Skip to main content

Product, Platform, Responsible AI, and Sustainable Decisions

Competency Contract​

  • Primary competency: SEP-ETHICS-001
  • Mastery: explain, apply, develop, and evaluate as required by the semester assessment
  • Status: required evidence extension; included explicitly in the reconciled semester workload

Required Work​

Use product discovery and socio-technical analysis; design a platform golden path; apply NIST AI RMF reasoning; evaluate cost and software-carbon tradeoffs per functional unit.

Evidence Contract​

  • stakeholder and harm map
  • product assumption test
  • platform contract
  • AI risk decision
  • cost/carbon comparison.

Assessment Integration​

  • Project: produce the evidence in the semester repository rather than as detached prose.
  • Checkpoint: show the evidence exists and identify its highest-risk gap.
  • Cumulative review: connect the work to current, previous, and two-semesters-back knowledge.
  • Exam: evaluate a new case and reject an attractive but unsafe or unsupported alternative.
  • Rubric: every required evidence category must pass; strengths cannot average away a missing category.
  • Answer key: calibrate the shape of evidence without prescribing one implementation.

Oral Defense​

Defend who benefits, who bears risk, what the platform standardizes, and why an alternative was rejected.

AI-Use Declaration​

Record tools and models used, material output accepted, rejected suggestions, and independent verification. Undeclared generated evidence fails the integrity gate.

Reading Route​

Start with the linked semester concepts and labs. Use local books selectively for conceptual depth and current official standards for normative requirements. Stop reading when you can produce and defend the required evidence.

P1–P5 practical evidence and calibration​

Budget 40–55 hours: P1 7–10, P2 7–9, P3 9–12, P4 9–13, P5 8–11. Reuse module work and count only additional investigation. Score each 0 absent, 1 prompted, 2 independent evidence with limitations, 3 fresh transfer. Every category requires at least 2.

P1 — Stakeholders and harm (7–10 hours)​

Extend the existing ethical-impact record with rider, driver, support/operator and excluded-user perspectives. Trace two concrete harm scenarios through a technical decision, safeguard, accountable owner and remediation trigger. Obtain an actual consenting review from a different perspective; a role-play is preparation. Record disagreement without personal identifiers. Calibration: a generic fairness statement with no affected group or trigger is incomplete.

P2 — Product assumption test (7–9 hours)​

Choose one assumption from the design, such as whether operators can safely use the replay tool. State a falsifiable hypothesis, participant/task scope, benefit, safeguard and stop condition before testing. Run a small synthetic task study with a consenting peer, preserving failed attempts and help given. Report observed behavior and selection limits; do not claim a representative customer study. Reuse TL4 feedback where relevant, while adding the task observations. Calibration: mandatory adoption or positive opinions alone do not establish usefulness.

P3 — Platform contract (9–12 hours)​

Implement a narrow local golden path for the existing synthetic service: documented setup, configuration validation, a health check and a safe failure/recovery path. Define consumer, version, supported inputs, escape hatch, support owner and deprecation expectation. Have another person follow it from a clean checkout without hidden local state; retain the failed step and repair. Exercise one invalid configuration and demonstrate that no live credentials or customer records are required. This is a small platform capability, not a mandate to adopt Kubernetes. Calibration: a script working only on its author's machine does not pass consumer usability.

P4 — AI risk decision (9–13 hours)​

For an optional dispatch-ranking or operator-assistance feature, compare no-AI and bounded-AI alternatives. Use NIST AI RMF 1.0 as the identified teaching version: govern ownership, map intended use/affected groups, measure a scoped risk and manage safeguards. Create synthetic cases covering normal, ambiguous and harmful inputs; declare whether outputs are actual model observations or hypothetical tabletop responses. Never label hypothetical outputs model evaluation. Run a local policy-boundary test showing that an unsafe proposal cannot trigger the business action, that a legitimate permitted case still works, and that human review/disable paths remain available. Name residual risks, an owner and a reconsideration trigger. A no-AI decision can pass when supported by this comparison and tested boundaries; framework headings alone cannot.

NIST's official AI RMF page provides the versioned framework and playbook. Record the version/date used; do not claim certification or legal compliance from this exercise.

P5 — Cost and carbon per functional unit (8–11 hours)​

Compare two configurations for the same declared workload, quality target and boundary. Record costs with assumptions and date, and energy/embodied-emission inputs as measured, modeled or unavailable. Use a functional unit such as one completed synthetic trip; include unsuccessful work's resource use in the boundary. Calculate an illustrative SCI-style intensity as (energy × carbon intensity + allocated embodied emissions) / completed units. Explain allocation and uncertainty, test sensitivity and retain total consumption alongside the rate. Do not infer emissions directly from CPU time or cloud price.

Teaching example: A uses 2 kWh at 400 gCO2e/kWh plus 200 gCO2e allocated embodied emissions for 1,000 completed units: 1 gCO2e/unit. B uses 3 kWh at 100 plus 200 for the same 1,000 units: 0.5 gCO2e/unit. B uses more energy but this assumed mix yields lower emissions. If B's intensity is 400, its result becomes 1.4. These are fictional estimates, not measured sustainability claims. Lower unit intensity does not guarantee lower total emissions as demand grows; preserve latency, availability and privacy constraints when considering region/time shifts.

Use the Green Software Foundation SCI specification for boundaries and units. Disclose missing data instead of claiming compliance with the complete specification.

Independent defense​

Apply all five categories to a changed scenario after at least 48 hours: support load rises, telemetry is missing, and the lowest-cost region has different user constraints. Explain which decision changes and which evidence is still absent. Preserve original real service, security and ethical-impact obligations.

Concrete assistance/release comparison​

Use the required comparison and calibration within the stated category allocation. It specifies existing evidence rather than replacing other strand requirements. Keep hypothetical fixtures separate from actual tool observations.

Common statistical-learning requirement​

Complete the statistical-learning bridge before P4. SL1–SL4 each require level 2, alongside all existing gates. Its 20–30 hours are additional to P4 and included in the revised 510–725-hour total. Reuse its evidence in integration; no additional timed exam section is introduced.