Milestone 6 of 8

Inspect the evidence numerically and visually

Create count and distribution plots with visible denominators and inspect unusual documents by ID without assigning linguistic importance.

The figures are another view of saved evidence. Inspect the counts and document identities before deciding what a pattern means.

Goal

Create document-length, line-length, and finding-count plots, inspect selected documents by ID, and state observations without assigning linguistic importance.

Inputs

Use document_measurements.csv, corpus_summary.json, source_group_summary.csv, inspection_findings.csv, and selected_examples.json. Keep document IDs and the configured state and group labels available to the plotting code.

Plot data must come from saved measurements and findings. Show denominators or document counts where a distribution or finding count could otherwise be misread. Do not use a word cloud.

Deliverables

Implement plot-data preparation and figures in src/report.py. Save figures under output/, including document-length, line-length, and finding-count views, plus explicit plot data that retains document identities or stable group/rule keys.

Inspect at least one document with an unusual measurement and one with a finding. Record two observations in the report with IDs and exact counts. A possible explanation must be labelled as a hypothesis, not presented as a fact about language or document quality.

Checks

Check that each plotted count comes from its saved artifact, that document IDs behind selected points can be recovered, and that group and finding counts retain their denominators. Read plot data back and use a deliberate measurement mutation to confirm the audit catches stale figures.

Check that no plot silently excludes rejected records when the figure claims to show corpus state, and that the word-cloud pattern is not used as a substitute for counts or identity.

Workspace

Keep plot-data and figure code in src/report.py. Read saved measurements and findings from output/ and leave them unchanged. Keep the report at the project-level report.md for the final milestone.

Hints

HintA point needs an identity
Store document ID or a stable group/rule key with every plotted value so a reader can find its source.
HintState the denominator
A finding count without the number of inspected documents can make a small corpus look larger than it is.

Review

Read a figure alongside its source table. Identify one point or bar, find its document or group, and state only what the saved count shows.

How to check your work

Checks compare plot data, labels, denominators, and figure identity with the supplied fixture. The supplied fixture avoids linguistic interpretation based on a decorative graphic.

LLM PrimerInspect the evidence numerically and visuallyhttps://llmprimer.com/python/projects/inspect-a-text-corpus/inspect-the-evidence-numerically-and-visually© 2026 LLM Primer