diff --git a/specs/database/_builder_architecture.md b/specs/database/_builder_architecture.md
new file mode 100644
index 0000000..22e6911
--- /dev/null
+++ b/specs/database/_builder_architecture.md
@@ -0,0 +1,168 @@
+# PostgREST Query Builder — Architecture
+
+<!-- Area-wide architecture doc, not a single-feature spec. Describes the
+     builder chain shared by every database.* capability, independent of any
+     one SDK's class names or generics. This is the source of truth for how
+     the builder should be composed in new and existing SDK implementations. -->
+
+## Overview
+
+Every `database.*` capability is a step in one linear builder chain: pick a
+target (table/view or RPC function), optionally narrow rows with filters,
+optionally reshape/paginate the result, then execute. The chain has five
+logical stages. An SDK may implement each stage as a distinct object or reuse
+one object with narrowing types (the JS reference implementation does the
+latter) — either is valid as long as the external chaining behavior below
+holds.
+
+Reference implementation: [supabase-js `packages/core/postgrest-js`](https://github.com/supabase/supabase-js/tree/21e410f5622c163c3e64fc1a3bb78ee3b9ec86c9/packages/core/postgrest-js)
+(`PostgrestClient` → `PostgrestQueryBuilder` → `PostgrestFilterBuilder` →
+`PostgrestTransformBuilder` → `PostgrestBuilder`).
+
+## Graph
+
+```mermaid
+flowchart TD
+  Client["Client"]
+  QueryBuilder["QueryBuilder\n(scoped to one table/view)"]
+  FilterBuilder["FilterBuilder\n(row-matching predicates)"]
+  TransformBuilder["TransformBuilder\n(shape/order/paginate result)"]
+  Execute["Execute\n(await / .then() -> HTTP request)"]
+
+  Client -->|"schema(name)\ndatabase.query.schema_selection"| Client
+  Client -->|"from(table)\ndatabase.query.from_table"| QueryBuilder
+  Client -->|"rpc(fn)\ndatabase.query.rpc"| FilterBuilder
+
+  QueryBuilder -->|"select()\ndatabase.query.select"| FilterBuilder
+  QueryBuilder -->|"insert()\ndatabase.mutate.insert"| FilterBuilder
+  QueryBuilder -->|"update()\ndatabase.mutate.update"| FilterBuilder
+  QueryBuilder -->|"upsert()\ndatabase.mutate.upsert"| FilterBuilder
+  QueryBuilder -->|"delete()\ndatabase.mutate.delete"| FilterBuilder
+
+  FilterBuilder -->|"eq/neq/gt/gte/lt/lte/like*/ilike*/is/is_distinct/in/not_in/\ncontains/contained_by/range_*/overlaps/text_search/\nmatch/not/or/filter/regex*\ndatabase.using_filters.*"| FilterBuilder
+  FilterBuilder -->|"(no more filters needed)"| TransformBuilder
+  FilterBuilder -->|"await / .then()"| Execute
+
+  TransformBuilder -->|"order/limit/range/abortSignal/rollback/maxAffected\ndatabase.using_modifiers.order,limit,range,\nrequest_cancellation,dry_run,max_affected_rows"| TransformBuilder
+  TransformBuilder -->|"select()\ndatabase.mutate.select_after_mutation"| FilterBuilder
+  TransformBuilder -->|"single()\ndatabase.using_modifiers.single_row"| Execute
+  TransformBuilder -->|"maybeSingle()\ndatabase.using_modifiers.maybe_single_row"| Execute
+  TransformBuilder -->|"csv()\ndatabase.using_modifiers.format_csv"| Execute
+  TransformBuilder -->|"geojson()\ndatabase.using_modifiers.format_geojson"| Execute
+  TransformBuilder -->|"explain()\ndatabase.using_modifiers.explain"| Execute
+  TransformBuilder -->|"await / .then()"| Execute
+
+  Client -.->|"construction config: auto_retry, request_timeout\ndatabase.configuration.*"| Execute
+```
+
+## Stages
+
+### Client
+
+Holds connection state: base URL, headers, `fetch` implementation, and
+construction-time config (`database.configuration.auto_retry`,
+`database.configuration.request_timeout`). Not itself a query — it is the
+factory for the other stages via three entry points: `schema()`, `from()`,
+`rpc()`.
+
+### QueryBuilder
+
+Scoped to one table/view. Exposes the five entry verbs
+(`select`/`insert`/`update`/`upsert`/`delete` — `database.query.select`,
+`database.mutate.insert`, `database.mutate.update`, `database.mutate.upsert`,
+`database.mutate.delete`). Each verb picks the HTTP method and hands off to a
+FilterBuilder; the verb choice does not change which filter/modifier
+capabilities are available downstream. `rpc()` bypasses this stage entirely —
+there is no table and no verb choice, so the Client hands off straight to a
+FilterBuilder.
+
+### FilterBuilder
+
+Row-matching predicates (`database.using_filters.*`). Structurally uniform:
+every filter appends a predicate and returns the same stage, so any number of
+filters can be chained in any order. `or()` and `not()` are filters like any
+other at this stage — they don't terminate or branch the chain, they just
+combine/negate predicates. Once no more filtering is needed, the chain either
+executes directly or continues into TransformBuilder — a filter-only query
+(e.g. `select().eq(...)`) is a complete, executable chain without ever calling
+a modifier.
