A run that stops on an error restarts itself from the step that failed, after a short countdown (set it in Settings → Recovery, or press Restart now). If the same step fails again, the next attempt switches to OreML’s other compute engine automatically. Only if that fails too — or after three attempts — does it stop and ask for you: Settings → Only when you ask — diagnostics → Send diagnostics. It sends the run’s own record, never your data, and shows a reference to quote at support@oreml.com. We read these and fix what they show, fast.

You pick the mode once, at the start. Auto runs end to end on the engine’s measured recommendation at every judgement call. Manual runs the same engine but hands you the decisions where a human answer matters — columns to remove or rename, each outlier column, the target, the fill strategy per column — with the recommendation marked. Contests further down (feature set, projection, balancing) are settled by cross-validation in both modes. In manual, the value you choose is the value applied, stored in the step’s audit record and read downstream, and no step continues with a column left undecided. Choose manual when domain knowledge should beat a statistic; auto when you want the statistic.

Up to two targets can run together on any hardware, but in manual mode we strongly advise one at a time: two targets double your decisions and fork every downstream stage. One target means cleaner results, lower memory use and output that is easier to read — use two only for a side-by-side comparison.

In manual mode you choose upfront whether a run is supervised (with a target) or unsupervised. Guards catch an obviously wrong choice, but this one decision shapes target discovery, leakage checks, balancing and evaluation — give it a moment’s thought.

Give auto mode two plausible targets and it will model both, back to back — it sizes the work to your machine, so this is not a bug. For quieter fans and one result at a time, use manual mode and run them separately.

OreML writes analysis- and ML-ready tables at six checkpoints — after schema validation, both outlier stages, feature unification, dimensionality reduction and balancing — in standard formats you can open straight away. That default is the one we recommend. Exporting after every step is available but verbose (many more files, much more disk): use it to debug one stage, not as a habit.