Desktop Evaluate & train workflow

Evaluate & train runs the five commands the terminal otherwise keeps to itself: skryba corpus, skryba benchmark, skryba train, skryba database check and skryba database drop-schema. One segmented picker chooses which; the fields are the command's own positionals and flags, and every number is a field you type — none starts filled in.

Form

Subcommand	Control	UI value	CLI argument
Verify corpus	curated corpus file	non-empty	positional PATH
Verify corpus	samples (empty = all)	typed; omitted when empty	--samples
Benchmark	curated corpus file	empty, or documentation roots instead	--corpus
Benchmark	documentation roots, one per line	used when no corpus is named	positional roots
Benchmark	report destination	non-empty	--output
Benchmark	policy file (optional)	empty, omitted when blank	--policy
Benchmark	samples	typed, required	--samples
Benchmark	candidates per sample	typed, required	--candidates
Benchmark	fewest / most words per passage	typed, required	--min-words / --max-words
Benchmark, Train policy	temperature (nonnegative)	typed, required	--temperature
Benchmark, Train policy	minimum quality	typed, required	global --minimum-quality
Train policy	source rows	non-empty	positional SOURCE
Train policy	text column	empty, omitted when blank	--text-column
Train policy	starting policy (optional)	empty, omitted when blank	--policy
Train policy	trained policy destination	non-empty	--output-policy
Train policy	epochs, generations per group	typed, required	--epochs, --generations
Train policy	learning rate, beta, clip epsilon, inner iterations	typed, required	--learning-rate, --beta, --clip-epsilon, --inner-iterations
Train policy	detector weight, quality weight, quality failure penalty	typed, required	--detector-weight, --quality-weight, --quality-failure-penalty
Verify corpus, Benchmark, Train policy	seed	typed, required	--seed
Database check	none	—	—
Drop schema	schema to drop	non-empty	--schema
Drop schema	the same schema again, to confirm	must equal the schema to drop	not sent

Run is blocked with the app's own sentences until the required fields are filled: Choose the curated corpus file., Name a curated corpus or at least one documentation root., Name a corpus or documentation roots, not both; the CLI refuses both., Say where the report is written., Choose the source rows to train on., Say where the trained policy is written., and State the <setting>. for the first empty number, for example State the seed that orders the sample. or State the learning rate. Database check has nothing to validate and always runs. Drop schema runs only once the schema is named (Name the schema to drop.) and typed a second time to the same name (Type the schema's name again to confirm: dropping it cannot be undone.): a dropped schema and its rows do not come back. The CLI itself refuses the default schema, and that refusal shows here verbatim.

Exact invocation shape

skryba corpus PATH --seed N [--samples N]
skryba --minimum-quality Q benchmark (--corpus PATH | ROOT...) --output PATH --samples N --candidates N --min-words N --max-words N --temperature T --seed N [--policy PATH]
skryba --minimum-quality Q train SOURCE --output-policy PATH --epochs N --generations N --learning-rate F --beta F --clip-epsilon F --inner-iterations N --temperature T --detector-weight F --quality-weight F --quality-failure-penalty F --seed N [--text-column NAME] [--policy PATH]
skryba database check
skryba database drop-schema --schema NAME

The app sends each entered numeric value without selecting a default. Global options from the window (--detectors) are sent before the command as for every workflow; benchmark and train use the same Brama configuration as Rewrite. Training also uses its detector credentials; benchmark judges quality without detector calls. Skryba imposes no temperature ceiling: the selected Brama route reports its own restrictions.

Result

The panel shows the command's own JSON, pretty-printed, as it printed it: the corpus verdict, the benchmark summary and report path, the trained policy path, the database tables and counts, or the dropped store with "dropped": true. The app computes nothing from it. A refusal is the CLI's stderr verbatim, so every refusal on the corpus, benchmark, train and database pages appears here unaltered.