+
+### TransformBuilder
+
+Reshapes, orders, or paginates the result (`database.using_modifiers.*`).
+Split into two kinds of methods:
+
+- **Self-looping** — `order`, `limit`, `range`, `abortSignal`
+  (`request_cancellation`), `rollback` (`dry_run`), `maxAffected`
+  (`max_affected_rows`). These return the same stage, so they compose freely
+  with each other in any order.
+- **Terminal-narrowing** — `single` (`single_row`), `maybeSingle`
+  (`maybe_single_row`), `csv` (`format_csv`), `geojson` (`format_geojson`),
+  `explain` (`explain`). These commit the response shape; no further
+  filter/modifier calls are meaningful after one of these, only execution.
+
+`select()` at this stage is distinct from QueryBuilder's `select()`: it's the
+post-mutation embed (`database.mutate.select_after_mutation`), called after
+`insert`/`update`/`upsert`/`delete` to return affected rows, and it hands
+back to a FilterBuilder since the embedded selection can itself be filtered
+by relationship modifiers.
+
+Relationship embedding (spread/nested selects) is a syntax within the
+`select()` column argument (`database.using_modifiers.relationship_embed`),
+not a separate builder method — it doesn't add a node to this graph.
+
+### Execute
+
+Triggered by `await` or an explicit `.then()` on the chain — not by a
+dedicated method call. Performs the HTTP request, applies
+`auto_retry`/`request_timeout` from Client construction, and normalizes the
+response. A chain can reach Execute directly from FilterBuilder (no modifiers
+used) or from TransformBuilder.
+
+## Implementation Notes
+
+The JS reference implementation builds exactly one object per query and
+narrows its exposed type at each stage (`FilterBuilder extends
+TransformBuilder extends Builder`, with terminal methods doing `return this as
+unknown as Builder<...>`). This is an implementation choice to get static
+typing of "what's callable next" — it is not a requirement. An SDK may instead
+construct a genuinely new object at each stage transition, as long as:
+
+- filters can be chained in any order and any quantity before execution,
+- modifiers can be chained in any order after filters,
+- a terminal-narrowing modifier prevents further filter/modifier calls in that
+  SDK's own idiomatic way (type system, runtime error, or simply undocumented),
+- `rpc()` does not require a table/verb selection step first.
+
+## Legend — Capability ID → Graph Edge
+
+| Capability ID | Stage / Edge |
+|---|---|
+| `database.query.schema_selection` | Client → Client (`schema()`) |
+| `database.query.from_table` | Client → QueryBuilder (`from()`) |
+| `database.query.rpc` | Client → FilterBuilder (`rpc()`) |
+| `database.query.select` | QueryBuilder → FilterBuilder (`select()`) |
+| `database.mutate.insert` | QueryBuilder → FilterBuilder (`insert()`) |
+| `database.mutate.update` | QueryBuilder → FilterBuilder (`update()`) |
+| `database.mutate.upsert` | QueryBuilder → FilterBuilder (`upsert()`) |
+| `database.mutate.delete` | QueryBuilder → FilterBuilder (`delete()`) |
+| `database.mutate.select_after_mutation` | TransformBuilder → FilterBuilder (`select()`) |
+| `database.using_filters.*` (eq, neq, gt, gte, lt, lte, like, like_all, like_any, ilike, ilike_all, ilike_any, is, is_distinct, in, not_in, contains, contained_by, range_gt, range_gte, range_lt, range_lte, range_adjacent, overlaps, text_search, match, not, or, raw, regex, regex_icase) | FilterBuilder → FilterBuilder |
+| `database.using_modifiers.order` | TransformBuilder → TransformBuilder (`order()`) |
+| `database.using_modifiers.limit` | TransformBuilder → TransformBuilder (`limit()`) |
+| `database.using_modifiers.range` | TransformBuilder → TransformBuilder (`range()`) |
+| `database.using_modifiers.request_cancellation` | TransformBuilder → TransformBuilder (`abortSignal()`) |
+| `database.using_modifiers.dry_run` | TransformBuilder → TransformBuilder (`rollback()`) |
+| `database.using_modifiers.max_affected_rows` | TransformBuilder → TransformBuilder (`maxAffected()`) |
+| `database.using_modifiers.single_row` | TransformBuilder → Execute (`single()`) |
+| `database.using_modifiers.maybe_single_row` | TransformBuilder → Execute (`maybeSingle()`) |
+| `database.using_modifiers.format_csv` | TransformBuilder → Execute (`csv()`) |
+| `database.using_modifiers.format_geojson` | TransformBuilder → Execute (`geojson()`) |
+| `database.using_modifiers.explain` | TransformBuilder → Execute (`explain()`) |
+| `database.using_modifiers.relationship_embed` | Syntax within `select()`'s column argument — no dedicated edge |
+| `database.configuration.auto_retry` | Client construction → Execute (applied during HTTP request) |
+| `database.configuration.request_timeout` | Client construction → Execute (applied during HTTP request) |
+
+## Related
+
+- [capabilities/database.yaml](../../capabilities/database.yaml) — feature registry this graph traces