DraxCodes revisó este gist . Ir a la revisión
1 file changed, 70 insertions
Dual-Model Visual Verification.md (archivo creado)
| @@ -0,0 +1,70 @@ | |||
| 1 | + | ## 2. Dual-Model Visual Verification | |
| 2 | + | ||
| 3 | + | Introduce independent visual cross-comparison between DeepSeek and GLM 5.3 Flash. | |
| 4 | + | ||
| 5 | + | This is particularly relevant because Drax makes extensive use of Playwright and screenshot-based verification. | |
| 6 | + | ||
| 7 | + | ### Proposed Workflow | |
| 8 | + | ||
| 9 | + | 1. Playwright (or whatever tool we're using to automate) performs browser interactions. | |
| 10 | + | 2. Screenshots and visual evidence are captured. | |
| 11 | + | 3. DeepSeek independently analyses the screenshots. | |
| 12 | + | 4. GLM 5.3 Flash independently analyses the same screenshots. | |
| 13 | + | 5. Both models produce structured findings. | |
| 14 | + | 6. DeepSeek receives GLM's findings and performs a cross-comparison. | |
| 15 | + | 7. DeepSeek retains final decision-making authority. | |
| 16 | + | ||
| 17 | + | ### DeepSeek – Primary Visual Judge | |
| 18 | + | ||
| 19 | + | - Inspect screenshots. | |
| 20 | + | - Identify visual regressions and usability problems. | |
| 21 | + | - Assess layout, spacing, alignment and consistency. | |
| 22 | + | - Review GLM's findings. | |
| 23 | + | - Challenge unsupported observations. | |
| 24 | + | - Produce the final visual ruling. | |
| 25 | + | ||
| 26 | + | ### GLM 5.3 Flash – Independent Reviewer | |
| 27 | + | ||
| 28 | + | - Independently inspect the same screenshots. | |
| 29 | + | - Identify issues DeepSeek may have overlooked. | |
| 30 | + | - Highlight layout, typography and visual hierarchy problems. | |
| 31 | + | - Identify potential inconsistencies and regressions. | |
| 32 | + | - Provide evidence for each finding. | |
| 33 | + | - Avoid simply agreeing with DeepSeek. | |
| 34 | + | ||
| 35 | + | ### Important Principle | |
| 36 | + | ||
| 37 | + | GLM must not see DeepSeek's initial conclusions before producing its own assessment. | |
| 38 | + | ||
| 39 | + | This prevents anchoring and encourages genuinely independent analysis. | |
| 40 | + | ||
| 41 | + | DeepSeek subsequently receives both assessments and the original visual evidence. | |
| 42 | + | ||
| 43 | + | It must not blindly accept GLM's findings. It should reject unsupported observations where appropriate. | |
| 44 | + | ||
| 45 | + | **DeepSeek retains the final say.** | |
| 46 | + | ||
| 47 | + | ### Suggested Finding Format | |
| 48 | + | ||
| 49 | + | Each model should report: | |
| 50 | + | ||
| 51 | + | - Finding ID | |
| 52 | + | - Screenshot reference | |
| 53 | + | - Affected element or region | |
| 54 | + | - Description of the issue | |
| 55 | + | - Expected appearance | |
| 56 | + | - Actual observed appearance | |
| 57 | + | - Severity | |
| 58 | + | - Confidence | |
| 59 | + | - Supporting evidence | |
| 60 | + | ||
| 61 | + | ### Cross-Comparison Decision Matrix | |
| 62 | + | ||
| 63 | + | | DeepSeek | GLM | Required Action | | |
| 64 | + | |---|---|---| | |
| 65 | + | | Finds issue | Finds issue | High-confidence candidate finding; verify evidence | | |
| 66 | + | | Finds issue | Doesn't find issue | DeepSeek reviews its finding | | |
| 67 | + | | Doesn't find issue | Finds issue | DeepSeek re-examines the screenshot | | |
| 68 | + | | Neither finds issue | Neither finds issue | Accept provisionally, subject to other tests | | |
| 69 | + | ||
| 70 | + | The purpose is not majority voting. The purpose is to identify overlooked defects and improve the accuracy of the final visual ruling. | |
Siguiente
Anterior