GMP Bench
← Back to Test Cases

EM Excursion Investigation Report

Task CompletionEm Reportmedium
System Prompt
You are a Quality Assurance specialist in a pharmaceutical
manufacturing facility operating under EU GMP and FDA regulations.
User Prompt
An action-level excursion was detected during environmental monitoring
in Grade A Fill Room 1 on January 16, 2026. Write a deviation investigation
report that includes:
- Deviation description and classification
- Immediate actions taken
- Root cause investigation (use Ishikawa/fishbone approach)
- Impact assessment on product batches
- CAPA (Corrective and Preventive Actions)
- Conclusion and approval section
Rubric
DimensionWeightCriteria
Investigation Rigor30%Is the root cause analysis thorough? Does it consider multiple potential causes? Is the Ishikawa/fishbone approach properly applied?
Regulatory Compliance25%Does it follow GMP deviation handling requirements? Are batch impact assessments properly documented? Are timelines mentioned?
CAPA Quality25%Are corrective actions specific and immediately implementable? Are preventive actions designed to prevent recurrence? Are effectiveness checks defined?
Professional Quality20%Is the format appropriate for a GMP deviation report? Is the language precise and unambiguous?
Source Materials
  • excursion_data

Cross-Model Comparison

ModelScoreLatencyTokens InTokens Out
Qwen3.8-FlashN/AN/A--
GLM-5.3-FlashN/AN/A--
GLM-5.3N/AN/A--
Muse Glimmer 30BN/AN/A--
Muse Glimmer 30BN/AN/A--
MiniMax M3N/AN/A--
Qwen3.8 Max (2.4T-A95B)N/AN/A--
Qwen3.8 Max (2.4T-A95B)N/AN/A--
Qwen3.8 Max (2.4T-A95B)N/AN/A--
Qwen3.8-FlashN/AN/A--
Qwen3.8-FlashN/AN/A--
GLM-5.3N/AN/A--
GLM-5.3N/AN/A--
GLM-5.3-FlashN/AN/A--
GLM-5.3-FlashN/AN/A--
Hy4 PreviewN/AN/A--
MiniMax M3N/AN/A--
Hy4 PreviewN/AN/A--
Hy4 PreviewN/AN/A--
Muse Glimmer 30BN/AN/A--
MiniMax M3N/AN/A--
Claude Opus 4.892.4%57.4s4394,287
Qwen3.6 27B91.5%86.2s3405,299
MiniMax M2.791.5%95.1s2906,380
Claude Opus 4.690.9%186.4s3338,303
DeepSeek V4 Pro89.8%49.3s2822,973
Claude Sonnet 4.689.6%152.4s3338,388
Qwen3.6 35B A3B89.5%32.6s3405,285
Claude Haiku 4.589.5%57.2s3326,965
Qwen3.5-397B-A17B88.6%59.2s3404,168
GLM-5.387.5%162.9s29520,727
Gemini 3.1 Pro87.3%46.8s3223,903
Hy4 Preview87.1%291.7s30319,350
Qwen3.8 Max (2.4T-A95B)86.6%367.8s37813,875
DeepSeek-V3.285.9%80.0s2821,945
Qwen3.8 Max (2.4T-A95B)85.0%349.2s37813,812
GPT-5.6 Terra84.8%66.3s2904,454
Qwen3.8 Max (2.4T-A95B)84.8%325.1s37812,733
GPT-5.6 Luna84.8%39.8s2903,812
Qwen3.8 27B84.6%158.2s37812,024
Qwen3.8-Flash84.3%186.3s37811,124
Qwen3.8 27B84.2%347.1s37814,856
GPT-5.6 Sol83.7%66.3s2904,591
GPT-5.6 Terra83.5%64.5s2904,218
GPT-5.6 Luna83.4%46.6s2904,071
DeepSeek-R183.4%54.2s2821,965
Qwen3.5-35B-A3B83.4%30.6s3424,056
Claude Sonnet 583.4%62.2s4395,835
Qwen3.8-Flash83.3%151.8s37810,774
Hy4 Preview83.2%214.8s30315,354
GLM-5.3-Flash83.1%303.2s29512,130
Muse Glimmer 30B83.0%42.3s3032,819
Claude Sonnet 583.0%56.4s4394,861
GLM-5.3-Flash82.3%377.3s29511,221
GLM-5.281.8%32.2s2882,165
MiniMax M381.5%64.1s4446,136
