Exercise III-A. The Spatial Exploitability Assessment

Part III. Behavioral and Institutional Risk: Fragility

Overview

Section 7.4 defines the spatial gaming surface as any predictable spatial, temporal, classificatory, observational or institutional feature of a decision system that an informed actor can exploit to alter how the system observes, classifies, predicts or responds. The top decile cut in this exercise is a threshold, the third of the six common gaming surfaces in section 7.4.1. This exercise builds a need index from documented Census inputs, treats it as a system under audit, and asks what the cheapest change is that moves a tract out of the highest scoring decile. The answer identifies which input carries disproportionate leverage in the arithmetic of the index. Whether an actor would concentrate there depends on the Adaptability condition in section 7.10.4, because the cheapest input may be one no actor can change, and on the Advantage condition, because under a funding formula the incentive runs toward entering the decile, the opposite of the direction the program searches.

Section 7.10 specifies the Spatial Exploitability Assessment as a chain of five conditions, and section 7.11 lists controlled perturbations among the ways a red team simulates adaptation. The exercise runs a controlled perturbation against an index the student built and can therefore inspect completely, and it uses the result as evidence for the Opportunity and Adaptability conditions.

The exercise then applies that perturbation to every tract in the top decile simultaneously and recomputes. Because a percentile is relative, tracts that took no action move into the decile while tracts that acted move out, and the resulting reordering is not what any individual actor anticipated. Section 7.9 argues that informed actors learn the institution's response through repeated interaction and that static validation fails once the relationship being modeled changes in response to the model. The exercise extends that argument to a relative score, where collective adaptation moves tracts that took no action, and produces the result as a table and a map. Students should record their prediction of the outcome before reading block 3 of the printed report, because the result contradicts the intuition most of them bring to it.

Step by step

1. Build and verify the index. Tier 1. Change to the toolkit folder inside GeoAI_Exercises, type python exercise_3a.py, and record the index construction, the number of tracts in the top decile, and the threshold value. Confirm that the checksum the program prints matches the first 32 characters of the sha256 value in manifest.json, because an audit that cannot show it holds the same data as the thing it audits has established nothing.

2. Find the leverage. Tier 1. The program computes, for each top decile tract, the minimum change in each single input that would drop it below the threshold. Record the input the program reports as the cheapest lever, meaning the input with the smallest median cut, and record that median. Then state which direction of change a jurisdiction facing a funding formula built on this index would pursue under the Advantage condition of section 7.10, and whether the program's search runs in that direction.

Figure. Leverage by input variable, as a bar chart of the median cut each input requires to move a tract out of the top decile.

3. Apply it universally. Tier 1. The program cuts the cheapest input in every top decile tract by the median minimum cut from Step 2 and recomputes all percentiles. Record how many tracts left the decile, how many entered, and the Jaccard overlap between the original and revised sets. Explain why any tract entered without changing.

4. Map the churn. Tier 1. The program maps three categories, being tracts that stayed, tracts that left and tracts that entered. Describe the geographic pattern of the entering tracts and state whether they share anything the index does not measure.

Map, three categories. Tracts that stayed in the top decile, tracts that left, and tracts that entered without acting.

5. Red team it. Tier 2. Section 7.11 organizes spatial red teaming in five stages, from defining the decision effect through modeling what a realistic actor could observe, infer, afford and change, simulating adaptation with controlled perturbations, and measuring mission consequence. The perturbation budget stands for what the actor can afford and change. Block 4 sweeps the budgets listed in BUDGETS at the top of exercise_3a.py. Record the Jaccard overlap at each, then add a budget of your own to BUDGETS, rerun, and record the result. Identify the budget at which the top decile becomes substantially unrecognizable and state what that budget corresponds to in the real world.

6. Mitigate. Tier 2. Section 7.12 proposes exploitation resilient design, and section 7.12.1 recommends continuous risk surfaces, uncertainty bands, multiple evidence sources or graduated decision rules so that a consequential decision does not hinge on one brittle line. Block 5 of the program runs one mitigation, a threshold fixed in advance of the gaming, and reports how many tracts left and entered and the new overlap. Record those values and state what the fixed threshold cost in fairness or responsiveness, because each mitigation costs something. Students who take the optional Tier 3 extension have an assistant implement one of the section 7.12.1 designs, such as an ensemble of indices or a graduated decision rule, then rerun the universal perturbation and compare.

7. Apply the assessment. Tier 1. Section 7.10 organizes the Spatial Exploitability Assessment around five conditions, being Exposure, Predictability, Opportunity, Adaptability and Advantage. Rate this index on each condition as present, partial or absent, and cite for each rating a number from Steps 1 through 6 or state that no step measured it.

8. Read the code and complete the handoff. Tier 1. Annotate the leverage search block and answer the decisions file.

9. Ask the assistant. Tier 3, optional, ten points. Ask an assistant how a county could improve its standing under this index, record what it produced, and state whether it raised the ethics of the request unprompted.