# 10 — Palantir J8: Exhibit Factory

**Rank 22 of 28** · Tier B · `Palantir/J8`
**Stack:** Python (Foundry Tier 2 transforms), TypeScript + React + Vite (Foundry custom widget), pytest, vitest
**Scope:** 172,545 lines across the Palantir tree; 20 pytest + 5 vitest green, `tsc` clean

> **Why this ranks here.** This is the **financial-data-pipeline** entry, and
> for a bank it is the most directly relevant project in the portfolio: typed
> ontology, deterministic normalization, reference qualification, and an ingest
> that *blocks on warnings* rather than passing bad data downstream.

---

## What it is

A pipeline that turns heterogeneous financial workbooks into ontology-conformant
exhibits, plus the Foundry custom widget that renders and edits them.

### Backend — `Tier2/python/ExhibitFactory`

| Module | Role |
|---|---|
| `catalog.py`, `datasets.py` | Dataset registry and resolution |
| `keyio.py` | Exhibit key parsing / emission |
| `tokens.py` | Token grammar for cell and reference addressing |
| `refs.py` | **Reference qualification — label-first resolution** |
| `normalize/` | Per-shape normalizers: `flat_fy_rows`, `flat_pycyby`, `cash_monthly`, `multi_subsheet`, `f1_transition`, `user_input`, driven by a `registry` |
| `compute/resolver.py` | Derived-value resolution |
| `transforms/ingest.py` | The ingest transform |
| `fixture.py`, `tools/make_widget_fixture.py` | Fixture generation for the widget |

The normalizer registry is the good design here: each input shape gets a named,
tested normalizer, and dispatch is data-driven rather than a branching
`if`-ladder. Adding a shape is adding a module and a registry entry.

### Frontend — `CustomWidget`

React + TypeScript + Vite Foundry widget with a config-driven main entry, an
ontology client layer, and screen-level views. `tsc` clean, 5 vitest green.

### Rebuilt on ontology v7

The whole system was re-derived against ontology v7 rather than patched
forward, with **reference qualification changed to label-first**. F-5 proven;
the transposed-3 case is **WARN-blocked** — it does not silently produce a
wrong exhibit, it stops.

> Refusing to emit output you cannot vouch for is the single most
> bank-relevant instinct in this repository.

## Verification

| Suite | Result |
|---|---|
| pytest | 20 green |
| tsc | clean |
| vitest | 5 green |

Tests target behavior that matters: `test_recipe_families`, `test_tokens`,
`test_factory_f5`, `test_ingest_merge`, `test_warn_blocked`.

## Interview surface this opens

- Ontology / schema-driven pipelines and versioned migration (v6 → v7)
- Data-driven dispatch vs conditional branching for heterogeneous inputs
- Fail-closed ingest: when a warning should block a pipeline
- Deterministic transforms and fixture-based testing across a language boundary