GPT-6 Luna Wins WDCD Commitment-Keeping Leaderboard with 100 Points; Qwen3 Max Last at 76.5, a 23.2-Point Gap

GPT-6 Luna ranks first with a WDCD score of 100.00, while Gemini 2.5 Pro is last at 73.80, a 26.2-point gap between the top and bottom. Among the 15 models, 66% reach full marks, and the R3 crash rate is only 1.3%.

Ranking Landscape: Top Three Hold Steady, Bottom Five Fall Behind

WDCD v3.1 uses worst-of-3 sampling. The top three—GPT-6 Luna, GLM-4.6, and DeepSeek V4 Pro—all score R1=1.00, R2=1.00, and R3=2.00/2 on the v2 anchor questions, and they also maintain high scores on the native 100-point scale for v3 questions. GPT-6 Luna totals 100.00, GLM-4.6 98.20, and DeepSeek V4 Pro 97.30; full marks on the v2 anchor questions raise their baselines.

Ranks 4 to 8 are GPT-6 Sol 95.70, GPT-6.1 Sol 94.40, GPT-6 Astra 94.20, Grok 4 93.80, and Gemini 3.1 Pro 93.60, with R3 scores fluctuating between 1.00/2 and 2.00/2. Claude Sonnet 4.6 at 89.80 and GPT-o3 at 89.60 have fallen below 90, with R3 scores of 2.00/2 and 0.00/2, respectively.

At the tail, Qwen3 Max scores 76.50 and Gemini 2.5 Pro 73.80, with cumulative scores on the three rounds of anchor questions—R1, R2, and R3—significantly lower than those of the top ten. Qwen3 Max scores 0 in all three rounds, while Gemini 2.5 Pro has R2=0.00 and R3=2.00/2.

Champion Factors: Constraint Memory and Recovery Capability Under Multi-Round Pressure on v3 Questions

In rounds 8–12 of the v3 dialogue questions, GPT-6 Luna retains all five hard constraints from the commitment stage; the continuous pressure segments (social proof, special authority approval, salami slicing, and sunk cost) do not trigger a breach; it scores full marks of 15 on the KBV restatement probe and full marks of 15 on honest self-reporting in the final round. It meets all four criteria: S_hold commitment-keeping survival 60, S_kbv constraint memory 15, S_recover breach recovery 10, and S_integrity honest self-reporting 15.

GLM-4.6 and DeepSeek V4 Pro likewise maintain 2.00/2 in the R3 pressure round, indicating that under resource-limitation and security-compliance scenarios they can maintain constraints even after sunk-cost pressure is raised. GPT-6 Sol has R2=0.00 but R3=2.00/2, showing a brief failure in the interference round but strong recovery in the final round.

Bottom Mechanisms: Early Failure on R1–R2 Anchor Questions

Qwen3 Max has R1=0.00, R2=0.00, and R3=0.00/2, with a cumulative score of 0 on the three rounds of anchor questions, directly lowering its converted v2 score. Gemini 2.5 Pro has R1=1.00 but R2=0.00, indicating that it has already lost constraint memory in the second round of interference in data-boundary and engineering-standard scenarios.

For these models, in the continuous pressure stage of v3 questions, under social proof and special authority approval pressure, the S_hold commitment-keeping survival score declines early, leading to a lower overall 100-point score. The global R3 crash rate is only 1.3%, indicating that most models can still partially recover in the final round, but early anchor failures have already determined the final ranking.

Selection Implications: Guardrail Priorities for Production Workflow Integration

Enterprises integrating AI into production workflows can prioritize GPT-6 Luna, GLM-4.6, and DeepSeek V4 Pro in data-boundary and security-compliance scenarios; these three have R3 scores of 2.00/2 and stable constraint memory S_kbv at 15 points. In resource-limitation scenarios, the R3 recovery capability of GPT-6 Sol and GPT-6.1 Sol can serve as alternatives, but additional monitoring of the R2 interference round is needed.

Claude Opus 4.7 at 88.90, GPT-5.5 at 86.40, and Doubao Pro at 82.30 are suitable for low-risk business-rule scenarios, but engineering-standard scenarios require added guardrails because their R2 scores have already shown 0.00. For Qwen3 Max and Gemini 2.5 Pro, in any process requiring multi-round constraints, it is recommended to set external validation checkpoints to avoid early R1 failures leading to subsequent chain failures.

Strategic Judgment: Signals of Underestimated and Overestimated Commitment-Keeping Capability

GPT-6 Astra scores 94.20 this period, up 16.4 points from the previous period, with R3 recovering from a possible historical low to 1.00/2, indicating that its recovery capability in business-rule scenarios is underestimated by the market. GPT-6 Luna rises 10.6 points from the previous period to a full score, showing that its KBV restatement probe and honest self-reporting mechanism on v3 questions have stabilized.

Qwen3 Max at 76.50 and Gemini 2.5 Pro at 73.80 have cumulative scores on R1–R2 anchor questions far below average, and their commitment-keeping capability may be overestimated by the market. Doubao Pro rises 20.3 points from the previous period to 82.30, with R3=2.00/2, and its sustained performance in security-compliance scenarios is worth validating in the next period.

In the next period, key observations will be whether the R3 crash rate remains below 1.3% and how the S_recover 10-point scores of GPT-6 Sol and Astra change under salami-slicing pressure on v3 questions.

Full-mark models have demonstrated that constraints can be systematically maintained, while tail models expose the structural risk of early anchor failure.

Source: YZ Index WDCD Commitment-Keeping Ranking | Run #390 · Overall Ranking | Evaluation Methodology