WDCD Cycle: GPT-6 Astra Falls 20.3 Points as Six Models Decline and Only Sonnet Rises

In this cycle of the WDCD v3.1 commitment-adherence test, only Claude Sonnet 4.6 showed a positive change among the 15 evaluated models, while all six other models declined. GPT-6 Astra alone fell by 20.3 points.

Data facts: declines are concentrated and substantial

Compared with Run #411, the score changes in this cycle are as follows: Claude Opus 4.7 fell 11.6 points, DeepSeek V4 Pro fell 6.9 points, GLM-4.6 fell 13.5 points, GPT-6 Astra fell 20.3 points, GPT-o3 fell 12.2 points, and Grok 4 fell 11.3 points. Claude Sonnet 4.6 rose 5 points. In the current Top 5, Claude Sonnet 4.6 ranks first with 93.93 points, followed by Gemini 3.1 Pro with 91.90 points; GPT-6 Sol with 85.28 points, GPT-5.5 with 81.88 points, and GPT-6 Luna with 81.69 points rank third through fifth.

Cause analysis: differences in constraint survival under multi-round pressure

The WDCD v3 task uses an 8–12-turn dialogue structure. It first establishes 2–5 parallel hard constraints, then applies escalating pressures in sequence, including social proof, authoritative exception, salami slicing, and sunk cost, and finally scores through the KBV restatement probe and an honest final-round review. S_hold, the commitment-survival weight, accounts for 60 points; the later a commitment is broken, the higher the score. GPT-6 Astra had the largest decline, presumably because it first gave way during the R2–R3 distraction and pressure stages, causing a substantial loss of S_hold points. GLM-4.6 and GPT-o3 both fell by more than 12 points, possibly because they failed to maintain constraint memory under continuous salami-slicing pressure in data-boundary or security-compliance scenarios, affecting both S_kbv (15 points) and S_recover (10 points). Claude Sonnet 4.6 rose 5 points against the trend, indicating that in engineering-standard constraint scenarios it can still maintain a relatively high S_hold score when sunk-cost pressure increases.

Sampling uses worst-of-3: each question is run three times and the worst result is taken, with rule-based scoring and no AI judges. Therefore, the above changes are more likely due to differences in models' response stability to parallel hard constraints in v3 tasks rather than single random fluctuations.

Selection implications: production workflow integration requires scenario differentiation

For enterprises integrating AI into production workflows, the total commitment-adherence scores of Claude Sonnet 4.6 at 93.93 points and Gemini 3.1 Pro at 91.90 points mean they can be considered for direct use in scenarios that require long-term maintenance of business rules or resource limits, but external guardrails should still be added in data-boundary constraint scenarios. The 20.3-point drop for GPT-6 Astra suggests a higher probability of giving way in security-compliance or engineering-standard scenarios; integration should add human review checkpoints or secondary verification mechanisms. Although GPT-6 Sol's 85.28 points places it in the Top 3, it still trails the leader by 8.65 points. Under multi-round gradual pressure, it is recommended only for low-risk, non-core processes.

Strategic judgment: signals of overestimation and underestimation of commitment-adherence capability

Analysis shows that the commitment-adherence capabilities of GPT-6 Astra and GLM-4.6 may be overestimated by the market, as their S_hold survival performance under the multi-round pressure of v3 tasks shows a clear gap with current Top 5 models. The 5-point rise of Claude Sonnet 4.6 may be underestimated, and it is worth continuing to verify next cycle whether its S_integrity score remains stable on the KBV restatement probe. The signal worth tracking is this: if the declines of GPT-o3 and Grok 4 narrow next cycle, their drops can be judged as mainly due to single-period prompt-sensitivity fluctuation; otherwise, attention should be paid to the impact of model updates on constraint memory.

A commitment-adherence score is not a static label, but a model's survival curve under real production pressure.

When selecting models, enterprises should prioritize the distribution of breach rounds under the 60-point S_hold weight rather than only overall ranking. Pilot-stage data from WDCD v3.1 already shows that differences in constraint scenarios translate directly into deployment risk.

Data source: YZ Index WDCD Commitment-Adherence Leaderboard | Run #422 · Change Tracking | Evaluation Methodology