Lab 7: AI ethics and governance¶

Outcomes: SLO 6. Planning time: two sessions of 45–75 minutes for a first attempt; this is a planning estimate, not a measured novice completion time.
Environment: your dedicated local course database. Synthetic data only.

At a glance¶

Before starting	Labs 1–6
You will run	Calculate the supplied prediction metrics and inspect the synthetic evaluation records.
You will write	One worksheet containing metric interpretation, an initial governance decision and a required change review.
Done when	The technical evidence, one checked calculation and all three worksheet sections are complete.
Safe stopping point	After a completed section; save your draft before closing the editor.
Submission format	Use the shared template plus the completed decision worksheet; refer to worksheet sections instead of duplicating them.

Follow the steps below in order. Keep configuration, generated logs and submissions private. Return to the required course path when this lab is complete.

Why this lab matters¶

Improving one AI error rate can worsen another. A model or policy can look better on one metric while creating a different cost for another group or stakeholder.

In this lab, you inspect a small synthetic evaluation fixture, check one error-rate calculation yourself, then make and revise a governance decision. The second decision changes the proposed purpose, so you must decide whether the earlier evidence is still sufficient rather than assume that one evaluation supports every future use.

Learning objectives¶

These instructor-developed objectives support the outcomes listed above. You will:

Calculate and interpret group error rates with the correct denominators.
Explain bias, transparency and accountability throughout the data lifecycle.
Distinguish metric improvement from a complete fairness conclusion.
Defend and revise a deployment decision using NIST AI RMF 1.0.

Skills you will practice¶

Read metric evidence, interpret a small CSV fixture, check arithmetic, assess tradeoffs, distinguish evidence from uncertainty, assign accountable roles and write a short change review.

Completion: Automation passing means the technical calculation completed. The governance decision, lifecycle reasoning and change review remain required.

Terms you need for this lab¶

Term	Meaning here
Prediction / label	The supplied selection decision / the reference eligibility outcome. This lab evaluates existing predictions; it does not train AI.
Selected	For this exercise, a prediction value of `1` means the record is selected for support priority; `0` means it is not selected.
False negative	A record whose reference outcome is eligible (`eligible = 1`) but whose prediction is not selected (`0`).
False positive	A record whose reference outcome is ineligible (`eligible = 0`) but whose prediction is selected (`1`).
Label bias / representativeness	Reference outcomes may contain systematic errors; a small fixture may not represent the intended population.
Proxy / audit group	A field indirectly associated with another characteristic / a grouping used to compare results. Auditing groups does not mean using them as prediction inputs.
Metric parity	Two groups having the same value on one metric. Parity on one metric does not establish overall fairness.

Groups A and B are synthetic audit groups. Do not infer a real demographic attribute such as race, sex, disability or another protected characteristic from these labels.

Use the glossary for more detail; this is reference support, not another assignment.

Your route and deliverables¶

[ ] Part A: run the metric calculation, read the report and verify one group-B false-negative-rate calculation.
[ ] Worksheet section 1: interpret the metrics and their limitations.
[ ] Worksheet section 2: complete the lifecycle record and make an initial governance decision.
[ ] Worksheet section 3: complete the required changed-purpose review.
[ ] Submission: include narrow technical evidence plus the completed worksheet; do not write a duplicate essay.

You submit this one worksheet plus the relevant execution evidence. Read the fictional library decision if you need an example of the expected depth. It demonstrates structure and scope; it is not a retail answer key.

Before you begin¶

Complete Labs 2–6 for governance, lineage, access and monitoring context. Use your existing Ubuntu environment; setup need not be repeated.

If needed, begin with Student start here. No AI programming experience, external account or model training is required. All commands below run from the repository root.

Part A: Follow the metric example¶

A1. Calculate the supplied metrics¶

Run:

.venv/bin/python scripts/course.py lab 7

Expected:

PASS: connection, dedicated database, schemas and non-superuser builder.
PASS: AI metrics calculated. Complete the governance decision; metric parity is not proof of fairness.
LAB 7 TECHNICAL CHECKS COMPLETE. Return to labs/core/lab07-ai-governance/README.md for interpretation and deliverables.

This command reads the supplied synthetic prediction file, calculates the four group/policy metric sets and writes:

.local/ai-evaluation.json

It does not train a model or make a new prediction.

Checkpoint: retain the command and the two relevant PASS/completion lines privately. Technical completion does not complete the written task.

