All Five Dimensions Missing in GLM-4.6 Smoke Evaluation; API Failure Prevents Ranking

GLM-4.6 had a complete lack of data across all five dimensions in today's Smoke evaluation—execution, evidence linkage, judgment, integrity, and communication—and has entered automatic rerun, so it will not participate in this period's ranking.

Data Facts

In the score comparison table, code execution, material constraints, engineering judgment, task expression, and the main leaderboard all show "- → - (-)". The Smoke evaluation consists of 10 questions per day, and single-day fluctuations are normal, but this is not a score fluctuation; rather, all dimension records are empty.

Cause Analysis

Fluctuations from question sampling usually only cause individual question scores to rise or fall, not the loss of records for an entire dimension. Genuine model degradation would manifest as a continuous decline in scores or a drop in specific dimensions, not as API response timeouts or failures. This anomaly is explicitly labeled as an API failure/timeout, so the most likely cause is a connection interruption at the interface layer, not a change in model parameters or reasoning capability.

The execution dimension and the material constraints dimension are missing at the same time, further pointing to a call-chain issue rather than degradation in the model's internal judgment or expression capabilities.

Implications for Users

Teams that rely heavily on code execution should, when calling GLM-4.6, add API timeout retries and fallback model switching mechanisms to prevent a single Smoke-level quick-test anomaly from affecting production tasks.

In scenarios sensitive to material fidelity, developers should add manual spot checks or a second review to critical workflows to prevent output truncation caused by API-layer failures.

Engineering judgment and task expression depend on stable API responses. The current failure indicates that these side-leaderboard capabilities are temporarily unavailable, so a fallback plan should be prepared during model selection.

Strategic Assessment

This anomaly was directly caused by an API failure and is unrelated to the model's actual capabilities; there is no need to worry about GLM-4.6's core performance. If the next Smoke evaluation returns to normal score records, it can be confirmed that this incident was an isolated technical issue.

It is recommended to continue monitoring the automatic rerun results. Only if multiple dimensions are still missing after the rerun should further investigation into the stability of the model's server-side be needed.

Data source: YZ Index | Run #322 | View raw data