Write and audit the patch-dataset card

The final card explains the spatial dataset boundary and its evidence. It should make the edge policy and reconstruction status easy to find.

Goal

Write and audit a patch-dataset card that states source identities, axes, grid and edge rules, pad values, patch and mask shapes, groups, batches, coverage, reconstruction, artifacts, replay, and limits.

Include

Describe:

• source and transformation identities and the shared channel contract;
• row-major origin generation and half-open source bounds;
• exact drop and pad rules, strides, and bottom/right padding;
• composite (image_id, top, left) identity and patch indexes;
• patch shape, mask shape, dtype, pad values, and valid-pixel meaning;
• aligned array and metadata invariants;
• source-times-coverage accumulation and complete/incomplete reconstruction;
• image groups, shared indexes, kept remainders, and zero-patch cases;
• saved arrays, tables, figures, manifests, and separate replay evidence; and
• which later vision work may change only through an explicit versioned rule.

Do not claim

Do not claim that one patch size, stride, overlap, padding policy, or group assignment is best for a model. Do not call patches embeddings or features. Do not imply that exact reconstruction proves useful representation, visual quality, model accuracy, fairness, or suitability for every image. Do not turn operational group names into training or validation claims.

Deliverables

Write the project-root report.md and audit each sentence against a saved artifact. Include image and configuration identities, grid counts, one traced patch and batch remainder, coverage and reconstruction status, one replay result or the first mismatch, and a short limitations section. Name the evidence beside each observation.

Checks

Use the prose rubric to reject unsupported embedding, attention, convolution, augmentation, sampling, and model-quality claims. Check that the report names row-major order, both edge policies, padding and mask semantics, coverage denominators, uncovered coordinates, group conservation, final remainders, 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 examples retain image, patch, origin, and group identities.

Workspace

Keep the card and audit in src/report.py or a clearly separated report writer. Do not turn a coverage map into a claim about information quality.

Hints

Review

Trace one statement in the card to a patch record, mask, coverage map, batch manifest, PNG/NumPy artifact, or replay record. Then name one thing the project cannot establish. Can a reader who skipped IMG-01 still understand the array and source boundary?

How to check your work

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