The planned workflow
Trace every decision back to its evidence.
The database foundation is complete. The steps below describe intended product capabilities; imports, retrieval, LLM execution, workspace screens, and comparison are not available yet.
- 01 / Planned
Freeze a version
Keep dataset membership, dev/test splits, and human ground truth (GT) fixed. Corrections create a new version so later comparisons use the same evidence.
- 02 / Planned
Import detector outputs manually
Bring existing CLIP or CleanVision exports into a compatible dataset version. Preserve raw scores and source settings; automated detector execution is outside the current scope.
- 03 / Planned
Retrieve candidates
Use a text request and effective threshold/top-K settings to select candidate images from the imported outputs. Retrieval selects what to inspect; it does not establish a final defect label.
- 04 / Planned
Inspect actual images
A vision LLM examines the candidate images themselves, using recorded model and prompt settings. It returns decisions and explanations for the requested defect.
- 05 / Planned
Review the evidence
Inspect images, decisions, and explanations in shared signed-in conversations. An image may have multiple defects. Valid night scenes need judgment, not an automatic dark label.
- 06 / Planned
Compare against human GT later
Compare methods against human-verified labels on a frozen version. Unknown GT is excluded for that label. Metrics, evaluation storage, and public benchmarks are deferred.
An absent answer is not a negative label.
- Unknown GT
- The human reference has no known answer for that label, even if other labels are verified.
- Missing prediction
- A detector or model has not supplied a decision. It is not evidence that the defect is absent.
- Rejected candidate
- The LLM inspected a retrieved image and decided that the requested defect does not apply.
- Failed execution
- The run did not complete successfully. Failure remains distinct from a valid negative decision.
Account connection is available when Google sign-in is configured.