Course 25 | Module 12 of 12

Capstone Integration and Engineering Decision Defense

Assemble a small but complete workflow from stakeholder need through evidence, AI-assisted review, change impact, and a defended decision.

MAP

Module map

Learning outcomes

  • Scope a tractable mechanical subsystem and define its decision and context of use.
  • Integrate requirements, architecture, model, simulation data, measurement, uncertainty, surrogate, and optimization.
  • Build an evidence map and perform AI-assisted review with manual verification.
  • Make and orally defend a bounded recommendation, limitations, V&V plan, and change-impact analysis.

Evidence standard

Complete all four lessons, reproduce the worked checks, run the lab, and correct the weekly quiz. Treat AI output as candidate evidence until independently verified.

12.1

Capstone scoping, decision, and architecture baseline

Why this lesson matters

An oversized capstone becomes a collection of shallow artifacts. A strong project narrows one consequential decision while preserving the full evidence chain.

Learning objectives

  • Define and distinguish Question of interest and Project boundary.
  • Apply the lesson method to the worked capstone scoping, decision, and architecture baseline case.
  • Evaluate evidence, uncertainty, and AI-assisted output before making a claim.

Readiness check

Before continuing, explain what decision this topic supports and name one upstream source that must be controlled.

Check your response

A sound answer names a specific engineering decision, its configuration, and a controlled requirement, model, dataset, interface, or standard that constrains the work.

Core idea

Choose a subsystem with two to five meaningful design variables, one analytical or numerical model, obtainable test or synthetic data, explicit uncertainty, and a decision that can change based on evidence.

Key concepts

Question of interestThe specific engineering question the project must answer.
Project boundaryThe declared subsystem, interfaces, lifecycle stages, and exclusions.
Decision gateThe point and criteria at which evidence supports proceed, revise, test, or stop.
Evidence planThe required artifacts, checks, owners, and timing for the decision.

Step-by-step explanation

  1. Select a subsystem and one primary decision.
  2. Define stakeholders, operating scenarios, interfaces, hazards, and exclusions.
  3. Write six to twelve controlled requirements and verification intent.
  4. Create functional and physical architecture with configuration identity.
  5. Plan model, data, uncertainty, surrogate, optimization, AI review, and decision evidence before implementation.

Worked example

Capstone question: choose thickness and rib height for an aluminum electronics bracket that meets mass, static deflection, first-mode, and manufacturability requirements under uncertain load and material properties.

  1. 1

    Boundary: bracket and its bolted interfaces, excluding redesign of the electronics box and vehicle frame.

  2. 2

    Decision: release one geometry for prototype test or request more analysis.

  3. 3

    Variables: thickness and rib height; outputs: mass, deflection, stress, frequency.

  4. 4

    Evidence plan: hand calculation, small simulation dataset, synthetic or measured deflection data, UQ, GP surrogate, constrained trade study, trace graph, and AI review.

Result. The scope is small enough for depth but contains every required evidence relationship.

Independent check. Every artifact supports the primary decision; optional features that do not change the decision are removed.

Common misconceptions

MisconceptionCorrection
A capstone must be industrial scaleA small, rigorous, reproducible evidence chain teaches more than a broad unverified architecture.
A tool output closes the questionA result remains a candidate until its inputs, method, configuration, uncertainty, and relevance have been checked.

Diagnostic questions

How many models are required?

Enough to support and challenge the decision, including at least one analytical or numerical model and a surrogate; complexity is not the goal.

What would make this work reproducible?

Controlled inputs, method or code, versions, assumptions, outputs, and a stated interpretation tied to the decision.

Practice ladder

Basic

Reduce a full vehicle digital twin proposal to one testable subsystem decision.

Intermediate

Write a capstone charter with boundaries, variables, outputs, evidence, and exclusions.

Advanced

Design decision gates that can stop work when model credibility or data quality is insufficient.

AI-assisted engineering task

Ask AI to critique scope creep and missing decision dependencies, not to invent requirements.

