COURSE 1 · CUMULATIVE PROJECT

An evidence-based AI decision.

Audit the data, compare a baseline, allocate review and defend a recommendation. A reasoned “no AI” is a valid outcome.

Saved locally with version 4 My Work backups. Report submission, first-attempt transfer and human review are separate. This is a classroom flower-label service proposal, not deployment evidence.

Try: should a classroom collection use label suggestions?

A school collection has four flower measurements per record. A team proposes a nearest-neighbor classifier that suggests one of three Iris species. Your task is to advise the collection coordinator: build a small educational tool, investigate a supplier, defer for evidence, or keep a human process. No option automatically earns a better grade.

Predict: could a repaired split lower accuracy yet improve the quality of the evidence? Could sending more predictions for review reduce automated mistakes while increasing workload?

Audit
Source, units, missing values, split

Compare
Baseline, mistakes, human review

Decide
Supported claim and next evidence

Entering midway? A feature is a measured input; a label is the known target. Training data chooses the predictor. Development data helps select settings. Repeated development results cannot serve as an untouched final test. Review data splits or review error costs.

Experiment: capture the audit

Use the shared lab below to capture two runs: (1) teaching view, row split, fit all data, ignore unit tags, mean fill; (2) teaching view, grouped split, fit training only, convert units, mean fill. Captures also appear in the existing data-audit project. The decision project copies the two configurations when you attach them; later edits elsewhere do not silently change your brief.

Loading the local audit workbench…

Capture both specified configurations, then attach them here.

Understand: compare evidence and review costs

Read the source, license and alteration manifest. Fisher (1936), Iris, UCI Machine Learning Repository, CC BY 4.0. The 5,977-byte local CSV contains 150 rows. Duplicates, missing cells and mixed units are teaching alterations. Identical measurement vectors define groups; these are not authenticated specimen identities. No personal identifiers are present.

Attach the two audits to see recomputed results. You may draft the brief or open the transfer case at any time.

Review policy and explicit teaching costs

Minimum closeness for an automatic labelCost per wrong automatic labelCost per referred measurement group

Closeness = 1 / (1 + distance). Distance is the nearest-neighbor distance after training-only standardization, with no physical unit. At distance 0, closeness is 1; at distance 1, it is 0.5. Values below the chosen threshold go to review; equality receives an automatic label. This is an uncalibrated distance rule, not a probability of correctness.

Apply: write a reviewable decision brief

Saved artifactEvidence variantRecommendation

Rubric: trace a numerical result, explain the split repair and development boundary, compare error and workload consequences, identify missing coverage, and justify the next evidence request. No decision option is automatically correct. A supplier has not been evaluated by this lab.

Task and affected usersDefine flower-label suggestions for a classroom collection, measurement timing, who checks uncertain predictions and a measurable acceptance condition.Data card and corrected splitCite the source/license, teaching alterations, measurement groups, training-only fitting and units. These plant records have no personal identifiers; explain retention and permission checks for any future data.Baseline, errors and uncertaintyCompare the leaky and repaired audit, majority baseline and review policy. Include counts, closeness threshold, costs, interval assumptions and missing coverage.Decision and operating assumptionsDefend build, buy, defer or no AI. State staff capacity, review cost, imperfect human review, supplier evidence if buying and who owns the decision.Supported claim and next evidenceSeparate repeated development from independent evaluation. Name excluded uses, an unfamiliar collection to evaluate, a stop condition and what would change your recommendation.Implementation and reproductionEngineering: paste run instructions, test output and your comparison with browser results. Other variants: describe how another learner can reproduce your choices.

Report: draft. Human artifact review: pending. Any edit reopens submission. Use a My Work backup to restore editable progress; this report is an inspection artifact.

Independent transfer: a different task

A fixed new library scenario checks split design, arithmetic, coverage and claim boundaries without calling the Iris model. It is self-administered and inspectable, not secret or proctored. Opening records exposure; your first submission is retained across reload/import. Up to two retries are practice and never replace it. Unsaved answer drafts are temporary. Written reasoning still requires review.

Continue with the evidence you need

Bring the brief to a peer or teacher for review. Ask them to challenge one numerical claim, one use assumption and one missing population. A completed form and a passed objective case do not demonstrate that the proposal is ready to deploy. Learner pilots remain pending.

Review the cumulative learning path · Practice additional audit transfer cases