fix: increase edge inference confidence threshold to reject non-fish images

[sks-007]

Resolves #99

Problem Statement

When running the application in offline mode, taking a picture of a non-fish object (like a human face or generic object) often resulted in the app blindly accepting the image and returning a bogus "Fresh" score, rather than explicitly rejecting it.

Root Cause

The NOT_A_FISH_THRESHOLD in src/fusionInference.js was configured to 0.36. Because the Stream A MobileNet model has 3 classes, pure random chance yields a baseline probability of 0.33. This meant the threshold was so low that almost any out-of-distribution image's maximum class probability would pass the check, successfully bypassing the guardrail.

Solution Implemented

• Increased NOT_A_FISH_THRESHOLD from 0.36 to 0.55.
• Added an explanatory comment above the constant.

This requires the model to be significantly more confident than random chance before accepting an image as a valid fish, effectively filtering out generic objects and faces without the payload overhead of downloading a secondary object-detection model.

Testing Performed

• Verified offline edge inference with various fish images successfully passes the new threshold.
• Verified offline edge inference with non-fish objects (faces, coffee mugs) correctly falls below the 55% threshold and triggers the NOT_A_FISH error state.

Potential Impacts

Very blurry or poorly lit legitimate fish images might occasionally fall below the 55% confidence threshold, requiring the user to rescan (which is the intended safe fallback behavior).

[vercel]

@sks-007 is attempting to deploy a commit to the karan3431's projects Team on Vercel.

A member of the Team first needs to authorize it.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
fresh-scan-ai	Ready	Preview, Comment	Jun 22, 2026 8:45am

[jpdevhub]

Thanks for contributing