Scope: three models, one application, four dependent requests. A second planned application was withdrawn before the models saw it. The earlier comparison had six distinct short problems for each of six models. Neither is a broad, repeated survey of everything a model can do.
Unfinished work: Ornith did not complete request two; Flint did not complete request four. Requests beyond those points have no completed result to grade. Both repetitions were stopped. The completed review found a monitoring failure but did not isolate why the repetition began.
Specific limits: Flint’s changed correction files are one continuing problem, not multiple separate penalties. Its early manual reset affected tables that were already empty; no erasure of populated records was demonstrated. Raon restored its temporarily lost credit, and the preserved old reports were not lost. Its broader backup weakness was established by reading the saved work, not by running a new failure example.
Recommendation: the observed follow-through favors Raon for a longer assignment. The earlier focused-repair evidence still favors Ornith. There is not enough matched long-job evidence to rank MiniCPM or Apodex against Raon.