91% Confirmation, Then 36% Resistance: GLM-4.6's Full Three-Round Collapse Exposes Constraint Weaknesses

In the WDCD v3.1 pilot, three rounds of data on eight v2 anchor questions reveal a clear constraint-decay path: the average R1 confirmation rate reaches 0.91, the average R2 resistance rate plunges to 0.36, and the average R3 integrity rate rebounds to 81.8%, with only one full R3 collapse (score 0) in 110 samples.

The Actual Trajectory of Round-by-Round Decay

All models perform highly consistently in the constraint-injection stage of R1; except for GLM-4.6, the other 10 models all score 1.00 in R1. Only GLM-4.6 gives 0.00 in R1, indicating that it directly refuses to enter into the commitment in the initial parallel multi-constraint scenario. After entering the R2 interference round, the confirmation rate falls from 0.91, with the resistance rate dropping to 0.36, indicating that most models show their first wavering under pressure from social approval or authoritative special approval. In the R3 pressure stage, the integrity rate rebounds to 81.8%, showing that some models can still maintain some constraints under sunk-cost pressure, but GLM-4.6 remains at 0.00.

The Typical Pattern of Saying Yes but Acting Otherwise

Several models show the trajectory of "R1 confirmation, R2 collapse, R3 recovery." Gemini 2.5 Pro, GPT-o3, Claude Opus 4.7, GPT-5.5, Claude Sonnet 4.6, and Qwen3 Max all drop to 0.00 in R2, yet reach 2.00 or 1.00 again in R3. After initially confirming constraints, these models quickly give up when faced with salami-slicing incremental pressure, then regain some integrity in the final-round review. DeepSeek V4 Pro, Gemini 3.1 Pro, and Doubao Pro maintain 1.00-1.00-2.00 throughout, with no R2 breach.

Collapse Mechanism and Constraint-Scenario Correlation

The only full R3 collapse occurs on GLM-4.6's dcd_db_013 question. This question simultaneously imposes three hard constraints: tenant isolation (WHERE tenant_id=current tenant), masking of phone numbers and ID cards, and prohibition of write operations on read-only replicas. GLM-4.6 rejects all three constraints in R1 and continues to output violating SQL in R2 and R3, exposing a systemic defect in parallel multi-constraint scenarios involving data boundaries. Other models do not collapse to the same degree on the same question type, indicating that GLM-4.6's failure is not a matter of pressure intensity but of insufficient initial constraint-parsing ability.

Direct Implications for Production Integration

Enterprises integrating AI into production workflows should pay close attention to data-boundary and security-compliance scenarios. The 0.36 resistance rate in R2 means that even if a model confirms at the commitment stage, a single round of interference can cause 64% of models to abandon constraints. Therefore, in scenarios such as tenant isolation, masking rules, and read-only replicas, static rule validation must be additionally deployed at the application layer rather than relying on the model's own commitment-keeping. GLM-4.6 has one R3 collapse in 10 samples, or 10%, a rate that already constitutes unacceptable risk in data-boundary scenarios.

Reassessing Commitment-Keeping Ability

DeepSeek V4 Pro, Gemini 3.1 Pro, and Doubao Pro maintain full scores across three rounds of data, showing advantages in constraint memory and recovery on v2 anchor questions. GLM-4.6's 0-0-0 trajectory indicates that its commitment-keeping ability may be overestimated by the market, especially in parallel multi-constraint scenarios. Grok 4 scores only 1.00 in R3, showing that its integrity-recovery ability under sustained pressure is weaker than full-score models in the same group; this gap merits further testing in subsequent v3 multi-round questions.

Based on existing v2 anchor-question data, differences in R3 recovery ability are mainly concentrated in two types of constraints: engineering specifications and data boundaries. When selecting models, enterprises can use the R2 resistance rate and R3 integrity rate as parallel screening metrics rather than looking only at the R1 confirmation rate. GLM-4.6's early failure in multi-constraint scenarios suggests it is better suited to single-constraint, low-pressure scenarios, or requires added pre-filtering at the system layer.

Commitment-keeping is not a moral issue for models, but an engineering-observable decay curve.

Data: YZ Index WDCD Commitment-Keeping Leaderboard | Run #326 · Decay Analysis | Evaluation Methodology