Gemini 3.1 Pro Tops WDCD Constraint-Adherence Leaderboard with 89.83, Qwen3 Max Last at 58.85

Gemini 3.1 Pro ranks first in the WDCD v3.1 constraint-adherence test with 89.83 points, while Qwen3 Max places last with 58.85 points. The equal-weight average scores of the 15 participating models show a 31-point gap between the top and bottom, mainly stemming from constraint-survival performance in the R3 pressure stage.

Ranking Landscape: Concentration at the Top and a Break at the Bottom

The top five models all score above 80: Gemini 3.1 Pro (89.83), Grok 4 (83.64), GPT-5.5 (82.46), GPT-6 Astra (81.68), and GLM-4.6 (80.48). Ranks 6 through 12 cluster in the 73–79 range, with a clear decline after rank 13: GPT-6 Luna at 69.58, Doubao Pro at 69.45, and Qwen3 Max at 58.85. The overall perfect-score rate is 41%, and the R3 collapse rate is 8.5%, indicating that most models can maintain constraints in the first two rounds, but after entering the third round of multi-turn pressure, about one in ten models suffers a full breach.

Champion Analysis: Gemini 3.1 Pro's R3 Advantage

Gemini 3.1 Pro scores R1=1.00, R2=0.88, and R3=1.75/2 on the v2 anchor questions, for a total of 89.83. Its R3 score of 1.75 exceeds Grok 4's 1.50 and GPT-5.5's 0.63, indicating stronger constraint-survival capability under sustained pressure (social proof, special authority approval, salami slicing, and sunk cost). The S_hold weight in the v3 questions accounts for 60 points; this model breaches latest in 8–12 turn dialogues, directly lifting its total score. By contrast, Qwen3 Max scores 0 in R1, R2, and R3, meaning it cannot establish effective adherence during the first-round constraint-injection stage, and all subsequent rounds fail.

Mechanism Defects of the Bottom Model

Qwen3 Max scores 0 on all v2 anchor questions, showing that it fails at the initial constraint-establishment stage in data-boundary and security-compliance scenarios. Tail models such as Doubao Pro (69.45) score only 0.75 in R1, indicating a 25% probability of breach already at the commitment stage. Claude Opus 4.7 scores 73.36 this period, down 9.7 from the previous period, mainly because its R3 score fell from a potentially higher level to 1.25, reflecting weakened recovery capability under engineering-standard constraints.

Causes of the Gap Between Top Tier and Tail

The gap most likely comes from the sustained-pressure rounds in the v3 questions and the KBV paraphrase probe. Gemini 3.1 Pro still maintains 1.75 in R3, indicating higher scores in S_recover (10 points for recovery from breach) and S_integrity (15 points for honest self-reporting). For tail models, S_hold drops directly to zero after sunk-cost pressure is raised, causing total scores to fall far behind. In the three-round design of the v2 anchor questions, R3 is weighted at 2 points; top models average above 1.2 in R3, while tail models average below 0.8, confirming the discriminative power of multi-round incremental pressure for constraint adherence.

Implications for Enterprise Selection: Practical Risks of Integrating AI into Production Workflows

Enterprises integrating AI into production workflows should prioritize models scoring above 80 on WDCD for security-compliance and data-boundary scenarios. Gemini 3.1 Pro's high R3 score means a lower probability of breach under resource limits and business-rule constraints, which can reduce manual review costs. GPT-5.5 (82.46) and GPT-6 Astra (81.68) score 0.63 and 1.13 in R3 respectively, a 0.5-point gap, suggesting that different versions in the same series perform inconsistently in engineering-standard scenarios and require separate testing before deployment. For tail models such as Qwen3 Max, R1 is already 0 in all constraint scenarios; if enterprises use them for high-risk compliance processes, they must add external rule engines and manual approval, otherwise a single breach can cause data leakage or process violations.

Strategic Judgment: Underestimated and Overestimated Signals

Grok 4's combination of 83.64 points with R2=1.00 and R3=1.50 may be underestimated by the market; its constraint-memory capability in distraction rounds is superior to most GPT variants. GPT-o3 scores 75.04 this period, down 5.4 from the previous period, with R3 at only 0.50, showing a decline in S_hold capability under multi-round pressure; this signal is worth verifying next period to see whether it is a temporary fluctuation during version iteration. Claude Sonnet 4.6 (74.24) and Claude Opus 4.7 (73.36) have close R3 scores, but the former has a lower R2, indicating systematic differences among models of different sizes from the same vendor in constraint memory. Doubao Pro rises 5.8 points this period, mainly from R3 improving to 1.25, possibly reflecting targeted optimization for security-compliance scenarios, but it still needs to be observed for stability next period.

Overall, the WDCD v3.1 results show that constraint adherence has become a key metric distinguishing production-grade models. Enterprises should not rely solely on general benchmarks; they should separately validate R3 performance against their own five categories of constraint scenarios. Gemini 3.1 Pro's lead may continue, but the more than 10-point divergence within the GPT series is a reminder that selection requires version-by-version testing.

Constraint adherence is not an add-on feature of a model but the baseline for production deployment; whoever breaches first in the R3 stage is the first to plant a compliance landmine for the enterprise.

Data source: YZ Index WDCD Constraint-Adherence Ranking | Run #428 · Overall ranking | Evaluation methodology