GLM-4.6 Scores Perfect 100 on WDCD, Doubao Pro Last at 62: What Does the 38-Point Gap Expose?

GLM-4.6 ranked first with a WDCD score of 100.00, while Doubao Pro ranked last with 62.00, a gap of 38 points.

Ranking Landscape: The Top Six and a Sharp Tail Drop-off

A total of 15 models participated in this WDCD v3.1 pilot, and the question pool contained 10 questions. GLM-4.6 achieved perfect scores on all v3 questions and v2 anchor questions, with R1=1.00, R2=1.00, and R3=2.00/2. Grok 4 followed with 97.50, scoring 1.00/2 on R3. Gemini 3.1 Pro with 95.20, DeepSeek V4 Pro with 94.00, GPT-5.5 with 92.60, and GPT-o3 with 92.00 ranked third through sixth, respectively.

The scores of the models ranked 7th through 13th were concentrated in the 91.80 to 77.80 range. Claude Opus 4.7 at 88.00 and Qwen3 Max at 76.50 had clearly fallen behind. Doubao Pro at 62.00 still had a 14.5-point gap with the 14th-place Qwen3 Max, forming a clear break at the tail.

Why the Champion Won: Zero Breaches Across All Rounds

GLM-4.6 achieved perfect scores on all four v3 metrics—S_hold, S_kbv, S_recover, and S_integrity—and passed all three rounds of the v2 anchor questions. Other models showed at least one score loss during the R3 stage. Grok 4 and Gemini 3.1 Pro each scored 1.00/2 on R3, indicating that each had one instance of constraint loosening during the final pressure round. GLM-4.6's zero loss indicates that under four types of escalating pressure—social proof, special authority approval, salami slicing, and sunk cost—its constraint memory and recovery capabilities remained stable.

Bottom-Rank Analysis: A Mechanism Flaw That Fails at R1

Doubao Pro scored 0 on R1 of the v2 anchor questions, meaning it did not establish effective commitment-keeping during the initial constraint injection stage. Qwen3 Max also scored 0 across R1, R2, and R3, indicating it could not maintain constraints during either the establishment or subsequent interference stages. By contrast, 12th-ranked Gemini 2.5 Pro scored R1=1.00, R3=2.00/2, and only R2=0.00, showing that its constraint-establishment ability remains, but it is fragile during the intermediate interference stage. The low scores of tail-end models mainly stem from early constraint memory failure, rather than only an R3 collapse.

Selection Implications of the Top Tier vs. the Tail Gap

For enterprises integrating AI into production processes, GLM-4.6's score of 100 shows that it can maintain the longest commitment-keeping survival time across five types of constraint scenarios: data boundaries, resource limits, business rules, safety compliance, and engineering standards. GPT-5.5 and GPT-o3 scored 92.60 and 92.00, respectively, with differences in R3 performance; they can be used in non-core business rule scenarios, but safety compliance scenarios require an extra layer of secondary checks.

Doubao Pro and Qwen3 Max scored below 70 and failed at the R1 stage, indicating they are not suitable for direct integration into processes that require long-term constraint maintenance. If enterprises must use such models, they should mandatorily add external guardrails in engineering standards and safety compliance scenarios, rather than relying on the models' own commitment-keeping ability.

Volatility Signals from Comparison with the Previous Period

Claude Sonnet 4.6 scored 89.30 this period, up 8.5 points from the previous period; DeepSeek V4 Pro rose 9.0 points to 94.00. GPT-5.5 rose 11.8 points. Doubao Pro fell 22.8 points. GPT-6 Luna rose 10.7 points. These increases and decreases all come from this period's data; no continuous historical records were provided, so trend stability cannot be judged.

Strategic Assessment: Underestimated Signals and Signals Requiring Further Study

DeepSeek V4 Pro scored R1=1.00, R2=0.00, R3=2.00/2 this period, with a score of 94.00. It may excel in the constraint-establishment and recovery stages, but has a weakness at R2, making it worth focusing on next period to examine its performance under salami-slicing pressure. Claude Opus 4.7 scored 88.00 this period, with R3=0.00/2, down 6.0 points from the previous period; its recovery ability under final-round pressure may be underestimated by the market, and it needs to be observed whether its R3 score rebounds next period.

GLM-4.6's perfect performance may stem from zero loss in the KBV paraphrase probes and the final-round honest self-report stage, but the pilot-stage sample size is limited, and whether this advantage is stable still requires follow-up study.

In this pilot, GLM-4.6 achieved zero breaches across all constraint scenarios, while Doubao Pro failed across the board at the initial constraint stage. The 38-point gap indicates that before integrating models into production environments, enterprises must distinguish the models' actual commitment-keeping boundaries.

Data source: YZ Index WDCD Commitment-Keeping Leaderboard | Run #379 · Overall ranking | Evaluation Methodology