Chicago Schelling — a transit intervention that made segregation worse

studies/chicago_schelling/ · reference: true · Path C (wrapped legacy simulator)

Starting from the 2010 census, how does the Black–White dissimilarity index D_bw evolve over three steps when a CTA rail station opens in a South-Side tract at step 2, paired with a fair-housing policy broadcast? The hypothesis was that better transit access draws a more diverse inflow and lowers D_bw. It did not. This case is the pattern for wrapping a validated legacy simulator (a Mesa-style Chicago segregation model) behind the four core interfaces, driving it with real census data, and letting a scheduled intervention plus audience-scoped broadcasts play out on top — with the trajectory recording what happened rather than what was hoped for.

Needs the SocioVerse-ABM checkout

The study's adapter, artifacts and report ship in SocioVerse2; the legacy Chicago model and its census data ship in the companion SocioVerse-ABM repository. Clone it next to SocioVerse2 and install the chicago extra (pip install -e ".[dev,chicago,workbench]"), or point SV_ABM_ROOT / SV_CHICAGO_LEGACY at it — see companion repositories. A no-token dry run uses the adapter's DeterministicLLMClient. The Chicago data carry their own terms, including a non-commercial field; see the SocioVerse-ABM data notice.

P / E / B

	what it is	where it comes from
P	570 persistent households, materialized from 2010 census tracts (chicago.census, scale: small). agent_id encodes race / income bracket / family type / neighborhood type plus the tract GEOID, so every panel row is geo-locatable. Interaction is spatial_adjacency over Queen contiguity; propagation independent.	population/population.json + population/roster.jsonl
E	Four layers on the two axes — see the table below. Provider chicago.env (scale: small, init_mode: census).	environment/environment.json
B	Decision chicago.schelling: each household emits an Action of kind move or stay, carrying ranked_targets. Moves are executed inside the environment's apply, so relocation is an environment mutation, not a bookkeeping trick.	adapter/decision.py

The four environment layers:

layer	modality	scope	dynamics	content
city_race_share	physical	macro	endogenous	city-wide racial shares
tract_local	physical	local	endogenous	per-tract demographics + 1-hop Queen surroundings
news	information	macro	scheduled	citywide policy broadcasts
ward_notice	information	local	scheduled	tract-scoped notices

The intervention is three things firing together at step 2:

• a scheduled event on tract_local — tract 17031612000 (New City, South Side), cta_stations += 2;
• broadcast policy_A_citywide on news, audience: "all", ttl 3 — the station opening plus a citywide fair-housing / anti-displacement program;
• broadcast policy_B_tract on ward_notice, audience = that one GEOID, ttl 3 — the same news addressed to the block.

That pairing is the point of the design: the same event reaches everyone as macro information and reaches the affected tract as local information, so a later version can vary each channel independently.

One step, and the horizon

One step is one relocation round for the whole city: every household sees its tract and its Queen neighbours, decides move or stay, and the moves settle before metrics are computed. n_steps: 3, interaction_rounds: 1, seed: 42 — four metric rows (t=0 baseline plus three decided steps) over 570 tracked households, 2280 panel rows in study.duckdb.

display_metrics are D_black_white, D_hispanic_white, D_asian_white, Isolation_black; the store carries more (isolation and exposure indices, mover diagnostics).

Results

step	D_bw	D_hw	D_aw	Isolation_black	n_movers	pct_pop_moved	mover_alignment_rate
0	0.8160	0.5582	0.3282	0.6137	—	—	—
1	0.8068	0.5730	0.3364	0.6240	41	0.0349	0.220
2 (intervention)	0.8447	0.5891	0.3085	0.6083	50	0.0292	0.280
3	0.8464	0.6112	0.3237	0.6162	31	0.0123	0.387

The hypothesis was falsified. D_bw rose by +0.0304 over the horizon, with the jump concentrated exactly at the intervention step. The mechanism is visible in the panel: in the target tract, the White population went 13 → 0 while the Hispanic population went 0 → 241, with the Black stock nearly unchanged. The accessibility shock lowered the cost of moving, but the direction of moving stayed governed by same-group neighbourhood share — so the nearest majority group, not a returning White population, took the opportunity, and the tract became more non-White than before.

Two secondary readings survive the same data. mover_alignment_rate climbs monotonically (0.220 → 0.280 → 0.387): movers get more like-seeking over time, the classic self-reinforcing tipping signature. And n_movers peaks at the intervention step and then falls to 31, so the system is re-equilibrating after a one-off shock, not being permanently stirred.

This is the interesting result, not a failed run

A study whose numbers contradict its own hypothesis field is working as designed. The artifacts record the falsification, the report explains the mechanism, and the adjustable_params tell you which knob to turn next (the report's own next step: rerun on a tract that starts near the tipping band rather than one already ~100% segregated).

Grounding

grounding/grounding.json carries three sourced facts, three declared assumptions and two implementation references — and the split is exactly where you would want it: the magnitudes are sourced, the scenario choices are declared.

entry	basis	role
Chicago metro Black–White dissimilarity = 76 (0–100), 2010 census — Brown University US2010 (Logan/Stults)	sourced	sanity band for the t=0 D_bw = 0.816
CTA Red Line Extension — 5.5-mile South-Side rail expansion	sourced	the real shock the step-2 event is modeled on
Schelling tolerance threshold 30–50% same-group neighbours	sourced	bounds the decision model's satisfaction regime
intervention tract = 17031612000	assumed	a plausible South-Side site; the real RLE terminates further south
cta_stations += 2 as the station opening	assumed	no public metric maps a station onto this accessibility field
the paired fair-housing / anti-displacement policy text	assumed	stylized, not a specific enacted ordinance

Implementation references: Schelling (1971), Dynamic Models of Segregation — the per-household relocation rule; and a 2025 JAPA quasi-experimental study of transit-induced gentrification — which is precisely why the environment pairs the station with an anti-displacement broadcast, since the literature says the ethnoracial effect can go either way.

resources.json names the two datasets the run actually consumed (chicago_tracts.geojson, cta_stations.geojson) without recording any endpoint or key — see External Capabilities.

What you can change

adjustable_params — read each as "change this to ask that question":

param	the question it opens
n_steps	does D_bw stabilize at the higher level or slowly relax? Three steps cannot say.
seed	how much of the +0.0304 is a single seed's mover draw?
intervention tract & step	put the station in a tract inside the tipping band — does accessibility integrate a mixed neighbourhood even though it cannot integrate a segregated one?
policy broadcast text	the current text promises anti-displacement but gives no one a reason to move in. Add a citywide targeted-attraction message and A/B it.
metrics subset	swap Isolation_black for the exposure indices when the question is contact rather than concentration.

Forking is the reuse path, not editing: this study is reference: true, so sv-init routes a matching question to a fork of it — a new study, not an edit of this one.

Every change here is a version, never a branch

A counterfactual branch replays the parent run's stored actions for the steps before the fork step, and the engine can only do that for a from-scratch study running one interaction round. This study wraps a legacy simulator, so it is not branch-eligible: each change above becomes a new version that re-runs cold from step 0, including the ones that only touch the environment. See Iterate: versions and branches for how a version is snapshotted, and Wrap a Legacy Simulator (Path C) for the seam-and-adapter mechanics.