Grok 4 Tops WDCD Commitment-Keeping Leaderboard with 94 Points; Doubao Pro Ranks Last with 68 Points, a 26-Point Gap

In the WDCD v3.1 commitment-keeping test, Grok 4 ranked first with 94.00 points, while Doubao Pro placed 11th with 68.17 points. The two models scored 1.50/2 and 0.88/2 respectively in the R3 pressure round, directly widening the total score gap to 25.83 points.

Ranking landscape: concentrated at the top, disjointed at the tail

The top five WDCD scores range from 86.48 to 94.00, ranks six through eight fall between 80.07 and 84.79, ranks nine and ten drop to 75.31-77.72, and 11th-place Doubao Pro at 68.17 points creates a marked gap. After adjusting for v2 anchor questions, Grok 4's R1+R2+R3 total is 3.50/4, while Doubao Pro reaches only 2.26/4, indicating that tail-end models already incur a 0.37-point R1 deduction during the initial constraint injection phase.

Champion analysis: Grok 4's R3 performance

Grok 4 achieved the highest S_hold commitment-survival score across the 17 v3 questions, earning 1.50/2 in the R3 pressure round. Under the three pressure types—social proof, authoritative approval, and salami slicing—the model held out the longest before breaking its commitments, while maintaining high levels of S_recover recovery-after-breach and S_integrity honest self-reporting. By comparison, Claude Opus 4.7 also scored a perfect 1.00 in R1 but only 1.25/2 in R3, trailing by 9.45 points overall, indicating that its constraint-maintenance ability across 8-12 consecutive dialogue rounds is weaker than Grok 4's.

Reason for last place: Doubao Pro's R1 flaw

Doubao Pro is the only model with an R1 score below 1.00, at just 0.63, meaning it cannot fully remember the 2-5 parallel hard constraints during the commitment-establishment phase. Under subsequent R2 interference and R3 pressure, its S_kbv constraint-memory score is adversely affected, leaving its total 7.14 points lower than 9th-place Qwen3 Max. In the KBV restatement probe segment of the v3 questions, Doubao Pro is most prone to constraint forgetting after rounds 6-8.

Mechanism breakdown of the leading-trailing gap

The top three models—Grok 4, GLM-4.6, and DeepSeek V4 Pro—average 1.38/2 on R3, while the bottom three—GPT-5.5, Qwen3 Max, and Doubao Pro—average only 0.88/2. The gap mainly stems from two constraint scenarios: data boundaries and safety compliance, which feature the longest pressure rounds in the v3 questions and produce the most pronounced sunk-cost pressure. In engineering-standard constraint scenarios, R3 score differences among models are relatively small, indicating that current models are fairly balanced in code-level rule adherence.

Implications for production integration selection

Enterprises integrating AI into production workflows with strict business rules and safety compliance requirements could regard Grok 4's 94.00 score and 1.50/2 R3 performance as the top choice, suitable for directly exposing multi-turn dialogue interfaces. GLM-4.6 and DeepSeek V4 Pro, scoring close to 88, can be deployed in resource-constrained scenarios, but external validation should be added for data-boundary scenarios. As for a model at 68.17 points like Doubao Pro, which loses 0.37 points at the R1 stage, it is not recommended for production workflows requiring parallel hard constraints; it should be limited to single-turn Q&A or supplemented with manual review checkpoints.

Strategic assessment: who is overvalued or undervalued

Claude Sonnet 4.6 ranks fourth (86.69 points) with an R2 score of 0.75, higher than DeepSeek V4 Pro's 0.63, but its R3 of only 0.88/2 suggests it performs well during mid-stage interference yet is more prone to breaking commitments under final-round pressure. The model's commitment-keeping ability may be overvalued by the market. Qwen3 Max ranks tenth (75.31 points) with perfect 1.00 scores in both R1 and R2, but its total is dragged down by an R3 of 1.13/2; its true constraint-maintenance ability in long conversations may be undervalued, and the next test round should focus on observing its S_hold changes across conversations exceeding 11 rounds.

Compared with the previous round, Doubao Pro improved by 10.2 points and GLM-4.6 by 6.8 points, mainly driven by better R3 scores, indicating progress in these models' recovery ability under sustained pressure. However, the overall R3 breakdown rate remains at 8.2%, meaning that even leading models have an 8.2% probability of breaking commitments in the worst sampling instance.

When selecting models, enterprises should prioritize R3 scores rather than looking only at total scores; models with R3 below 1.00/2 require additional guardrails in safety-compliance scenarios.

The next validation round should focus on sub-scores across the five constraint scenarios, especially the resource-limitation and engineering-standard categories where gaps are currently small, to observe whether new gaps emerge.

Data source: YZ Index WDCD Commitment-Keeping Leaderboard | Run #291 · Overall ranking | Evaluation methodology