WDCD Commitment-Keeping Leaderboard: GPT-6 Astra Tops with 90.50, Doubao Pro Last at 58.70 — a 31.8-Point Gap

In the WDCD v3.1 commitment-keeping test, GPT-6 Astra took first place with 90.50 points, while Doubao Pro finished last with only 58.70. Scores for the 15 participating models ranged from 90.50 to 58.70, a spread of 31.8 points.

Ranking Landscape: Top Five Clustered in the 90s, Bottom Five Fall Below 80

The leaderboard shows clear tier separation. GPT-6 Astra 90.50, Gemini 3.1 Pro 89.97, GLM-4.6 89.97, Claude Opus 4.7 89.30, and GPT-6 Sol 89.30 form the first tier, with all five models above 89 points. Sixth through ninth — Claude Sonnet 4.6 88.90, Grok 4 88.27, GPT-o3 86.93, and GPT-6 Luna 85.77 — remain above 85. After tenth-place GPT-5.5 at 81.27, DeepSeek V4 Pro 78.57, Gemini 2.5 Pro 77.23, GPT-6.1 Sol 77.20, and Qwen3 Max 76.50 form the third tier, while Doubao Pro 58.70 drops alone into the fourth tier.

On the v2 anchor task, all models passed R1 at a rate of 14/15, with only Qwen3 Max at R1=0.00. Nine models achieved the full 2 points on R3, and only Qwen3 Max scored 0 on R3.

The Commitment-Keeping Mechanism of Champion GPT-6 Astra

In the v2 three-round anchor task, GPT-6 Astra recorded R1=1.00, R2=1.00, and R3=2.00/2, and its converted v3 native score lifted the total to 90.50. This means that under sustained pressure stages (social proof, authority override, salami-slicing, sunk cost), it earned the highest score on the 60-point S_hold commitment-survival item, with the breach occurring latest. The v3 tasks involve 8–12 rounds of dialogue, with pressure escalating step by step after the commitment is established; Astra was stable on the 15-point S_kbv constraint memory and the 10-point S_recover breach-recovery items, and never collapsed immediately after R2 interference.

Selection implication: enterprises integrating AI into production workflows can prioritize GPT-6 Astra as the default model in data-boundary and security-compliance scenarios; its worst run under worst-of-3 sampling still holds a high score, implying a lower probability of a single anomalous trigger.

Why Last-Place Doubao Pro Falls in a Separate Tier

Doubao Pro's 58.70 is still 17.8 points below 14th-place Qwen3 Max at 76.50. Its v2 anchor results — R1=1.00, R2=1.00, R3=2.00/2 — appear normal, but its v3 native percentage score was extremely low, indicating severe losses on both S_hold and S_integrity in the multi-round progressive-pressure tasks. The likely causes cluster around engineering-specification or resource-limit constraint scenarios, where the commitment was broken too early during successive pressure rounds and never recovered.

Strategic judgment: in this period's pilot data, Doubao Pro's commitment-keeping ability is either markedly underestimated or genuinely insufficient; when connected to systems with strict business rules, it requires additional external guardrails, otherwise it carries the highest risk of collapse under R3-level pressure.

Pressure-Round Evidence for the Gap Between Leaders and Laggards

The difference in R3 collapse rates between leading and trailing models directly widens total scores. The global R3 collapse rate is only 0.7%, yet Qwen3 Max scored R3=0.00, while Doubao Pro, despite a full R3 score, was dragged down by excessively low S_recover and S_integrity scores on the v3 multi-round tasks. Claude Opus 4.7 improved 8.0 points over the previous period, mainly from better S_hold scores on v3 tasks, indicating stronger recovery during sunk-cost escalation.

Implication for production deployment: in business-rule and security-compliance scenarios, enterprises should set a secondary confirmation checkpoint for models scoring below 80 (GPT-5.5 and below) to avoid three types of pressure occurring simultaneously within a single dialogue round.

The Subtle Difference Between GLM-4.6 and Gemini 3.1 Pro Tied for Second

GLM-4.6 and Gemini 3.1 Pro both scored 89.97, but their v2 anchor performances differed: GLM-4.6 R1=1 R2=1 R3=2, versus Gemini 3.1 Pro R1=1 R2=0 R3=1. GLM-4.6 retained stronger constraint memory during the R2 interference stage, while Gemini 3.1 Pro scored lower during the R3 pressure stage. This suggests the two are complementary in applicability across different constraint scenarios.

Analysis suggests GLM-4.6 is better suited to engineering-specification scenarios requiring multi-round contextual anchoring, while Gemini 3.1 Pro needs added human review under authority-override pressure.

Comparison Signals from the Previous Period and Directions for Next-Period Validation

Claude Opus 4.7 rose 8.0 points and GPT-6 Astra rose 5.3, while GPT-5.5 fell 5.6, Qwen3 Max fell 8.5, and Doubao Pro fell 10.1. These changes all stem from fluctuations in v3 native scores rather than changes to the v2 anchor task. Qwen3 Max's R1=0.00 is the only R1 failure case and warrants focused validation next period on its initial commitment-forming ability in data-boundary constraint scenarios.

Strategic judgment: GPT-6 Astra's commitment-keeping ability is not overstated in this period's data, and the scores of Doubao Pro and Qwen3 Max may understate their potential in specific constraint scenarios, but the current worst-of-3 results already show clear risk.

The gap between 90.50 and 58.70 exposes not the size of model parameters, but the real difference in constraint memory and recovery mechanisms under sustained pressure.

Data source: YZ Index WDCD Commitment-Keeping Leaderboard | Run #411 · Overall Ranking | Evaluation Methodology