Gemini 3 Flash81.5%8.5s3201,286
DeepSeek-V3.280.8%46.0s2822,152
Qwen3.6 35B A3B80.7%25.4s3405,144
Qwen3.5-35B-A3B80.6%88.2s3404,626
GPT-5.4 mini80.5%12.3s2902,194
GPT-5.480.5%48.4s2903,345
Claude Sonnet 580.3%61.8s4395,641
GPT-5.6 Sol80.1%71.8s2904,806
GLM-5.279.5%50.1s2952,524
Qwen3.6 35B A3B79.4%20.1s3404,805
Qwen3.6 27B78.9%169.0s3405,495
Gemma 4 31B IT78.4%68.5s3521,754
Mistral Large 3 675B78.1%48.9s3253,107
Muse Glimmer 30B77.8%44.2s3023,087
Muse Glimmer 30B77.8%45.5s3033,037
MiniMax M2.777.8%61.3s2874,984
Qwen3.5-35B-A3B77.1%31.4s3403,975
DeepSeek-R177.1%150.8s2832,956
Qwen3.8 27B76.8%165.6s37812,529
Gemma 4 26B A4B IT76.6%12.0s3381,537
DeepSeek-V3.276.6%47.2s2821,536
GLM-5.376.1%24.5s2962,818
GPT-5.6 Sol75.8%63.1s2904,292
GLM-5.375.7%39.1s2964,401
GLM-5.275.4%38.4s2952,606
GLM-5.3-Flash75.1%449.5s29514,294
GPT-5.4 nano75.1%18.6s2903,105
Qwen3.6 35B A3B74.9%38.0s3404,748
GLM-5.274.3%89.8s2953,105
Qwen3.6 27B74.2%145.4s3405,344
DeepSeek-R174.1%166.7s2833,300
Qwen3.6 27B73.7%157.4s3405,765
Hy4 Preview73.4%241.8s30316,357
Qwen3.5-35B-A3B73.1%50.5s3403,931
DeepSeek-V3.273.1%53.2s2821,421
Claude Opus 573.0%277.5s43920,669
GPT-5.6 Terra72.7%53.3s2903,569
Qwen3.5-397B-A17B72.5%117.6s3404,367
Kimi K372.0%364.2s37511,182
DeepSeek V4 Pro72.0%95.5s2824,137
DeepSeek V4 Flash71.9%29.8s2823,357
Gemma 4 26B A4B IT71.9%28.8s3381,479
Gemma 4 26B A4B IT71.1%12.4s3381,502
Qwen3.8-Flash70.6%268.4s37819,475
Gemma 4 31B IT70.6%51.6s3371,565
DeepSeek V4 Pro70.6%56.9s2823,378
Mistral Small 260370.5%8.1s3371,376
Gemma 4 26B A4B IT70.3%20.2s3371,500
Gemma 4 31B IT69.8%51.6s3371,586
Gemma 4 31B IT69.3%61.8s3371,692
Gemini 3.1 Flash-Lite68.9%7.0s322969
Mistral Large 3 675B68.8%40.6s3252,913
DeepSeek V4 Pro68.1%93.0s2825,096
Qwen3.5-397B-A17B67.3%55.9s3404,550
Claude Opus 567.3%278.3s43921,334
Claude Opus 567.1%301.5s43922,961
Kimi K366.3%531.6s37516,233
Mistral Large 3 675B66.3%59.4s3252,983
MiniMax M366.0%33.9s3924,392
GPT-5.6 Luna65.8%40.1s2903,592
Llama 4 Maverick65.8%43.2s2891,006
DeepSeek-V3.265.6%39.6s2821,810
Kimi K365.0%345.8s37511,058
MiniMax M364.8%99.9s4444,659
Qwen3.5-397B-A17B64.6%46.9s3404,157
MiniMax M2.764.5%85.7s2877,368
MiniMax M2.764.1%103.0s2878,687
DeepSeek V4 Flash62.7%59.1s2824,049
DeepSeek-R1-Distill-Qwen-32B62.3%93.7s3201,688
Llama 3.3 70B Instruct62.3%7.8s295881
DeepSeek-R161.8%98.3s2831,724
Mistral Small 260361.5%21.7s3372,245
DeepSeek V4 Flash60.4%81.3s2822,791
Claude Haiku 4.560.3%72.2s3329,054
Claude Haiku 4.559.8%97.9s33212,224
Mistral Large 3 675B59.6%32.2s3252,756
Mistral Small 260357.8%16.4s3371,724
Llama 4 Maverick57.5%39.8s289873
Claude Haiku 4.557.5%76.0s3329,670
Llama 4 Scout57.0%11.2s289855
Llama 4 Scout56.3%26.4s289818
Llama 3.3 70B Instruct55.8%12.3s2951,016
Mistral Small 260354.5%16.2s3372,044
Llama 4 Scout51.7%25.1s289873
Llama 4 Maverick51.2%17.4s289902
Llama 4 Maverick51.0%42.2s289930
Llama 3.3 70B Instruct50.0%28.7s314851
Llama 4 Scout46.5%23.2s289730
Llama 3.3 70B Instruct44.8%12.1s295967
DeepSeek V4 Flash32.9%70.5s28216,484

Tags

environmental_monitoringdeviationinvestigation