Graded Work

Loaded: graded-work.md — rubric ingestion and submission handling.

Read this before anything else when the task is assessed work: coursework, a capstone, a certification project, or any deliverable with a scoring guide.

The rubric outranks this skill

Where the rubric names sections this spine omits, or weights them differently, follow the rubric and say where it diverges. The spine is a house style; the rubric is the actual specification being graded against.

Ask for it up front, before any code. If it isn't available yet, say what's being assumed in its absence — usually the template's own section headers — and revisit when it arrives. A rubric surfacing mid-session is normal; when it does, re-check completed sections against it rather than only applying it going forward.

Distinguish the scoring rubric (marks per section) from the submission format table (full-code vs low-code, file type, deliverable). They're different documents and are often confused. The format table governs the artefact; the rubric governs the content.

Reading a rubric properly

Three things a rubric encodes that are easy to miss:

"With comments" means the commentary is the deliverable, not the code. A line reading "Missing value treatment (if needed) with comments" is satisfied by a reasoned markdown cell explaining that no treatment was needed and why. Silence scores zero even when the analysis is correct. This applies equally to outlier treatment, dropped variables, and feature engineering — all commonly phrased as "if needed," all requiring an explicit statement either way.

Separately-listed criteria are separately marked. "Comment on model performance" and "Check the Feature Importance and comment on important features" appearing as two lines means two distinct write-ups, not one combined paragraph.

Named techniques are mandatory, not suggested. "Tune the model hyperparameters using GridSearchCV" means GridSearchCV specifically. Substituting RandomizedSearchCV for compute reasons is a judgement call that has to be surfaced and agreed, not made silently — see the compute budget rule in modeling-and-tuning.md.

Restate the rubric as a point-weighted checklist at the start, then map it to spine sections and name any divergence. Sections carrying the most marks get the most attention; a heavily-weighted tuning section deserves more depth than a lightly-weighted build section, even though the build feels like the bigger task.

Code quality is usually graded

Most rubrics for full-code submissions include a criterion covering structure, flow, and commenting. That makes the conventions in the spine's Code Idioms section directly worth marks rather than stylistic preference — particularly the natural-language comment on every cell, imports all at the top, and no unused imports.

Before submission, check three things that cost marks for no analytical reason:

1. Unused imports — anything imported but never called.
2. Mid-notebook imports — anything imported outside the top cell.
3. Cells run out of order — execution counts that jump around suggest a notebook that won't reproduce top-to-bottom. Restart-and-run-all before exporting.

Submission format

Confirm the required output format early, because it changes the last step: .html export, .ipynb, a PDF, a deck, or a written report. Check whether the submission is the notebook itself or a document derived from it — a "business report" variant is a different deliverable from a code notebook, with different weighting on visual presentation.

If the submission is a rendered export, the export happens after a clean restart-and-run-all, not from the working session's state.

Academic integrity

Assist with understanding, structure, technique, and interpretation. The analysis and the decisions belong to the person submitting it.

Where a course has its own AI mentor with instructions not to reveal answers to graded questions, respect the same boundary: explain what a technique does and why it applies, help debug what they've written, and let the interpretation of their own results be theirs. Writing an entire graded notebook end to end without their involvement is not what this skill is for, and the step-gated loop exists partly to make that difficult by design.