Última actividad 1795213390

Revisión 6e997fac1538f0137e9074c8087c760aaaf5fe54

Dual-Model Visual Verification.md Sin formato

2. Dual-Model Visual Verification

Introduce independent visual cross-comparison between DeepSeek and GLM 5.3 Flash.

This is particularly relevant because Drax makes extensive use of Playwright and screenshot-based verification.

Proposed Workflow

  1. Playwright (or whatever tool we're using to automate) performs browser interactions.
  2. Screenshots and visual evidence are captured.
  3. DeepSeek independently analyses the screenshots.
  4. GLM 5.3 Flash independently analyses the same screenshots.
  5. Both models produce structured findings.
  6. DeepSeek receives GLM's findings and performs a cross-comparison.
  7. DeepSeek retains final decision-making authority.

DeepSeek – Primary Visual Judge

  • Inspect screenshots.
  • Identify visual regressions and usability problems.
  • Assess layout, spacing, alignment and consistency.
  • Review GLM's findings.
  • Challenge unsupported observations.
  • Produce the final visual ruling.

GLM 5.3 Flash – Independent Reviewer

  • Independently inspect the same screenshots.
  • Identify issues DeepSeek may have overlooked.
  • Highlight layout, typography and visual hierarchy problems.
  • Identify potential inconsistencies and regressions.
  • Provide evidence for each finding.
  • Avoid simply agreeing with DeepSeek.

Important Principle

GLM must not see DeepSeek's initial conclusions before producing its own assessment.

This prevents anchoring and encourages genuinely independent analysis.

DeepSeek subsequently receives both assessments and the original visual evidence.

It must not blindly accept GLM's findings. It should reject unsupported observations where appropriate.

DeepSeek retains the final say.

Suggested Finding Format

Each model should report:

  • Finding ID
  • Screenshot reference
  • Affected element or region
  • Description of the issue
  • Expected appearance
  • Actual observed appearance
  • Severity
  • Confidence
  • Supporting evidence

Cross-Comparison Decision Matrix

DeepSeek GLM Required Action
Finds issue Finds issue High-confidence candidate finding; verify evidence
Finds issue Doesn't find issue DeepSeek reviews its finding
Doesn't find issue Finds issue DeepSeek re-examines the screenshot
Neither finds issue Neither finds issue Accept provisionally, subject to other tests

The purpose is not majority voting. The purpose is to identify overlooked defects and improve the accuracy of the final visual ruling.