A2. Read the report¶

Run:

.venv/bin/python -m json.tool .local/ai-evaluation.json

This command only prints the saved JSON report in a readable format.

The report is nested in this order:

policy
→ audit group
→ metric

Use this key:

JSON field	Meaning
`baseline`	The original supplied prediction policy
`mitigated`	The supplied alternative prediction policy
`A`, `B`	Synthetic audit groups
`n`	Number of records in the group
`selection_rate`	Selected records divided by all records in that group
`false_negative_rate`	Eligible records not selected divided by eligible records
`false_positive_rate`	Ineligible records selected divided by ineligible records

Expected values for the unchanged synthetic fixture:

Policy	Group	Records	Selection rate	False-negative rate	False-positive rate
Baseline	A	8	0.500	0.25	0.25
Baseline	B	8	0.125	0.75	0.00
Mitigated	A	8	0.500	0.25	0.25
Mitigated	B	8	0.625	0.25	0.50

Rates are proportions: 0.25 means 25%.

Read the denominators correctly¶

Selection rate uses all records in the group.

False-negative rate uses only eligible records:

eligible records not selected
-----------------------------
all eligible records

False-positive rate uses only ineligible records:

ineligible records selected
---------------------------
all ineligible records

In this fixture, each group contains:

8 total records
4 eligible records
4 ineligible records

Therefore, one record changes an FNR or FPR by:

1 / 4 = 0.25

This is one reason the fixture should be treated as a small teaching example, not strong population evidence.

What to notice before moving on¶

Group A's reported metrics are unchanged between baseline and mitigated. That is expected for this fixture.
Group B's false-negative rate decreases from 0.75 to 0.25.
Group B's false-positive rate increases from 0.00 to 0.50.
After mitigation, groups A and B both have an FNR of 0.25, but that is parity on one metric only.
Equal FNR does not prove fairness: FPR still differs, the sample is tiny, reference labels may have limitations and other harms or affected groups are not measured.

Do not describe the mitigated policy simply as “better.” Explain the tradeoff.

A3. Check one calculation against the source records¶

Open the synthetic CSV on GitHub or locally:

nano data/ai_predictions.csv

Read only. Leave Nano with Ctrl+X without saving changes.

How to read the CSV¶

Each row is one synthetic evaluation record.

Column	Meaning
`id`	Synthetic record identifier
`group`	Synthetic audit group `A` or `B`
`eligible`	`1` = reference outcome says eligible for support priority; `0` = ineligible
`baseline`	`1` = baseline selected the record; `0` = baseline did not
`mitigated`	`1` = mitigated policy selected the record; `0` = mitigated policy did not

The eligible column is the reference outcome used for this exercise. The fixture does not establish how those labels were originally produced or whether they are unbiased. You are not expected to prove label bias. Treat label quality as a possible limitation and state what further evidence would be needed.

Worked example: baseline group A¶

For baseline group A, four records are eligible. One of those four eligible records was not selected:

FNR = 1 / 4 = 0.25

Your calculation: baseline group B¶

Calculate baseline group B's FNR yourself.

Record these five items in your worksheet notes:

Eligible group-B records: ___
Eligible group-B records missed by baseline: ___
FNR calculation: ___ / ___ = ___
Does the result match A2? ___
What could this error mean in the support-priority scenario? ___

For the final question, reason from the scenario:

eligible + not selected = false negative
ineligible + selected   = false positive

Explain a plausible cost of the error. You do not need to change the data or write Python.

Part A complete: you have the Lab 7 execution evidence, the expected metric comparison and one calculation you verified directly from the CSV.

Part B: Make and revise a governance decision¶

B1. Create your private writing worksheet¶

Run these commands one at a time:

mkdir -p .local

cp -i labs/core/lab07-ai-governance/ai_decision_template.md .local/lab7-decision.md

If asked to overwrite on a repeat attempt, answer n to preserve your answers.

Then open the worksheet:

nano .local/lab7-decision.md

It contains three deliverables, not code:

metric interpretation;
initial governance decision and lifecycle record;
required changed-purpose review.

Use short answers. Bullets are acceptable. You may also use a word processor if that is easier than editing in Nano.

The worksheet uses vertically labeled lifecycle sections rather than a wide Markdown table so that it is easier to edit in a terminal.

Put your Part A calculation and interpretation in section 1. Complete the lifecycle and decision record in section 2. Complete the change-review amendment in section 3.

Save in Nano with Ctrl+O, Enter, then Ctrl+X.