How to prove the AI output yourself

  1. Trace every requirement to an authoritative project assumption or stated teaching allocation.
  2. Have the instructor approve scope and data feasibility.
  3. Verify that the final decision can change under plausible results.

Retrieval and spaced review

Answer closed-notes today, then again after 1, 3, 7, and 30 days.

Define Question of interest.

The specific engineering question the project must answer.

What role does Project boundary play here?

The declared subsystem, interfaces, lifecycle stages, and exclusions.

What must a reviewer be able to reconstruct?

Every artifact supports the primary decision; optional features that do not change the decision are removed.

End-of-lesson summary

Choose a subsystem with two to five meaningful design variables, one analytical or numerical model, obtainable test or synthetic data, explicit uncertainty, and a decision that can change based on evidence.

Student notes

Freeze a one-page charter before building models. Record every scope change and why it improved the decision evidence.

Recommended readings

Instructor notes

Reject projects with no real decision or with unavailable validation evidence unless a transparent synthetic-data teaching assumption is justified.

12.2

Integrating model, data, uncertainty, surrogate, and optimization

Why this lesson matters

Capstone quality depends on consistent configurations and quantities across artifacts, not on each analysis looking polished in isolation.

Learning objectives

  • Define and distinguish Analytical baseline and Simulation dataset.
  • Apply the lesson method to the worked integrating model, data, uncertainty, surrogate, and optimization case.
  • Evaluate evidence, uncertainty, and AI-assisted output before making a claim.

Readiness check

Before continuing, explain what decision this topic supports and name one upstream source that must be controlled.

Check your response

A sound answer names a specific engineering decision, its configuration, and a controlled requirement, model, dataset, interface, or standard that constrains the work.

Core idea

Build the workflow in increasing fidelity: analytical baseline, verified computation, simulation dataset, measurement comparison, uncertainty propagation, validated surrogate, and constrained optimization. Carry identifiers, units, assumptions, and configuration through every transformation.

Key concepts

Analytical baselineA transparent low-order model used for scale, limits, and independent checking.
Simulation datasetControlled input-output runs with model and solver provenance.
Synthetic test dataArtificial observations used only under explicit teaching assumptions and never misrepresented as measurement.
Integrated workflowA reproducible sequence whose outputs and evidence links preserve identity and meaning.

Step-by-step explanation

  1. Implement and test the analytical baseline.
  2. Generate source simulations with verification status and a designed input set.
  3. Acquire measurement or declare synthetic data generation and uncertainty.
  4. Calibrate only when justified, then validate on independent cases.
  5. Train the surrogate, evaluate decision metrics, optimize, and confirm candidates with source models.

Worked example

The bracket analytical model predicts 2.06 mm deflection. A higher-fidelity model predicts 2.03 mm, and a calibrated test reports 2.12 +/- 0.07 mm expanded uncertainty under matched load.

  1. 1

    Differences are analytical-test -0.06 mm and FEA-test -0.09 mm using prediction minus measurement.

  2. 2

    Check whether expanded uncertainty basis and numerical uncertainty permit a meaningful comparison.

  3. 3

    Use all three as complementary evidence, not an average truth value.

  4. 4

    Train a surrogate only on controlled higher-fidelity runs and confirm any optimum with the source model.

Result. The full chain provides scale check, numerical prediction, physical comparison, uncertainty, and an efficient trade model with separate roles.

Independent check. All results share load, geometry, material, coordinate, quantity definition, and configuration or explicitly document differences.

Common misconceptions

MisconceptionCorrection
Agreement among three models proves realityModels may share assumptions and inputs; independent measurement and uncertainty are needed to challenge them.
A tool output closes the questionA result remains a candidate until its inputs, method, configuration, uncertainty, and relevance have been checked.

Diagnostic questions

When may synthetic data be used?

For transparent teaching and workflow testing when generation model, noise, and limitations are declared.

What would make this work reproducible?

