# ADR-0004 — Stamdata as code (config-as-code, not a production database)

Status: Accepted · Date: 2026-07-20

## Context

The business needs to control certain inputs that change over time — the clearest example
being **which professions link to which diplomas** (`geneeskunde → Arts`, …), but also
tunable thresholds, policy-question text, document-category definitions, and some letter
copy. Two hard constraints:

1. **No managing this through a production database.** A live admin surface writing DB rows
   means a bad value ships silently and is discovered in production.
2. **Issues must be caught at compile time.** A change should be typed, reviewed, and
   versioned before it can affect anyone.

The codebase already leans this way but had never named it as a pattern, and one key table
was neither isolated nor validated:

- All reference data and thresholds are **compiled-in C# constants**, served through
  screen-shaped BFF-lite endpoints; the frontend renders decisions and holds no reference
  data (ADR-0001).
- User-facing UI copy is already **`$localize`** (`apps/<app>/src/locale/*.xlf`) — git-tracked, and a
  second locale is a translation file, not a code change. That is already the compile-time
  model for text.
- The profession↔diploma map lived as a _private_ `Dictionary` inside `DiplomaRules`, mixed
  in with the rules that consume it, with **no cross-reference check**: a diploma whose
  program wasn't in the map silently rendered `"Onbekend"`.

## Decision

Treat business-tunable reference data as **stamdata-as-code**: typed, checked-in
configuration, changed through the normal git → PR → build → deploy pipeline. Never a
production database, never runtime-editable.

1. **One home, typed.** Business-editable reference data lives in the
   `BigRegister.Stamdata` namespace (`backend/src/BigRegister.Api/Stamdata/`), one file per
   concern. A table lives **either** as plain typed C# data (records / dictionaries) **or** as
   a typed JSON data-file deserialized into a record (`professions.json` → `ProfessionMapping`,
   loaded via `StamdataFile`). Both are checked-in config-as-code, gated the same way; the
   data-file trades the compiler's _value_ check (gate #1 sees only the shape, not a wrong
   `beroep`) for hand-editing ergonomics and the low-code editor below — the value gate becomes
   `StamdataValidationTests`. Separate the **data** (what the business tunes) from the **rules**
   (dev-owned logic that consumes it): the profession _table_ is `Stamdata.Professions`; the
   _rule_ "an English diploma needs a B2 question" stays in `DiplomaRules`. Tables may carry
   **valid-time** (`geldigVan`/`geldigTot`, half-open `[van, tot)`); `StamdataCatalog` +
   `StamdataTable.Of<T>` describe every table generically (columns reflected from the record)
   so one endpoint pair and one grid editor serve all of them.
2. **Served unchanged.** The existing BFF-lite endpoints keep serving this data
   (`/duo/diplomas`, `/intake/policy`, `/uploads/categories`, …). No frontend change — the
   FE still renders decisions.
3. **Two gates.** The **C# compiler** catches shape and type mistakes. A build-time
   **`StamdataValidationTests`** catches the referential integrity the compiler can't —
   every seeded diploma program resolves to a real profession, no blank keys/values,
   thresholds in range. CI runs it, so a bad edit fails the build and never merges.
4. **Business control = config-as-code (GitOps).** The business owns the content of these
   files; a change is a reviewed edit, not a live DB write. A future low-code editor could
   commit a PR on their behalf without changing this model (the compile-time gate stays).

### Where each kind of business-controllable thing lives

| Kind                                                                                                         | Home                                                                                          | Gate                                                                                                  |
| ------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------- |
| Reference tables + tunable numbers (professions↔diplomas, thresholds, policy questions, document categories) | `Stamdata/` typed C# **or** typed JSON data-file (`professions.json`), optionally valid-timed | compiler (shape; + values when C#) + `StamdataValidationTests` (values, references, validity windows) |
| User-facing UI copy                                                                                          | `$localize` → `apps/<app>/src/locale/*.xlf`                                                   | build (`i18nMissingTranslation: error`)                                                               |
| Letter / brief passage content                                                                               | config-as-code in the backend (seed content), **not** the DB                                  | compiler + endpoint tests                                                                             |

### The deliberate exception: operational configuration

"Never runtime-editable" above is the rule for **stamdata** — the shared reference tables
and business rules the whole register runs on. It is not a ban on all persisted
configuration. Some configuration is operational rather than business-rule, and belongs to
an admin persona at runtime.

