Milestone 8 of 8
Write and audit the report
Connect every statement to saved item-level or aggregate evidence and state undefined values and limitations plainly.
The report should make the evaluation definitions and evidence easy to follow. Every claim should point to an item-level or aggregate artifact.
Goal
Write and audit a concise report that states the definitions, totals, per-label and slice evidence, paired differences, undefined values, and limitations.
Inputs
Use aligned_items.csv, overall summaries, both confusion tables,
per_label_results.csv, slice_results.csv, paired_comparison.csv, category
counts, all figures, evaluation_manifest.json, and the replay record. Read the
saved files rather than relying on console output.
Deliverables
Complete the project-level report.md with these sections:
- Question and alignment: state the supplied records, ID join, configured label order, and why row position was not used.
- Definitions and totals: explain match fraction, confusion-table axes,
per-label counts and ratios, and
nullfor undefined denominators. - Slices and paired comparison: state denominators, describe one source group, and explain the four paired outcome categories.
- Evidence: trace one item through aligned fields and paired category, and point to the tables and shared-scale figures.
- Replay and limits: state agreement or mismatch and explain what this fixture cannot establish about significance, fairness, causality, or future performance.
Audit every numerical statement against a saved artifact. Use “match fraction for these supplied items” rather than “accuracy” unless a later subject has defined that term for a different boundary.
Checks
Check that the report names a real item, follows its labels through the aligned record and paired comparison, and agrees with overall, per-label, slice, and category artifacts. Confirm that undefined ratios remain visible as undefined and that replay claims match the replay record.
Check that the report states at least one concrete limitation and separates observations from hypotheses. Remove claims about ranking, significance, causes, fairness, model selection, or future performance that the project does not test.
Workspace
Keep the report at the project-level report.md. Do not duplicate every table
or trace in prose. The report explains durable files; the manifest, aligned
records, tables, figures, paired comparison, and replay record remain the
primary evidence.
Hints
HintFollow one item
HintUse the supplied names
match_fraction has a precise local meaning. Do not silently rename it to a broader claim.Review
Read the report as a new reader. Can they tell how rows were aligned, where a ratio came from, what a slice denominator was, which items differed, and what the evidence leaves unknown? Replace broad wording with an artifact-backed statement or remove it.
How to check your work
Checks compare the report structure and evidence links with the supplied fixture. The supplied fixture is an exemplar of traceable writing, not a conclusion to copy without checking your own artifacts.