Analyses

Eight of them. The agent picks whichever applies to the data's shape; you override with /run.

	asks
Descriptive summary	what do the numbers look like?
Value by segment	how does the money split across the book?
Frequency / exposure profile	where is the exposure?
Time profile	how do records, and value, run over time, and are there gaps?
Concentration / Gini	do a few risks carry the total?
Correlation screen	which numeric pairs move together?
Outlier audit	which columns have values outside 1.5·IQR?
Missingness / data-quality audit	where are the holes?

Seeing which apply

/list shows only the ones that make sense for the data in front of you, numbered:

analyses that apply to this data
▸ 1. Descriptive summary
  2. Value by segment
  3. Frequency / exposure profile
  4. Time profile
  5. Concentration / Gini
  6. Correlation screen
  7. Outlier audit
  … 1 more
  ↑↓ move · ⏎ pick · esc cancel

The list is a modal sub-list: while it is open, arrows move, Enter picks, typing filters, and Esc clears a filter first and then backs out. It owns the keyboard until you leave it, so a command typed while it is open lands in the filter instead of running.

Switching

/run gini             by name
/run 3                by the number /list just printed
3                     a bare number answers whichever menu was just printed

The HARD pane re-renders. The pipeline does not restart, and the dataset is not re-read.

Why only eight

Everything here profiles and screens. Model fits, chain-ladder, GLMs, Lee-Carter, stay in the Scelo IDE, where there is room to show diagnostics alongside the answer.

A terminal pane pretending to fit a GLM would produce a number nobody should trust, because the thing that makes a fit trustworthy is the residual plots and the convergence detail sitting next to it. Scelo's real catalog is around thirty entries; these eight are the ones that survive being read in a column.

Column heuristics

Which column is the money, and which categorical is worth splitting by, is decided by shared heuristics rather than by the model. That matters for two reasons:

• it is stable, so the same data gives the same grouping every run
• the export generators use the same code, so the groupby in the generated Python and R lands on the columns you actually saw on screen

A test walks the entire menu and fails if any analysis lacks an export story, so every one of the eight can be handed to pandas and base R.

One column at a time

/show <column> prints a single column's profile as a card: type, cardinality, missingness, range and the distribution. It is the quickest way to check the agent's reading against the actual data.