After this chapter, you should be able to
- Integrate decision, data, baseline, model and assurance evidence.
- Defend validation and uncertainty choices.
- Investigate consequential errors and domain shift.
- Operationalise oversight, abstention, monitoring and rollback.
- Communicate a proportionate accountable recommendation.
Engineering context and methodSource §Lesson 14 · Engineering context and method · GOV-01–08 · ALL-COURSE-SOURCES
Choose a structural/materials, geotechnical, water/environmental, transportation/pavement, construction/vision, geospatial or document-intelligence problem. Submit a decision statement, stakeholder and consequence analysis, baseline, data card, reproducible pipeline, model rationale, independent validation, error slices, uncertainty, five investigated errors, model card, risk register, oversight, abstention, monitoring, rollback and decision memo. Fully autonomous design approval is out of scope.
Verified worked exampleSource §Lesson 14 · Verified worked example · GOV-01–08 · ALL-COURSE-SOURCES
Decision synthesis under weak transfer
Grouped future-test pavement MAE improves from 0.25 to 0.20 m/km, but reaches 0.38 m/km in a climate stratum with only 12 sections.
- Average gain
0.25 − 0.20
0.05 m/km - Critical slice
0.38 m/km with n = 12
weak and uncertain transfer - Scope
route weak stratum to existing assessment
bounded pilot candidate
Result. Do not recommend broad deployment. A controlled pilot may be defensible only in represented strata with stop thresholds and new evidence collection.
Practical lab · 15 h lesson effortSource §Lesson 14 · Practical lab · 15 h lesson effort · GOV-01–08 · ALL-COURSE-SOURCES
- Produce the complete decision dossier and reproducible evidence package.
- Compare against a viable non-AI baseline and justify any added complexity.
- Investigate at least five errors, including the most consequential available case.
- Peer-review another project and defend intended use, split, uncertainty, abstention, monitoring and remaining professional review.
Failure modes to investigateSource §Lesson 14 · Failure modes to investigate · GOV-01–08 · ALL-COURSE-SOURCES
- Accuracy-first story with no decision chain.
- No viable baseline.
- Hidden cleaning or weak test set.
- No adverse error or subgroup analysis.
- Governance pasted from a template and recommendation exceeding evidence.
Knowledge checksSource §Lesson 14 · Knowledge checks · GOV-01–08 · ALL-COURSE-SOURCES
| Question | Answer rationale |
|---|---|
| What is the primary output? | A defensible engineering decision dossier, not merely a trained model. |
| Can average improvement override a failed critical slice? | No; consequence and uncertainty govern scope. |
| What if the baseline is equally useful? | Prefer the simpler auditable solution unless another benefit is justified. |
| Who closes remaining professional issues? | Named competent reviewers and the authorised owner under applicable procedures. |
| When may a course project operate? | Only after separate context-specific validation and authorisation; normally it remains educational. |
Key points
- Start from the accountable engineering decision and its consequence.
- Compare against a transparent non-AI baseline.
- Validate on a split that represents intended use and retain human authority.
Source references recorded by the supplied chapter
- All course primary governance sources and the domain-specific sources selected for the project.
- AI-CIVIL-ENGINEERING-ASSESSMENT-AND-RUBRIC.md.
- AI-CIVIL-ENGINEERING-DATASET-REGISTER.md.
- AI-CIVIL-ENGINEERING-TIERED-PRECISION-VERIFICATION.md.