Collect one failing request or training step, runtime/source revision, resolved config, artifact identity, logs and relevant outputs. Keep confidential inputs and credentials out of shared reports.
| Symptom | First checks |
|---|---|
| Side/rear geometry mismatch | Shared input preparation: physical camera mapping, synchronized WSM, crop and actual per-view text |
| Short output or caption transition failure | Input settings: total horizon, source coverage, sidecar bounds and first-chunk selection |
| Missing metadata / HF load rejected | Export: regular student, rolling profile, per-view text and matched tokenizers |
| Unexpected CFG or sampler behavior | Fixed four-step student; use the guidance-1 recipe, not bidirectional RF step/shift settings |
| Missing or misaligned LiDAR | HD-map controls versus measured prefix, 30/10 Hz clocks, source IDs; causal measured prefix is at most one sweep |
| Missing training camera/caption/control fields | Bundle mappings at every stage; inspect decoded and formatted samples |
| Wrong teacher or resume iteration | Checkpoints: new stage versus complete same-stage resume |
| OOM / compilation / distributed error | Verify setup/mesh, then an authorized smoke; preserve architecture/objective |
| Only one SF-DMD phase appears | Six iterations include one student and five critic updates; inspect phase metrics and both optimizer histories |
Preserve original inputs and the model's checkpoint/attention settings while isolating the issue. Apply the smallest correction and check the affected result. For reader errors, let bounded retries finish and check training progress. Stop the affected run on a fatal failure, exhausted retries, invalid data or its time limit; leave independent healthy jobs running and allow saves/exports to finish. Report observations, the confirmed cause or remaining hypothesis, the correction and checks performed.