Set a quality and verification bar

For diagnose a confusing dashboard screenshot using image evidence plus written context, define quality across accuracy, completeness, usefulness, safety, and reproducibility. Weight dimensions according to harm. A cosmetic miss can be revised; an unsupported claim, broken calculation, privacy leak, or rights violation blocks release.

Translate each dimension into observable checks. Accuracy means a claim, value, behavior, or frame agrees with an authoritative source. Completeness means every required field or stage appears. Usefulness means a product analyst preparing a bug report can take the intended action. Safety includes the boundary that you must crop and redact names, faces, addresses, notifications, account IDs, location clues, EXIF metadata, and confidential browser tabs before upload. Reproducibility means the prompt, input version, settings, and review evidence are saved.

Use checks from cheapest to strongest:

1. Contract check: required sections, schema, length, and prohibited content.
2. Source check: trace claims and values to supplied evidence.
3. Edge check: run normal, boundary, missing, and adversarial cases.
4. Independent check: calculate, test, rehearse, listen, inspect, or open the original.
5. Human gate: a responsible reviewer approves consequential use.

In this chapter, the concrete verification is to zoom and compare source pixels, transcribe critical values manually, reproduce in the product, inspect logs, and reject claims unsupported by visible or supplied evidence. The expected candidate is An observation table grounded in screen regions, clearly labeled hypotheses, and targeted follow-up evidence instead of a confident visual guess. Record actual evidence, not a checkbox copied from the prompt.

Score each criterion 0 (fails), 1 (partly), or 2 (passes). Any zero for factual correctness, permission, privacy, or required disclosure is an automatic stop. A total score is useful for comparing iterations, but it must never average away a blocking defect.

Inspect this redacted checkout dashboard screenshot. Goal: explain why the conversion total appears inconsistent. First inventory only visible labels and values with locations. Separate observation from inference. Then propose three hypotheses and the next screenshot or log needed to test each. Do not identify people, infer hidden fields, or invent unreadable text.

After generation, sample beyond the happy path. Failures such as hallucinated small text; mixing observation with inference; ignoring chart scale; accidental exposure around the crop; treating an image as current truth often survive a superficial review because the output has the right shape. Use a counterexample designed to expose the riskiest assumption.

Store the rubric result, reviewer, date, input version, failed cases, and unresolved limitations. If the artifact changes, rerun affected checks. Point to regions, labels, frames, or timestamps. When the model cannot read a detail reliably, provide a crop or transcription instead of asking it to guess harder. Quality assurance is part of the work, not an apology added at the end.