Controlled inputs, method or code, versions, assumptions, outputs, and a stated interpretation tied to the decision.

Practice ladder

Basic

Create a configuration table aligning analytical, FEA, test, and surrogate artifacts.

Intermediate

Define independent validation cases and decision-critical metrics.

Advanced

Resolve a case where the surrogate optimum passes but source-model confirmation fails.

AI-assisted engineering task

Ask AI to check configuration tables and identify inconsistent units, IDs, or conditions.

How to prove the AI output yourself

  1. Recompute analytical values.
  2. Inspect source-model verification.
  3. Reproduce data processing.
  4. Confirm surrogate candidates at higher fidelity.

Retrieval and spaced review

Answer closed-notes today, then again after 1, 3, 7, and 30 days.

Define Analytical baseline.

A transparent low-order model used for scale, limits, and independent checking.

What role does Simulation dataset play here?

Controlled input-output runs with model and solver provenance.

What must a reviewer be able to reconstruct?

All results share load, geometry, material, coordinate, quantity definition, and configuration or explicitly document differences.

End-of-lesson summary

Build the workflow in increasing fidelity: analytical baseline, verified computation, simulation dataset, measurement comparison, uncertainty propagation, validated surrogate, and constrained optimization. Carry identifiers, units, assumptions, and configuration through every transformation.

Student notes

Maintain one machine-readable manifest listing every artifact ID, version, checksum or path, input source, output, and status.

Recommended readings

Instructor notes

Require students to label synthetic evidence prominently. It can test a method but cannot claim real-world validation.

12.3

Evidence map, AI-assisted review, and change impact

Why this lesson matters

Integration becomes visible when a change can be traced to exactly the analyses, tests, claims, and decisions that need review.

Learning objectives

  • Define and distinguish Evidence map and Impact path.
  • Apply the lesson method to the worked evidence map, ai-assisted review, and change impact case.
  • Evaluate evidence, uncertainty, and AI-assisted output before making a claim.

Readiness check

Before continuing, explain what decision this topic supports and name one upstream source that must be controlled.

Check your response

A sound answer names a specific engineering decision, its configuration, and a controlled requirement, model, dataset, interface, or standard that constrains the work.

Core idea

Build an evidence graph with typed, versioned links and validity conditions. Use AI to suggest gaps or impacts only in candidate state. Verify every accepted suggestion and record reviewer disposition.

Key concepts

Evidence mapA graph or structured view of claims, artifacts, links, status, and configuration.
Impact pathA traversable route from a changed artifact to potentially affected evidence or decisions.
Review dispositionAccepted, rejected, modified, unaffected, rework, or obsolete status assigned by an accountable reviewer.
Audit completenessAbility to reconstruct inputs, actions, checks, decisions, and changes.

Step-by-step explanation

  1. Create nodes for needs, requirements, architecture, models, runs, data, UQ, surrogate, optimization, AI records, and decision.
  2. Add directional typed links with configuration and validity conditions.
  3. Run orphan, coverage, status, and change-impact queries.
  4. Use AI to rank candidate missing links or impacts with path rationale.
  5. Manually verify and preserve both accepted and rejected suggestions.

Worked example

The bracket load requirement changes from 2.0 to 2.4 kN after the surrogate and optimum were produced.

  1. 1

    Traverse from requirement to analytical and FEA inputs, simulation dataset, surrogate, optimization constraints, confirmation run, test plan, compliance assessment, and release decision.

  2. 2

    Mark these nodes suspect, not automatically invalid.

  3. 3

    Screen scale relations: linear stress and deflection suggest likely margin erosion, but frequency may be unaffected by static load under the chosen model.

  4. 4

    Rerun required analyses, update evidence, review AI-proposed impacts, and issue a new decision revision.

Result. The graph focuses rework and preserves why the prior decision was reasonable under the old baseline.

Independent check. Every affected or unaffected disposition has path, physics rationale, reviewer, date, and resulting artifact revision.

