# IV Desk

> Screen credit-risk variables before you model: drop a loan-level modeling sample and get the
> pre-model variable screen (abnormal months, missing rate, IV, PSI, noise, correlation) computed in
> the browser, plus an AI review of what to keep, what looks like leakage and which thresholds to
> change.
> https://iv-desk.skillsafe.ai/

IV Desk is a web app on SkillSafe. The screen runs free in the browser and the rows are never
uploaded; the review is a metered AI run (model alias gpt-terra) that sees only per-feature
statistics and is checked against the screen afterwards.

## Input

- A loan-level CSV (comma, semicolon, tab or pipe; up to 300,000 rows and 600 columns): an
  application date, a 0/1 bad flag, optionally an organization column, key columns, then candidate
  features. A larger file is screened on its first 300,000 rows / 600 columns and every output
  (page, report, workbook, review input, download names) is marked PARTIAL.
- Optional: out-of-sample organizations, columns to ignore, a note on the product and target
  definition, a question.
- Settings (defaults from the source skill): values treated as missing -1, -999, -1111; months need
  10 bads and 500 rows; missing-rate threshold 0.6; overall and per-organization IV 0.1 with 2
  low-IV organizations tolerated; PSI 0.1 in a third of months in 6 organizations; 10 label
  permutations for the noise screen; correlation 0.9 with the top 20 features by IV protected.

## The screen (in the browser, no AI, no upload)

Following the pipeline of the source skill, every step runs independently over the same modeling
sample and a feature survives when no step flags it:

- Load and format: rows without a 0/1 label or a readable date left out, duplicate keys removed,
  sentinel values treated as missing, constant columns dropped.
- Organization sample analysis and out-of-sample split.
- Abnormal months: fewer bads or rows than the thresholds.
- Missing rate overall and by organization.
- Information value from decision-tree bins (at most 5 leaves, 1% minimum leaf, missing as its own
  bin), overall and by organization.
- PSI month over month inside each organization (10 quantile bins of the earlier month plus missing).
- Noise screen: the feature's IV against its IV under label permutations (a univariate stand-in
  for the reference pipeline's LightGBM null importance).
- Correlation: pairs above the threshold drop the lower-IV feature (the reference uses LightGBM
  gain).
- Flags: IV above 0.5 (possible leakage), names that suggest performance data, sentinel collisions,
  too few bads, rules that cannot fire as set (more organizations required than exist), one-off
  PSI shifts, protected correlated pairs.

The IV, missing-rate, PSI and correlation arithmetic was checked against scikit-learn decision
trees, pandas and the source skill's own PSI function on the bundled examples.

Exports: the report as Markdown, the feature table as CSV, a multi-sheet Excel workbook
(SpreadsheetML) mirroring the reference report's sheets, the survivors and drops as Python lists
(features_keep / features_drop) to paste into a notebook, and the exact run input as JSON.

## The review (metered)

One JSON object: verdict (`ready`, `revise`, `blocked`), headline, TL;DR, one review per step
(`agree`, `adjust`, `question`), feature calls (`keep`, `drop`, `investigate`) with the screen's
result and IV quoted, leakage suspects with the check that would clear each, threshold changes, next
steps, a cleaning memo, and an answer to every flag. The browser reconciles it: every flag answered,
every feature named present in the file, IV and keep/drop quoted exactly, every survivor called,
confirmed leakage carried, threshold advice naming real settings with their current values, and a
verdict no looser than the flags left standing. Suggested thresholds can be applied back into
Settings with one button to re-screen for free.

## API

https://iv-desk.skillsafe.ai/api.html - base URL https://api.skillsafe.ai/v1/app-api, fields
`task` (always `review`), `dataset`, `context`, `facts` (a JSON string with the screen's
statistics), `question`.

## Source

Derived from the agent skill @github/datanalysis-credit-risk (github/awesome-copilot, MIT):
https://skillsafe.ai/skill/@github/datanalysis-credit-risk - license at
https://iv-desk.skillsafe.ai/LICENSE-GITHUB-AWESOME-COPILOT.txt