Grok 4 Tops WDCD Commitment Leaderboard with 91.97 Points; Qwen3 Max Trails at 61.55 by a Wide Margin

In the WDCD v3.1 commitment test, Grok 4 became the highest scorer among 15 participating models with 91.97 points, while Qwen3 Max placed last with 61.55 points. The gap between them is 30.42 points, and Qwen3 Max scored 0 on all v2 anchor questions at the R1, R2, and R3 stages, indicating that it completely failed to maintain its commitments after constraint injection.

Ranking Landscape: Concentrated at the Top, a Break at the Bottom

The top four—Grok 4 (91.97), Claude Opus 4.7 (90.66), DeepSeek V4 Pro (90.38), and Gemini 3.1 Pro (89.76)—formed a clear first tier, with no gap among them exceeding 2.21 points. From fifth place, GPT-o3 at 87.07, scores declined progressively down to 13th place Gemini 2.5 Pro at 78.34, 14th place Doubao Pro at 73.55, and 15th place Qwen3 Max at 61.55. The overall perfect-score rate was 54.9%, and the R3 collapse rate was 9%, indicating that most models could still partially honor commitments under sustained pressure, but a few had completely failed by the third round.

Why the Champion Won: Advantage in Maintaining Constraints at R2 and R3

On the v2 anchor questions, Grok 4 scored R1=1.00, R2=1.00, and R3=1.13/2; it lost zero points at the R2 stage and lost only 0.87 points at the R3 stage. By comparison, Claude Opus 4.7 scored R2=0.75 and R3=1.50/2, already showing a 0.25-point loss at the R2 stage. DeepSeek V4 Pro scored R2=0.88 and R3=1.38/2, likewise losing 0.12 points at the R2 stage. Grok 4's perfect R2 performance shows that under escalating pressures such as social proof, special authority approval, and salami-slicing, it was better able to maintain multiple parallel hard constraints at the same time, directly raising its S_hold commitment-survival score.

Mechanism Deficiencies of the Last-Place Model

Qwen3 Max scored 0 on all three rounds of the v2 anchor questions, meaning it failed to establish effective memory at the constraint-injection stage, and naturally fell to zero during the subsequent interference and pressure stages. On the v3 questions, its WDCD score was only 61.55, far below 14th-place Doubao Pro's 73.55, indicating that it also failed to recover in the KBV restatement probe and the final-round honest review. Among the five constraint scenarios (data boundaries, resource limits, business rules, security compliance, and engineering standards), tail-end models may break down earliest in the security compliance and engineering standards scenarios, causing low scores on both S_hold and S_recover.

Selection Implications for Production Workflow Integration

Enterprises integrating AI into production workflows should focus on models scoring above 1.0 at the R3 stage. Grok 4 and Claude Sonnet 4.6 (R3=1.25/2) can still maintain relatively high constraints after sustained pressure, making them suitable for tasks involving security compliance and data boundaries. Although GPT-o3 had perfect scores at R1 and R2, its R3 was only 0.50/2; it is advisable to add external guardrails in resource-limit and business-rule scenarios. Qwen3 Max and Doubao Pro scored low across all anchor-question stages and should not be used directly in production pipelines that require long-term constraint memory.

Strategic Assessment: Signals of Underestimated and Overestimated Commitment Capability

Based on this period's data, Grok 4's commitment capability may be underestimated by the market; its zero loss at R2 and high R3 score are unique among the 15 models. Although Claude Opus 4.7 ranked second, it had already lost 0.25 points at the R2 stage, indicating that its stability under multiple rounds of gradual pressure is inferior to Grok 4's. The three models in the GPT-6 Sol series (GPT-6 Sol 85.31, GPT-6 Astra 82.28, and GPT-6 Luna 81.48) generally scored low at R2, showing that different versions within the same series differ markedly in constraint maintenance. A signal worth verifying in the next period is whether models scoring above 1.2 at R3 can maintain an S_integrity honest self-report score of no less than 12 in 8-12 round conversations on the v3 questions.

When selecting models, enterprises are advised to list models with a WDCD score above 85 and R3 greater than 1.0 as first candidates; other models require additional deployment of rule engines in security compliance scenarios. Qwen3 Max's 61.55 score and all-zero anchor performance already constitute a clear warning signal.

Honoring commitments is not an add-on feature of a model but a hard threshold for production deployment; with 91.97 points, Grok 4 proves that true leaders can still remember the original commitment in the final round.

Data source: YZ Index WDCD Commitment Leaderboard | Run #384 · Overall leaderboard ranking | Evaluation methodology