StandardInternational professional-learning synthesis · ABET 2026–27 · IEA GAPC v4 · ASCE ethics and AI responsibility · NIST AI RMF 1.0
Source3 source files
Review stateTechnical and publication gates pending
After this chapter, you should be able to
- Distinguish retrieval, extraction and generation.
- Verify claims against cited passages and editions.
- Resist prompt injection from untrusted documents.
- Protect confidential and intellectual-property content.
- Design insufficient-evidence and human-reconciliation behaviour.
Engineering context and methodSource §Lesson 11 · Engineering context and method · GOV-07 · GOV-11 · GOV-12 · RES-10 · RES-11
Language models generate plausible sequences, not authoritative engineering truth. Retrieval narrows the evidence but does not guarantee correctness, completeness or current edition. Treat retrieved documents as untrusted data, never as system instructions. Record document ID, edition, page or section and supporting span. A competent person must verify standards values from lawfully accessed current sources.
Verified worked exampleSource §Lesson 11 · Verified worked example · GOV-07 · GOV-11 · GOV-12 · RES-10 · RES-11
Action-extraction quality
A verified record contains 10 actions; the extractor returns 9, of which 8 match.
- Precision
8 / 9
0.889 - Recall
8 / 10
0.800 - F1
2×0.889×0.800/(0.889+0.800)
0.842
Result. One required action was missed and one unsupported action was generated. Human reconciliation remains mandatory.
Practical lab · 5 h lesson effortSource §Lesson 11 · Practical lab · 5 h lesson effort · GOV-07 · GOV-11 · GOV-12 · RES-10 · RES-11
- Extract action, owner, due date and exact evidence from a synthetic meeting record and score against a hidden key.
- Retrieve only from openly licensed, versioned documents and cite the supporting passage.
- Return insufficient evidence when support is absent.
- Ignore a malicious embedded instruction and log queries/sources without confidential content.
Failure modes to investigateSource §Lesson 11 · Failure modes to investigate · GOV-07 · GOV-11 · GOV-12 · RES-10 · RES-11
- Invented clause or unsupported action.
- Wrong standard or document edition.
- Citation does not support the claim.
- Prompt injection changes the workflow.
- Confidential upload or generated record silently committed.
Knowledge checksSource §Lesson 11 · Knowledge checks · GOV-07 · GOV-11 · GOV-12 · RES-10 · RES-11
| Question | Answer rationale |
|---|---|
| Does retrieval remove hallucination? | No; it improves grounding but output still needs verification. |
| What is prompt injection? | Untrusted content attempts to alter authorised system behaviour. |
| What if evidence is absent? | Return insufficient evidence and escalate. |
| Can model memory supply a design clause? | No; verify the lawful, current source. |
| Why record edition and page? | Requirements and meaning change by version and context. |
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
- NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative AI Profile.
- OECD AI Principles.
- UNESCO AI competency framework for students.
- Peer-reviewed civil-engineering LLM reviews listed as RES-10 and RES-11.