Recognition analysis
When recognition is right but the click is wrong
Trace image keys, dimensions, offsets, and movable identities to distinguish a UI mapping bug from an incorrect recognition result.
By the easycaptcha team · Published
A live rejection does not establish where a coordinate became wrong. The image may have been resized, cropped, offset in a frame, or associated with a different task key. Build a local fixture with known source points so you can verify the transform without asking a model to solve anything.
Geometry fixtures
Use points near all edges and the center. A symmetric example can let a bad transform look correct. Include the actual display sizes and browser zoom conditions supported by your application.
Trace the coordinate systems once
For an illustrative image rendered at twice its source width and height without cropping, a source point at (100, 50) scales to (200, 100) within the rendered image. Page and frame offsets are separate additions. Applying the scale twice or adding the wrong origin changes the interaction even if the source answer was correct.
Write down source dimensions, rendered dimensions, coordinate convention, crop, and offset ownership. Consult the actual API schema rather than assuming every service returns normalized fractions or source pixels.
Preserve identity as well as numbers
A correct point on the wrong image is still a wrong interaction. Binary answers need original task keys; area results need their image association; drag results need movable-object identity. Keep mappings explicit instead of aligning arrays by chance.
Record limited mapping evidence for controlled tests. Avoid copying unrelated form contents or full browser profiles into general traces.
Separate the failure stages in feedback
When an interaction is rejected, investigate input capture, solver completion, mapping, interaction dispatch, and downstream acceptance. Repeated paid submissions cannot repair a deterministic offset error.
Maintain a local classification for application failures even when a feedback API accepts only a boolean. Otherwise bad mapping is mislabeled as recognition failure and contaminates evaluations.
The UI-validation solution explains how to make known answers a precondition for live evaluation. Recognition families separates result models, and feedback quality examines how the final observer’s evidence affects the data.
Frequently asked questions
Why can a correct recognition answer produce a wrong click?
Scaling, cropping, offsets, or mismatched image identity can move the applied coordinates away from their intended position.
How should I test coordinate mapping?
Use known fixture points across the image, preserve original dimensions, and apply scale and offsets exactly once.
How can easycaptcha fit into my integration tests?
Use the documented recognition or token API for the hCaptcha stage, preserve task IDs, and inspect terminal results. Measure verifier acceptance and application completion separately from job completion.