Workspace help

What each step does

Step 1, choose files: the loan extract from the core banking export (required), control totals (optional), a saved mapping profile and a project file. Files are read in the tab and hashed with SHA-256.

Step 2, check data: the row rules of the frozen engine run over the extract. Errors block the snapshot, warnings travel into the snapshot and every report, information is shown. Rejected rows go to the correction file with line, column, found value and the rule expected. Control totals are reconciled when supplied.

Step 3, open the risk review: the v1 dataset is built, the Tier 1 v2 engine runs with the default configurations and a validated snapshot is stored in the project. The review queue orders rows the engines already produced; nothing new is computed about risk.

Then: record actions in the register (maker and checker labels), generate the management brief and the committee packet, ask the five fixed questions, and export the project file to keep everything.

What the institution file must contain

A CSV with one row per centre level pooled account and month end: as_of, loan_id, branch_id, branch_name, district, province, product, sector, disbursed_npr, outstanding_npr, days_past_due, restructured, written_off, writeoff_npr, due_npr, collected_npr. Other column names can be mapped with a mapping profile.

Tokenised only: loan_id is a token your institution assigns. No member name, phone number, address or citizenship number may be in the file; a column that looks like one stops the import before any row is read.

Dates are ISO month ends (YYYY-MM-DD). Amounts are whole NPR. Control totals, when supplied, are a CSV with as_of, metric and value.

Units

Amounts are NPR with South Asian grouping (1,00,00,000). The Amounts selector in the header shows figures in lakh, crore or arba, or picks the scale automatically: 1 arba = 100 crore = 1,000 lakh = 1,00,00,00,000 NPR. One table uses one scale and the header names it.

1 lakh = 1,00,000 · 1 crore = 1,00,00,000 · 1 arba = 1,00,00,00,000

The selector changes the display only. Exports (correction file, queue CSV, management brief, committee packet, project file) keep the contract figures and the crore scale of the board pack. Shares are percentages, changes are percentage points, scores are out of 100.

Draft and reviewed

Every report is DRAFT when generated. It reads REVIEWED only when a report_reviewed note with a reviewer label is recorded for the snapshot and the reconciliation is not a variance, or every variance has a register item in closed or overridden state. The label is printed on every page.

Reviewer, maker and checker are workflow labels typed by the user, not authenticated access control: the workspace has no login and cannot verify who typed a label.

Sample and customer data never mix

The bundled sample is synthetic and opens on its own route with its own banner. A sample project never holds a customer snapshot, a customer project never holds a sample one, and sample documents are never listed with institution reports. Close one project before opening the other.

Nothing leaves the browser

Files are read in the tab, hashed and parsed locally. Results live in page memory and in the project file you export explicitly. The workspace writes nothing to browser storage, sends nothing to a server and calls no model. Reminders run as a dry run inside the page; email and sms delivery does not exist.

What this is not

Not a supervisory assessment, not a regulatory report, not a regulatory finding and not a forecast. Every threshold is an illustrative default, uncalibrated, pending calibration to the current NRB directive for D class institutions.

Not a lending decision and not a change to any loan: the register never writes to the core banking system. The assistant is deterministic templates over the calculation results, not machine learning. Rows are centre level pooled accounts: one days past due figure stands for several member loans.