# Architecture testing notes

This file collects testing-related facts implied by the architecture diagram and project context.

## Timing requirement

- Analysis + result description must fit within **15 seconds per image**.
- Measure end-to-end time: trigger → inference → DB write → UI update.

## Test data sets

### Initial dataset

- Real photos of polyurethane soles.
- Used for the first training run.

### Learning dataset

- Combined from:
  - initial dataset;
  - artificially generated / augmented images;
  - expert-verified production data.

### Separate test set

- Must not overlap with training data.
- Size:
  - 20 soles without defects;
  - 3 soles with defects.

## Artificial disturbances

Use the artificial image generator to simulate production conditions:

- lighting effects;
- PNG pattern overlays (dust, lens dirt, other obstacles);
- rotation (±5° to match allowed positioning tolerance);
- noise generator for dust emulation.

## Position tolerance

- Rotation: ±5° relative to camera.
- Displacement: up to 10% of frame.
- Test inference quality across the full tolerance range.

## Camera replacement study

- Compare recognition quality when swapping camera models.
- Document changes in preprocessing parameters needed after replacement.

## Testing environment

- Run under conditions close to real production.
- Initial development can happen at home / lab using fake input data and artificial disturbances.
- Physical dimensions of soles must stabilize before testing (fixed time after casting).

## Expert feedback loop

- Every production event should be reviewable by an operator / expert.
- Expert verdicts (correct / incorrect) feed the learning dataset.
- Retraining can be filtered by date to avoid including low-quality old data.

## Model versioning

- Each trained / retrained model must be versioned.
- Keep metrics for each version so performance can be compared.
- Allow rollback to a previous model if retraining degrades quality.