Shapes, color, and time field card

Use this card at every decode, metric, model, and save boundary. Write the answers beside the code; do not keep them in your head.

Array contract

• Shape: image H × W × C; clip T × H × W × C; model image often C × H × W.
• Axis meaning: name time, height, width, and channel explicitly.
• Dtype and range: uint8 0..255 or float with a declared range such as 0..1.
• Ownership: a slice may be a view; use .copy() before isolated annotation.
• Arithmetic: convert to a wider or floating dtype, then clip before conversion back.

Runtime: Browser lab · Pyodide with NumPy · 5 min.

import numpy as np

frame = np.zeros((1080, 1920, 3), dtype=np.uint8)
print(frame.shape, frame.dtype, float(frame.min()), float(frame.max()))

Color contract

• Pillow: ordinarily RGB.
• OpenCV decode: ordinarily BGR.
• Grayscale: record the conversion rule rather than saying “black and white.”
• Saved output: confirm the consumer’s expected color order.
• Blue/orange swaps are a contract clue, not a grading tweak.

Timeline contract

• Record duration, time base, stream/frame rate fields, and tool version.
• Prefer presentation timestamps for evidence-sensitive sampling.
• State uniform, boundary-aware, or event-driven sampling policy.
• Keep ordered timestamps beside extracted frame filenames.
• A nominal fps does not prove constant presentation cadence.

Five boundary questions

1. What exact data entered?
2. What are its shape, axes, dtype/range, and color order?
3. Which timestamp or sampling rule selected it?
4. Does this operation share or own memory?
5. What inspectable evidence leaves the boundary?