Common misconceptions

MisconceptionCorrection
Change impact is a text searchWord matches miss structural and physical dependencies and produce irrelevant results.
A tool output closes the questionA result remains a candidate until its inputs, method, configuration, uncertainty, and relevance have been checked.

Diagnostic questions

Should the old decision be deleted?

No. Supersede it while preserving its baseline, evidence, and rationale.

What would make this work reproducible?

Controlled inputs, method or code, versions, assumptions, outputs, and a stated interpretation tied to the decision.

Practice ladder

Basic

Draw the downstream impact path for the load change.

Intermediate

Separate structurally reachable nodes from physically affected nodes.

Advanced

Design an AI impact assistant evaluation with injected changes and known affected sets.

AI-assisted engineering task

Ask AI to propose impact paths and missing evidence using only the graph export, with exact node IDs and an abstain option.

How to prove the AI output yourself

  1. Traverse graph paths.
  2. Check physical causality.
  3. Review configuration and validity conditions.
  4. Record human disposition and rerun evidence.

Retrieval and spaced review

Answer closed-notes today, then again after 1, 3, 7, and 30 days.

Define Evidence map.

A graph or structured view of claims, artifacts, links, status, and configuration.

What role does Impact path play here?

A traversable route from a changed artifact to potentially affected evidence or decisions.

What must a reviewer be able to reconstruct?

Every affected or unaffected disposition has path, physics rationale, reviewer, date, and resulting artifact revision.

End-of-lesson summary

Build an evidence graph with typed, versioned links and validity conditions. Use AI to suggest gaps or impacts only in candidate state. Verify every accepted suggestion and record reviewer disposition.

Student notes

For every change, preserve before, changed item, suspected set, reviewed set, rerun evidence, decision revision, and residual limitations.

Recommended readings

Instructor notes

Score students on rejected suggestions too. Critical review is part of the deliverable, not evidence of AI failure alone.

12.4

Final recommendation, limitations, oral defense, and full evidence chain

Why this lesson matters

A technically correct analysis can still produce a weak decision if assumptions, alternatives, uncertainty, and contrary evidence are hidden.

Learning objectives

  • Define and distinguish Decision recommendation and Limitation.
  • Apply the lesson method to the worked final recommendation, limitations, oral defense, and full evidence chain case.
  • Evaluate evidence, uncertainty, and AI-assisted output before making a claim.

Readiness check

Before continuing, explain what decision this topic supports and name one upstream source that must be controlled.

Check your response

A sound answer names a specific engineering decision, its configuration, and a controlled requirement, model, dataset, interface, or standard that constrains the work.

Core idea

The final recommendation must state the decision, configuration, evidence basis, uncertainty, credibility, alternatives, limitations, verification and validation plan, change triggers, and accountable approvals. Oral defense tests whether the student understands the chain rather than merely assembling artifacts.

Key concepts

Decision recommendationA bounded proposed action justified by evidence and compared alternatives.
LimitationA condition, omission, uncertainty, or domain boundary that restricts interpretation.
Residual riskRisk remaining after planned controls and evidence.
Oral defenseStructured questioning used to test reasoning, ownership, and ability to navigate evidence.

Step-by-step explanation

  1. Restate question, configuration, and decision criteria.
  2. Summarize alternatives and why the recommendation dominates for the stated priorities.
  3. Present evidence chain with VVUQ and measurement credibility.
  4. Disclose limitations, contrary evidence, unresolved risk, and monitoring or test plan.
  5. Define change triggers and answer defense questions by navigating source artifacts.

Worked example

Full bracket chain recommends design BRK-C: mass 0.46 kg; analytical deflection 2.06 mm; FEA 2.03 mm; test 2.12 +/- 0.07 mm expanded; requirement 2.50 mm; GP trade study explored thickness and rib height; source-model confirmation passed; AI trace review found one corrected stale link.