B2. Complete the initial decision¶

Use the metrics from Part A as evidence. Your reasoning should address:

metric evidence
→ affected stakeholders
→ competing error costs
→ evidence limitations
→ proposed safeguards
→ accountable roles
→ decision
→ review trigger

Discuss:

the mitigation's increased false-positive rate;
the small sample;
possible label-quality or representativeness limits;
whose interests or harms may not be captured by the measured rates;
transparency and appeal arrangements;
accountable roles;
which safeguards are observed versus merely proposed.

Label safeguards you have not implemented as proposed.

NIST AI RMF scope for this lab¶

Use NIST AI RMF 1.0 only at the level requested by the worksheet:

Govern — who is responsible and accountable?
Map — what is the context, affected population and potential harm?
Measure — what evidence exists, and what uncertainty remains?
Manage — what action, mitigation, monitoring or review follows?

You do not need to summarize the entire NIST AI RMF or Playbook. Four short labeled bullets applying these functions to your decision are sufficient. The linked resources are reference material, not an additional reading assignment.

Safe stopping point: after section 2, save your worksheet. Resume by reopening the same file; you do not need to recalculate unchanged metrics.

B3. Complete the required change review¶

The proposed use now changes:

original purpose: support prioritization
changed purpose:  restricting refunds

The incoming population is also different, and no new performance or harm measurements are supplied.

Do not assume that the earlier support-priority evaluation automatically supports the new refund-restriction purpose.

Use this sequence:

What changed?
→ Which earlier evidence no longer answers the new question?
→ What new evidence is required?
→ Who must review and approve the change?
→ What happens while the evidence is missing?
→ How should notice and appeal change?
→ What happens to old predictions and retained copies?

Write the amendment requested in worksheet section 3. Do not invent performance results. Both the initial decision and the amendment are required; this is a writing task, not another script to execute.

Submit¶

Optional: use the local report checker after writing your submission. Replace the path below with your actual private report filename. Include both files if your worksheet is separate; omit the second path if it is embedded in your report. Warnings are suggestions, not grades or new submission requirements.

.venv/bin/python scripts/report_checks/lab07.py .local/lab7-report.md .local/lab7-decision.md

Use the submission template, but do not duplicate the completed worksheet in a second essay.

Your Lab 7 submission should contain:

the Lab 7 command and relevant A1 PASS/completion lines;
the relevant A2 metric evidence, either as a narrow excerpt or a reference to the printed values;
your A3 baseline-group-B FNR calculation;
the completed private lab7-decision.md worksheet, including sections 1–3;
recovery notes if an unexpected error occurred;
your assistance disclosure and how you verified your work.

How the shared submission template maps to Lab 7¶

Shared template section	What to do for Lab 7
1. Lab and execution evidence	Include the Lab 7 command, relevant PASS lines and narrow A2 metric evidence.
2. Prediction or initial expectation	Write `Not required for this activity` unless you actually recorded a prediction requested by your instructor.
3. Observed result and explanation	Include your A3 calculation or refer to worksheet section 1 if it already contains the same explanation.
4. Investigation and independent transfer	Attach or refer to the completed `lab7-decision.md` worksheet, sections 1–3.
5. Evidence limitations and questions	Refer to limitations already stated in the worksheet; do not rewrite them unless something remains unresolved.
6. Assistance and verification	State AI/tool assistance and how you checked the result, or `None`, following course rules.

Text evidence is sufficient. Screenshots are optional. Never include .env, credentials, tokens, full connection strings or personal terminal details.

Submit privately through the LMS; independent learners keep the same record locally.

Assessment¶

Assessment uses only the AI decision rubric:

lifecycle: 20%;
bias analysis / limitations: 25%;
mitigation tradeoffs: 20%;
transparency: 15%;
accountability / framework application: 20%.

The general core-lab rubric does not apply.

The suggested word ranges in the worksheet help scope the work. They are not a requirement to fill space. Concise answers that address the requested evidence and reasoning are sufficient.

Recovery¶

Use the setup troubleshooting table. Read errors before retrying; never delete tests or change authentication to make a result pass. There is no automatic destructive reset. For an isolated new start use a new DB/user pair.

If .local/ai-evaluation.json is missing, rerun Lab 7 and resolve any reported error before continuing. Do not manually edit the JSON to force the expected values.

If you accidentally modify data/ai_predictions.csv, stop and restore the repository version before using it as evidence. Do not change fixture values merely to reproduce an expected metric.

Back to all labs

Author: Isaac K. Nti. Citation and reuse terms.