This section states the **test** rather than a list, so the next surface can check itself
instead of arguing by analogy. Runtime-editable persistence is permitted only when all four
hold:

1. **The catalog lives in code.** What may be set — the keys, the schema, the defaults,
   the descriptions — is compiled in and reviewed through git. The store holds values, never
   the definition of what a value means.
2. **An unknown or unlisted key fails closed.** A row the code catalog does not know cannot
   invent a setting, enable a feature, or be written. A bad row is inert, not authoritative.
3. **The value is operational.** Per-organisation identity, or an on/off rollout switch —
   not a shared business rule whose wrong value breaks the register for everyone. This is the
   clause that keeps stamdata out.
4. **Writes are admin-capability-gated and audited.** The write path goes through an `Authz`
   capability gate, and the gate records the decision — allow as well as deny — in
   `AuthzAuditStore`.

**Two surfaces pass this test today.**

| Surface                       | (1) catalog in code                             | (2) fails closed                                             | (3) operational                   | (4) gated + audited             |
| ----------------------------- | ----------------------------------------------- | ------------------------------------------------------------ | --------------------------------- | ------------------------------- |
| `OrgTemplateStore` (WP-23/26) | the `OrgTemplateDto` shape + `OrgTemplateRules` | unknown `subOrgId` → `null` → the endpoint 404s              | one sub-organisation's letterhead | `OrgAdmin` → `orgtemplate:edit` |
| `FeatureFlagStore` (WP-47)    | `Domain/Features/FeatureFlags.Catalog`          | unknown key → `Set` returns false (404); `IsEnabled` → false | an on/off rollout switch          | `FlagsAdmin` → `flags:manage`   |

Clause (4) became true for both only with RB-07, which moved `AuditAuthz` from each gate's
deny branch into the gate itself so the allow path is recorded too. Before that, both
surfaces were gated and **not** audited, and this ADR would have ratified a control the code
did not implement.

Org-templates also carry publish/rollback versioning inside the app, which is stronger than
the test requires but not part of it.

Stamdata itself — the rules and reference tables — fails clause (3) by construction and
stays code.

## Consequences

- **+** Every change is typed, reviewed, versioned, and rollback-able through git; zero
  production-DB risk; a dangling reference fails the build with a clear message instead of
  reaching users.
- **−** A change needs the PR pipeline — not instant, and a non-developer may need dev
  assistance to edit C# (mitigated later by a low-code editor that emits a PR, or by a
  data-file format if hand-editing ergonomics ever outweigh maximal compile-time safety).
- **Shipped with this ADR:** the profession↔diploma map (`Stamdata/Professions.cs`) and the
  policy-question wording (`Stamdata/PolicyQuestions.cs`) extracted from `DiplomaRules`,
  which now consumes both (behaviour unchanged), guarded by `StamdataValidationTests`.
  Document-category definitions follow the same pattern as the obvious next step; not moved
  yet.
- **Shipped as a follow-on (WP-29):** the "future low-code editor" and "data-file format" this
  ADR floated are now real. `professions` moved to `professions.json` (typed, valid-timed) and
  the generic `StamdataCatalog`/`StamdataTable`/`StamdataFile` model plus read-only, admin-gated
  `GET /stamdata` endpoints back an Angular `beheer/stamdata` editor. It is **not** a runtime
  write path: the admin edits a grid and downloads the edited JSON to commit as a reviewed PR —
  the compile/validation gate stays the authority, so this ADR's core decision is unchanged.