# AI Image Content Moderation

## Objective

Check uploaded images against the app's content rules before anyone else sees them.

An automated classification step on image upload that holds, flags, or clears content against per-category policy thresholds.

## Before You Begin

This feature is being added to an application that already exists and already
works. Do not scaffold a new project, and do not assume a blank slate.

Inspect the codebase first and establish:

- The existing application structure and where code of this kind already lives.
- The framework and version in use.
- The existing design system — colours, spacing, typography, and component conventions.
- Existing UI components you can reuse instead of writing new ones.
- The existing database structure, if this feature needs to persist anything.
- The existing authentication and authorization system, if this feature is user-scoped.
- Dependencies already installed, so you don't add a library that duplicates one.
- The existing test setup and conventions.

Only start writing code once you understand the above. If the application
already implements part of this feature, extend it rather than replacing it.

## Implementation Instructions

1. Find every path by which an image becomes visible to someone other than its uploader, including avatars, attachments, embeds, and public links, and put the check on all of them. A single unguarded upload path defeats the feature.
2. Express policy as named categories with their own thresholds, each mapped to an action. A single unsafe score gives operators no way to be strict about one category and permissive about another.
3. Hold anything above the review threshold in a state where the uploader can still see their own file but nobody else can, and reuse the moderation queue, action vocabulary, and appeal flow shared with AI Text Content Moderation rather than building a second one.
4. Record every decision with the image reference, the categories and scores, the action taken, and who or what took it, and keep that record long enough to handle an appeal. Do not retain the image itself beyond the app's normal storage rules for the sake of the audit trail.
5. Do not surface category names, scores, or thresholds to the uploader. Tell them the content was held for review and how to appeal; publishing the classifier's reasoning teaches people how to get around it.

## UI and UX Requirements

Match the application's existing design system exactly. Reuse its components,
spacing, and typography. This feature should look like it was always there.

## Responsive Requirements

Works on mobile, tablet, and desktop. Touch targets are large enough to hit on a
phone, and nothing overflows horizontally at 320px.

## Accessibility Requirements

- Fully keyboard navigable.
- Correct semantic elements and ARIA roles.
- Visible focus states.
- Meets WCAG AA contrast.
- Dynamic changes are announced to screen readers.
- Respects prefers-reduced-motion.

## Edge Cases

- Per-category thresholds tuned to the app's actual policy are required. One generic score forces a choice between blocking harmless content and admitting content the policy forbids.
- Scores in the uncertain band must go to a human queue rather than defaulting to allow or to block, and that queue needs a service-level expectation so held content does not sit for a week.
- Content held for review must be genuinely inaccessible to everyone except the uploader and reviewers, including through direct storage links, cached copies, and thumbnails already generated.
- False positives are certain. Give the uploader a stated appeal path, let a reviewer overturn a decision, and make the overturn restore visibility without requiring a re-upload.
- Do not expose category names, confidence scores, or thresholds in responses, client code, or error messages, because each one is a hint for evading the check.
- When the moderation provider is unavailable or times out, choose a documented failure posture — hold for review rather than publish — and make sure held items are re-checked automatically when service returns.
- Rate limits and large batches need backoff and a queue on the app's existing background-job system, with the uploader told their image is pending rather than shown an error.
- Re-uploading the same file must reach the same decision without a second classification, and a policy change must be able to trigger a re-check of previously cleared content.

## Testing

Exercise the feature end to end in the running application. Cover every edge case
above, then run the existing test suite and confirm nothing regressed.

## Acceptance Criteria

- [ ] Every path that publishes an image runs the check before the image is visible to others.
- [ ] Policy is expressed as named categories with independently configurable thresholds and actions.
- [ ] Held images are inaccessible to everyone but the uploader and reviewers, including through direct storage and thumbnail URLs.
- [ ] Uncertain results reach a human review queue rather than defaulting to allow or block.
- [ ] Uploaders have an appeal path and an overturned decision restores visibility without a re-upload.
- [ ] Category names, scores, and thresholds are never returned to the uploader.
- [ ] A provider outage holds content for review and re-checks it automatically when service returns.
- [ ] The feature matches the existing design system.
- [ ] No existing functionality is broken.

## Adaptation Rules

- Match the existing design system. Do not introduce a new colour palette,
  spacing scale, or component library.
- Reuse existing components and utilities wherever they fit.
- Follow the naming, file layout, and code style already present.
- Do not upgrade, replace, or remove existing dependencies to make this
  feature fit. Adapt the feature to the app, not the app to the feature.
- Do not break existing functionality. If a change is genuinely required in
  existing code, make the smallest one that works and say so.
- If something in these instructions conflicts with how the application is
  built, follow the application and explain the deviation.

## Final Verification

Before you report the work as done:

1. Re-read the acceptance criteria above and check each one against what you
   actually built.
2. Run the application and exercise the feature end to end.
3. Run the existing test suite and confirm you have broken nothing.
4. Check the feature on mobile, tablet, and desktop widths.
5. Check keyboard navigation and focus handling.
6. Summarize what changed: files added, files modified, and anything you
   deliberately did differently because of how this application is built.

If any acceptance criterion is unmet, fix it before reporting completion.