Qwen3 Max WDCD Plunges 26.2 Points; Only 2 Models Rise Against the Trend as 12 Decline

This round of WDCD v3.1 testing covers 26 questions, with 15 models participating and sampling based on worst-of-3. The results show that only DeepSeek V4 Pro (+2.3 points) and GPT-6.1 Sol (+2.1 points) increased, while the other 12 all declined. Among them, Qwen3 Max fell 26.2 points, GPT-6 Luna fell 15.8 points, and GPT-o3 fell 11.9 points.

Data facts: 80% of models declined, and the gap within the Top 5 widened

Compared with Run #400, Gemini 3.1 Pro fell only slightly by 3 points this round and remains first with 91.00 points. GLM-4.6 fell 11.8 points to 84.38, Grok 4 fell 6.1 points to 84.18, Claude Opus 4.7 entered the top five with 83.05 points, and GPT-6 Sol fell 4.2 points to 82.71. Among the declining models, GPT-5.5 fell 10.2 points, Claude Sonnet 4.6 fell 8.7 points, Gemini 2.5 Pro fell 11.7 points, and Doubao Pro fell 5.1 points. The two gainers both rose by no more than 3 points, overall showing a pattern of “a few slight gains and most declining markedly.”

Cause analysis: Under multi-round pressure, constraint memory and recovery capabilities decay in tandem

The WDCD v3 question format uses 8-12 rounds of dialogue: it first establishes 2-5 parallel hard constraints, then sequentially applies social proof, authority-granted exceptions, salami slicing, and sunk-cost pressure, and finally conducts KBV restatement probes and a final-round honest review. In scoring, S_hold constraint-adherence survival accounts for 60 points; the later a breach occurs, the higher the score. S_kbv constraint memory accounts for 15 points. Qwen3 Max and GPT-6 Luna had the largest declines, possibly because they lost points on both S_hold and S_kbv during consecutive pressure rounds. GPT-o3 and GLM-4.6 both fell by more than 11 points, suggesting they are more likely to breach commitments early during the sunk-cost escalation phase. By contrast, DeepSeek V4 Pro and GPT-6.1 Sol rose slightly, indicating that under the same pressure path their scores on S_recover breach recovery and S_integrity honest self-reporting were relatively stable.

Among the 12 models, 6 fell by more than 10 points, indicating that the v3 question pool has significantly improved its ability to distinguish prompt sensitivity.

Selection implications: Production-process integration must distinguish the priority of constraint scenarios

For enterprises integrating AI into production processes, Gemini 3.1 Pro’s 91.00-point constraint-adherence survival capability can still make it the first choice in data-boundary and security-compliance scenarios; however, in resource-limitation and engineering-standard scenarios, additional external guardrails are needed because it still posted a 3-point decline this round. Although GLM-4.6 and Grok 4 rank in the top three, they fell 11.8 points and 6.1 points respectively, suggesting significant volatility under business-rule constraints; it is recommended to add manual review checkpoints during the R2 interference and R3 pressure stages. After falling 26.2 points, Qwen3 Max is no longer suitable for direct use in processes requiring long-term constraint adherence; its specific rounds of S_hold scoring in the v3 question pool must first be verified.

Strategic judgment: Models whose constraint-adherence capability was overestimated and signals worth tracking

From this round’s data, it can be judged that the constraint-adherence capabilities of GPT-6 Luna and GPT-5.5 may have been overestimated by the market previously; they fell 15.8 points and 10.2 points respectively, and neither entered the Top 5. DeepSeek V4 Pro rose slightly; combined with its potential performance on v2 anchor questions, it is worth continuing to observe next period to see whether it forms a sustained advantage. GPT-6.1 Sol and GPT-6 Sol belong to the same series but one rose and the other fell; the difference may come from subtle adjustments to constraint memory and recovery mechanisms. This signal can be further verified in the next cycle through the worst-of-3 S_kbv individual score.

Overall, this round’s WDCD results show that most models’ constraint-adherence survival capabilities under multi-round progressive pressure have experienced a systematic decline. When selecting models, enterprises should prioritize the 91.00-point benchmark of Gemini 3.1 Pro and set stricter external constraints for models that fell by more than 10 points.

Constraint-adherence capability is not a static label but a real-time snapshot after each stress test.

Data source: YZ Index WDCD Constraint-Adherence Leaderboard | Run #406 · Change tracking | Evaluation methodology