Write and audit the toolkit card

The final card is a short account of the image contract and the evidence it produced. It should be useful without making the toolkit sound more capable than it is.

Goal

Write and audit a toolkit card that states image modes, axes, dtype and range, coordinates, every operation and edge rule, plan order, checks, artifacts, replay evidence, and limits.

Include

Describe:

• the manifest boundary and supported L and RGB modes;
• the (height, width, channels) array shape and axis meanings;
• uint8 values from 0 through 255 and independent arrays;
• half-open crop coordinates and both flip directions;
• length-one channel selection;
• wider arithmetic, clipping, mask shape, and fill-value rules;
• nearest-edge repetition and floor division for mean_3x3;
• plan validation, stage records, pixel checks, and count conservation;
• PNG save/reload equality, manifests, and separate replay; and
• what later vision work may add only through a versioned boundary.

Do not claim

Do not call the mean filter learned, optimal, or convolution training. Do not claim that a crop, flip, brightness change, mask, or plan improves visual quality, model accuracy, robustness, or data augmentation. Do not call the fixtures representative of natural images or suitable for every later task. Do not imply that a successful replay proves perceptual or model quality.

Deliverables

Write report.md and audit each sentence against a saved artifact. Include source and plan identities, counts and ranges, one traced stage, one reload check, replay agreement or the first mismatch, and a short limitations section. Name the numerical evidence beside each observation.

Checks

Use the prose rubric to reject unsupported visual-quality, semantic, augmentation, learned-feature, model-improvement, and representativeness claims. Check that the report names unsupported modes, invalid-plan behavior, edge rules, channel shape, clipping, count denominators, source immutability, and replay limits.

Read the card after changing one artifact and confirm that the audit catches a stale statement. Confirm the report is readable without the interactive editor and its examples retain image and stage identities.

Workspace

Keep the card and audit in src/report.py or a clearly separated report writer. Do not turn a visual comparison into a claim about what a person or a model will prefer.

Hints

Review

Trace one sentence in the card to an inventory record, stage record, pixel check, table, PNG reload record, or manifest. Then name one important thing the project cannot establish. Can a reader who skipped DATA-03 still understand the array and source boundary?

How to check your work

Checks compare the card with the accepted image-toolkit contract and saved manifest. The card is complete when every public claim has bounded evidence and every important limit is stated plainly.