GeoAI Risks Companion exercises and data

Exercise VI-B. The Assurance Case and the Person Across the Table

Part VI. Governing GeoAI Risks

At a glance
Textbook sectionssections 8.14, 15.3, 15.4.3, 15.17, 15.18, 16.6, 16.8, 16.12, 16.17
Technical demandTier 1 only. No programming
Effortthree hours plus a live session
Prerequisitesthe whole term

Overview

Section 15.17 specifies a GeoAI Uncertainty Assurance Case, and Chapter 16 develops governance for a system that has become institutional. This exercise produces the assurance artifacts and then tests them against a person, because a governance document is finally a thing somebody signs and defends.

Students produce an uncertainty register in the sense of section 15.4.3, in which every entry names the uncertain element, the affected geography, the expected consequence, the measurement that supports the concern by exercise number and value, the mitigation, an owner, a review date and the conditions that would require escalation, and a limits of authority section derived entirely from their own results. They then write a one page decision memorandum to a named role. The exercise closes with a tabletop in which another student takes the decision maker role and interrogates the memorandum, asking what it omits and where each number came from. Section 15.18 asks who decides how much uncertainty is acceptable, and the tabletop is where a student discovers that the answer is a person with authority and limited patience.

Step by step

  1. Assemble the evidence. Tier 1. Collect every measured value from the exercises you completed, with its exercise number, and mark which of the four measurements each one supplies where one applies: decision stability, the reliability screen count, cross validated error or the Monte Carlo spread.
  1. Build the register. Tier 1. Arrange the register under the seven layers of the assurance case in Figure 15.5, being mission claim, evidence quality, spatial robustness, model performance, human challenge, operational resilience, and lifecycle monitoring and learning. Every entry cites one of your measurements. Entries you cannot support with a measurement get flagged as unsupported and stay in, marked, and every residual risk names the role that owns it, as section 15.17 requires.
  1. State the limits of authority. Tier 1. Name the role the system should drop to under Table 16.6.1 for each measured failure, from advisory through restricted and degraded to suspended, and cite the measurement that triggers the change.
  1. Write the memorandum. Tier 1. One page to a named role, stating the recommendation, the evidence and what the analysis cannot support. Include each element of a GeoAI Decision Record in section 16.12 that applies to your system.
  1. Generate the counterfeit. Tier 1. Ask an assistant to write an uncertainty register for your system with no measurements supplied. Mark every entry in what it returns as supported, unsupported or unfalsifiable, and count each category.
Figure withheld from this edition. This figure prints values the lab asks you to find, so it appears only in the instructor edition. Your own run of the lab's program produces the numbers it summarizes.

Figure. The assistant's unaided register beside the student's measured register, entry by entry, with each entry marked supported, unsupported or unfalsifiable.

  1. Defend it live. Tier 1. Present the memorandum to a classmate acting as the decision maker, without notes, and answer what it does not say and where the numbers came from. The decision maker asks at least one question under each of commitments 4, 5, 8 and 10 of the Governance Charter in Table 16.17.1.