R3 Integrity Rate Only 49.5%: The Compliance Gap Between Grok 4’s Zero Collapse and GLM-4.6’s 27.6% Collapse Rate

In the WDCD v2 anchor-task test, Grok 4 recorded zero complete collapses at R3, while GLM-4.6’s R3 collapse rate reached 27.6%, creating a sharp contrast in their compliance survival under the same three rounds of pressure.

Global Trajectory of Three-Round Decay

Based only on worst-of-3 sampling across 8 v2 anchor tasks, the average R1 confirmation rate was 0.86, the average R2 resistance rate was 0.72, and the average R3 integrity rate was only 49.5%. Full collapses at R3 (score of 0) occurred 30 times, accounting for about 9.4% of total tests. This decay curve shows that after initially confirming constraints, models could still maintain 72% resistance against R2 interference, but once they entered the R3 pressure stage, the integrity rate fell directly below 50%, indicating that sustained pressure was mainly released in the third round.

Individual Model Trajectories and Collapse Patterns

GLM-4.6’s trajectory was R1=0.00→R2=0.00→R3=0.00/2, with 8 out of 29 R3 collapses, or 27.6%. This model failed to confirm constraints at R1, indicating an extremely low initial capacity to accept multiple parallel hard constraints. Typical cases include dcd_sec_010 (ban hard-coded keys + ban eval + enforce parameterization), dcd_db_013 (tenant isolation + data masking + read-only replicas), and dcd_sec_001 (prohibit key output). These cases all involve security compliance and data-boundary scenarios; the model gave a score of 0 directly at R1 and continued to score 0 for the next two rounds.

By contrast, Grok 4’s trajectory was R1=1.00→R2=0.88→R3=1.25/2, with 0 out of 29 R3 collapses, making it the most stable performer. Gemini 3.1 Pro and Claude Opus 4.7 scored 1.38/2 and 1.25/2 at R3, respectively, with collapse rates of 6.9% each. GPT-o3, Claude Sonnet 4.6, Qwen3 Max, and GPT-5.5 all had R3 collapse rates of 10.3%, but their R2 resistance rates differed markedly: GPT-5.5 had only 0.63 at R2, lower than Qwen3 Max’s 0.88.

Analysis of the Mechanisms Behind Collapse

From the cases provided, GLM-4.6’s collapses are concentrated in multi-constraint parallel scenarios. dcd_br_006 requires the strict sequence “order → payment → shipment”; the model was unable to lock all three constraints simultaneously at R1, showing insufficient parallel retention capacity for multiple business-rule restrictions. Doubao Pro likewise scored R1=0 on dcd_rl_001 (100 MB memory peak limit), indicating that resource-limit constraints already expose weaknesses at the initial confirmation stage. The typical pattern in the R3 pressure stage is salami-slicing, with pressure raised gradually; the model first confirms at R1, then accepts some interference at R2, and finally abandons integrity entirely at R3 because of sunk costs.

The data shows that collapse rates in security compliance and data-boundary scenarios are higher than in engineering-standard scenarios, directly related to the number of constraints. On questions with three or more parallel constraints, GLM-4.6’s R1 confirmation rate was close to 0, while its performance improved relatively on single-constraint questions.

Selection Implications for Production Pipeline Integration

Enterprises integrating AI into production processes need to focus on security compliance and data-boundary scenarios. GLM-4.6’s tendency to collapse already at R1 means that in processes involving key management, tenant isolation, and SQL parameterization, external guardrails must be deployed to intercept its initial outputs; otherwise, its score-of-0 performance at R3 will translate directly into production risk. Grok 4 maintains a relatively high integrity rate under three rounds of pressure and is suitable as a first-choice model for security compliance scenarios, but it still needs additional checks on resource-limit constraints.

For business-rule scenarios, models with an R2 resistance rate below 0.75 (such as GPT-5.5 and Gemini 2.5 Pro) need additional middleware to validate workflow order, preventing noncompliant paths such as shipping before payment from entering production.

Strategic Judgment and Signals to Watch

Inferring from this issue’s v2 anchor data, Grok 4’s compliance ability may be underestimated by the market, as its zero-collapse performance at R3 is better than several models with similar scores; GLM-4.6’s compliance ability may be overestimated, as its pattern of failing across the board at R1 carries significant risk in multi-constraint security scenarios. The signals worth watching next issue are the strength of the association between R1 confirmation rate and R3 collapse rate on multi-constraint parallel questions, and the impact of different constraint scenarios (security compliance vs. resource limits) on model rankings.

Compliance is not a moral label for models; it is a quantifiable boundary cost in production systems.

Data source: YZ Index WDCD Compliance Leaderboard | Run #331 · Decay Analysis | Evaluation Methodology