  1. 1

    State that nominal and measured deflection evidence support margin to 2.50 mm for the tested configuration and load.

  2. 2

    Do not subtract expanded uncertainty blindly without its basis; report the uncertainty and evidence assumptions explicitly.

  3. 3

    Present stress, frequency, manufacturing, and load-uncertainty evidence before claiming total compliance.

  4. 4

    Recommend prototype release only if all required claims meet criteria; otherwise issue a conditional recommendation with named closure actions.

Result. A defensible recommendation is configuration- and claim-specific. It is allowed to be conditional when evidence gaps remain.

Independent check. Every number is traceable, every requirement has a disposition, AI use is disclosed, limitations are visible, and the decision owner can reproduce the reasoning.

Common misconceptions

MisconceptionCorrection
A strong report sounds certainProfessional strength comes from calibrated claims, visible limits, and explicit residual risk.
A tool output closes the questionA result remains a candidate until its inputs, method, configuration, uncertainty, and relevance have been checked.

Diagnostic questions

What is the oral defense testing?

Understanding of evidence relationships, assumptions, calculations, uncertainty, decisions, and response to change.

What would make this work reproducible?

Controlled inputs, method or code, versions, assumptions, outputs, and a stated interpretation tied to the decision.

Practice ladder

Basic

Write a one-paragraph recommendation that separates fact, inference, and action.

Intermediate

Answer why the test does not validate every bracket failure mode.

Advanced

Defend the decision after an examiner removes one evidence item and changes one assumption.

AI-assisted engineering task

Ask AI to act as a skeptical examiner using only the supplied evidence index. It may ask questions but cannot grade final technical correctness.

How to prove the AI output yourself

  1. Answer with source IDs and calculations.
  2. Correct any unsupported premise in the question.
  3. Have instructor or panel judge reasoning.
  4. Record changes made after defense.

Retrieval and spaced review

Answer closed-notes today, then again after 1, 3, 7, and 30 days.

Define Decision recommendation.

A bounded proposed action justified by evidence and compared alternatives.

What role does Limitation play here?

A condition, omission, uncertainty, or domain boundary that restricts interpretation.

What must a reviewer be able to reconstruct?

Every number is traceable, every requirement has a disposition, AI use is disclosed, limitations are visible, and the decision owner can reproduce the reasoning.

End-of-lesson summary

The final recommendation must state the decision, configuration, evidence basis, uncertainty, credibility, alternatives, limitations, verification and validation plan, change triggers, and accountable approvals. Oral defense tests whether the student understands the chain rather than merely assembling artifacts.

Student notes

Use a claim-evidence-limit-action table as the executive summary of the final report.

Recommended readings

Instructor notes

Ask oral questions that remove evidence or change context. Memorized definitions should not pass the defense.

LAB 12

Lab 12: Assemble and audit a capstone evidence package

Lab objective

Combine a manifest, graph, calculations, validation metrics, AI audit record, and change-impact report into one reproducible package check.

Engineering context

The lab uses small teaching records so students can adapt the audit script to any approved capstone option.

Input data

  • Artifact manifest
  • Typed links
  • Required deliverable types
  • One changed requirement

Step-by-step task

  1. Validate unique IDs
  2. Check required artifact types and link endpoints
  3. Find orphan requirements
  4. Run change traversal
  5. Print a release-readiness gap report

