WDCD v3.1 Cycle: Doubao Pro Falls 22.8, GPT-5.5 Rises 11.8 as Commitment-Adherence Divergence Widens

In this cycle of the WDCD v3.1 commitment-adherence test, GPT-5.5 rose 11.8 points from the Run #365 comparison baseline to 92.60; DeepSeek V4 Pro rose 9 points to 94.00; and Claude Sonnet 4.6 rose 8.5 points. During the same period, Doubao Pro fell 22.8 points and Claude Opus 4.7 fell 6 points. Six models rose and two declined, showing clear overall divergence.

Data Facts: Specific Scores and Ranking Shifts

In the current Top 5, GLM-4.6 has WDCD=100.00, Grok 4 has WDCD=97.50, Gemini 3.1 Pro has WDCD=95.20, DeepSeek V4 Pro has WDCD=94.00, and GPT-5.5 has WDCD=92.60. Both DeepSeek V4 Pro and GPT-5.5 entered the Top 5, rising 9 points and 11.8 points respectively from the previous cycle. GPT-6 Luna rose 10.7 points, Qwen3 Max rose 10.9 points, and GPT-6 Sol rose 8.1 points. The declining Doubao Pro and Claude Opus 4.7 did not enter the Top 5.

Cause Analysis: Possible Mechanisms of Pressure Rounds and Constraint Scenarios

The WDCD v3 task uses 8–12 rounds of dialogue: it first establishes 2–5 parallel hard constraints, then escalates pressure step by step through social proof, special authority approval, salami tactics, and sunk cost, and finally conducts a KBV recitation probe and an honest final-round debrief. Doubao Pro's 22.8-point drop may stem from a sharp decline in its adherence-survival S_hold score during the sunk-cost or salami-tactics stage, dragging down its overall 100-point score. GPT-5.5's 11.8-point rise and DeepSeek V4 Pro's 9-point rise may come from gains in S_kbv constraint memory and S_recover breach-recovery scores during the R3 pressure round or the KBV recitation probe stage.

Claude Sonnet 4.6 rose 8.5 points while Claude Opus 4.7 fell 6 points; the opposite performance of different sub-models in the same series suggests that differences in adherence capability are more likely to come from prompt sensitivity in specific constraint scenarios than from overall architecture. In the three-round design of the v2 anchor task, the R3 pressure component carries a weight of 2 points; if a model's R3 score fluctuates the most in resource-limited or safety-compliance scenarios, this will directly affect the converted equal-weight average score.

Selection Implications: Practical Risk Boundaries for Production Workflow Integration

Enterprises integrating AI into production workflows can directly use WDCD scores to determine guardrail strength. For GLM-4.6, with WDCD=100.00, they may consider reducing additional rule layers in data-boundary and safety-compliance scenarios; GPT-5.5 with WDCD=92.60 and DeepSeek V4 Pro with WDCD=94.00 are suited to engineering-standard tasks but still require human review in business-rule scenarios. After Doubao Pro's 22.8-point decline this cycle, it is advisable to add independent constraint-check nodes in workflows involving multi-round incremental pressure, to avoid S_hold reaching zero too early and causing downstream workflow violations.

For scenarios requiring long-term dialogue to maintain constraints, prioritizing Top 5 models can reduce the risk of S_integrity honest self-report dropping to 0. If an enterprise uses multiple models simultaneously, it should configure additional prompt templates for declining models to prevent authority-approval-style pressure from rendering constraints ineffective.

Strategic Judgment: Underestimated and Overestimated Signals

GPT-5.5 and DeepSeek V4 Pro both entered the Top 5 this cycle with gains of more than 9 points, suggesting that their adherence-survival capability under v3 multi-round pressure may have been underestimated by the market. If subsequent versions maintain this trend, it will be worth verifying the stability of their S_recover scores in the next cycle. Doubao Pro's 22.8-point single-cycle drop was the largest; combined with the fact that it did not enter the Top 5, this suggests that the model may have a systematic weakness in constraint memory in engineering-standard or resource-limited scenarios, and it should be closely watched to see whether it rebounds in the next cycle.

GLM-4.6 held at 100.00 and did not appear in the change list, indicating that its adherence performance is the most stable under worst-of-3 sampling in the current task pool. Qwen3 Max rose 10.9 points but did not enter the Top 5, showing that although its improvement was large, its S_kbv and S_integrity scores still need further verification in business-rule scenarios.

Divergence in adherence capability has evolved from a single score into a real threshold for production deployment. The next cycle will focus on verifying the R3 pressure performance of rising models in safety-compliance scenarios.

Data source: YZ Index WDCD Commitment-Adherence Leaderboard | Run #379 · Change Tracking | Evaluation Methodology