variance combines an imposed direct broadcast shift with subsequent endogenous contact; it must not be presented as an emergent discovery that contentious events polarize. The repulsion comparison tests sensitivity to a behavioral assumption, not evidence that real people repel.

Local-priority gatherings and mostly domestic routines can mechanically suppress baseline international contact. The random-invitation arm is therefore more informative than local-only comparisons, but none of these arms validates geography or mobility. Country inequality, synthetic initial opinions, inherited role distributions, fixed update order, invitation selection and the food floor may all affect the result. There is no human-data fit, social-media replay, causal estimate about actual countries, or population-scale extrapolation.

Methodological sources

• Hegselmann & Krause (2002), Opinion dynamics and bounded confidence models. The bounded-confidence concept is the reference; this implementation is not the original globally synchronous HK model.
• Grimm et al. (2020), The ODD Protocol for Describing Agent-Based and Other Simulation Models.