Python code

from collections import defaultdict, deque

artifacts = [
    {"id": "N1", "type": "need"}, {"id": "R1", "type": "requirement"},
    {"id": "A1", "type": "architecture"}, {"id": "M1", "type": "model"},
    {"id": "S1", "type": "simulation"}, {"id": "T1", "type": "test"},
    {"id": "U1", "type": "uncertainty"}, {"id": "GP1", "type": "surrogate"},
    {"id": "O1", "type": "optimization"}, {"id": "AI1", "type": "ai_audit"},
    {"id": "D1", "type": "decision"},
]
links = [
    ("N1", "derives", "R1"), ("R1", "constrains", "A1"),
    ("A1", "configures", "M1"), ("M1", "produces", "S1"),
    ("T1", "validates", "M1"), ("S1", "feeds", "U1"),
    ("S1", "trains", "GP1"), ("GP1", "supports", "O1"),
    ("AI1", "reviews", "D1"), ("U1", "supports", "D1"),
    ("O1", "supports", "D1"), ("T1", "supports", "D1"),
]
ids = [a["id"] for a in artifacts]
assert len(ids) == len(set(ids))
idset = set(ids)
assert all(s in idset and t in idset for s, rel, t in links)
required = {"requirement", "architecture", "model", "simulation", "test",
            "uncertainty", "surrogate", "optimization", "ai_audit", "decision"}
present = {a["type"] for a in artifacts}
print("Missing types:", sorted(required - present))
adj = defaultdict(list)
for s, rel, t in links: adj[s].append(t)
queue, affected = deque(["R1"]), set()
while queue:
    for target in adj[queue.popleft()]:
        if target not in affected:
            affected.add(target); queue.append(target)
print("Review after R1 change:", sorted(affected))

Explanation of code

Step 1 validate unique IDs Step 2 check required artifact types and link endpoints Step 3 find orphan requirements Step 4 run change traversal Step 5 print a release-readiness gap report

Expected output

No missing required types and a downstream review set from architecture through decision. The exact order is sorted for reproducibility.

Interpretation

Structural completeness is only the audit floor. The final review must evaluate technical quality, VVUQ, uncertainty, AI transparency, and decision logic.

Common errors

  • Equating node presence with evidence adequacy
  • Omitting rejected AI suggestions
  • Failing to pin configurations and file versions

Extension tasks

  • Add checksums and file paths
  • Add schema validation
  • Add rubric-based evidence gates
  • Package with an environment lock file

Reflection questions

  • Which required type could still be technically weak?
  • Why is the test linked both to model and decision?
  • What artifacts should be re-reviewed after R1 changes?
WEEK 12

Weekly quiz and concept check

Closed notes. Answer each item, then use the key to correct in a different color.

  1. What makes a capstone appropriately scoped?
  2. How should synthetic data be labeled?
  3. What does an evidence map enable?
  4. What is the output of impact traversal?
  5. What must a final recommendation disclose?
  6. What does oral defense test?
Answer key
  1. 1. One tractable decision with a complete, reproducible evidence chain and explicit exclusions.
  2. 2. Explicitly as synthetic, with generation model, noise, purpose, and limitations.
  3. 3. Coverage, navigation, change impact, audit, and bounded AI assistance.
  4. 4. A suspect review set that engineers disposition using physics and configuration.
  5. 5. Evidence, uncertainty, alternatives, limitations, residual risk, AI use, V&V plan, and change triggers.
  6. 6. The student's command of reasoning, calculations, evidence links, assumptions, and response to changed conditions.
SOURCES

Module source map

SourceHow it is used
NASA Systems Engineering Handbook, NASA/SP-2016-6105 Rev. 2Lifecycle processes, requirements, interfaces, technical decisions, reviews, verification, and validation.
Singh and Willcox, Engineering Design with Digital ThreadDigital thread as a lifecycle data architecture and sequential design-decision problem under uncertainty.
NASA-STD-7009, Standard for Models and SimulationsModel and simulation lifecycle, credibility products, acceptance criteria, and reporting. NASA-STD-7009B supersedes 7009A.
ASME V&V 40, Risk-informed Model CredibilityContext of use, model risk, and evidence rigor commensurate with decision consequence.
NIST AI Risk Management Framework 1.0Govern, Map, Measure, and Manage functions for trustworthy and responsible AI risk management.
Martins and Ning, Engineering Design OptimizationOptimization formulation, constraints, derivatives, MDO, uncertainty, and algorithm choice.

Access labels and full-course source notes are on the course home page. Paywalled standards are not paraphrased as if their full text were accessed.