Four Models Rise Collectively in WDCD with No Declines; Grok 4 Still Leads at 94 Points

In this WDCD v3.1 pilot, Doubao Pro, Gemini 2.5 Pro, GLM-4.6, and Qwen3 Max scored increases of 10.2, 8.1, 6.8, and 7.6 points respectively, with no other participating models recording declines. Grok 4 ranks first with 94.00 points, and GLM-4.6 ranks second with 88.93 points.

Data Facts: Increases Concentrated in Specific Constraint Scenarios

All four models rose by more than 6 points, concentrated in the later segments of rounds 8–12 in the v3 questions. GLM-4.6 climbed from its previous level to 88.93 points, entering second place in the Top 5, with DeepSeek V4 Pro following closely at 88.14 points. Claude Sonnet 4.6 and Gemini 3.1 Pro rank fourth and fifth with 86.69 and 86.48 points respectively. All rising models showed notable gains in the S_hold commitment-keeping component, which carries a weight of 60 points — the later a commitment is broken, the higher the score.

Cause Analysis: Constraint Memory Differences Under Multi-Round Pressure

The WDCD v3 question design includes continuous post-commitment pressure and KBV recitation probes. The rising models performed more stably in both resource-constraint and safety-compliance scenarios. Doubao Pro's 10.2-point increase likely stems from enhanced resistance to "salami-slicing" incremental pressure escalation, with a corresponding improvement in the 15-point S_kbv constraint-memory sub-score. Gemini 2.5 Pro rose 8.1 points, and Qwen3 Max rose 7.6 points, both showing fewer premature commitment breaches during authority-approval pressure rounds. GLM-4.6's 6.8-point increase moved it into the top two, with its 10-point S_recover breach-recovery sub-score likely improved.

In comparison, Grok 4 maintained its high position at 94.00 points, demonstrating continued advantages in sunk-cost pressure escalation and final-round honest self-reporting. The equally weighted average after v2 anchor question conversion shows that the rising models contributed more heavily in R3 pressure rounds.

Selection Implications: Practical Boundaries for Production Workflow Integration

For enterprises integrating AI into production workflows, WDCD scores directly correspond to trustworthy constraint-execution capability. Grok 4's 94.00 points indicate suitability for data-boundary and engineering-standard scenarios, usable without additional guardrails. GLM-4.6's 88.93 points and DeepSeek V4 Pro's 88.14 points are close, making them suitable for resource-constraint tasks, though safety-compliance scenarios still require secondary verification. Claude Sonnet 4.6's 86.69 points and Gemini 3.1 Pro's 86.48 points can be used for business-rule execution, but manual review checkpoints are recommended when multi-turn conversations exceed 8 rounds.

Doubao Pro's substantial rise this period signals improved commitment-keeping stability in business-rule scenarios, warranting expanded pilot scope within internal tool chains.

Strategic Assessment: Signals That Commitment-Keeping Ability Is Underrated

The pattern of four models rising collectively with zero declines shows that the current v3.1 version captures prompt-sensitivity changes more effectively. GLM-4.6's jump to 88.93 points suggests its constraint-memory capability may be underrated by the market; the next edition will need to verify its S_integrity honest self-reporting performance in engineering-standard scenarios. Grok 4's sustained high position at 94.00 points confirms its commitment-keeping ability has not been overrated.

The analysis suggests that the common factor among the rising models is improved KBV recitation probe pass rates, possibly reflecting strengthened multi-constraint parallel processing in model updates or training. When selecting models, enterprises can deploy models scoring above 94 points directly in high-risk compliance workflows, while models in the 85–90 point range require additional guardrails in business-rule scenarios.

The next edition should focus on whether GLM-4.6 and Qwen3 Max continue to improve their S_integrity scores in the final-round honest review segment of the v3 questions.

Overall, the WDCD v3.1 pilot reveals an intensifying divergence in commitment-keeping capability. Grok 4's 94.00 points remain the most reliable benchmark, while GLM-4.6's 88.93 points is building catch-up pressure. Enterprises should select models based on specific constraint-scenario types rather than relying solely on general leaderboards.

Data source: YZ Index WDCD Commitment-Keeping Leaderboard | Run #291 · Change Tracking | Evaluation Methodology