diff --git a/docs/adr/0004-configurable-chunked-odata-batch.md b/docs/adr/0004-configurable-chunked-odata-batch.md
new file mode 100644
index 0000000..781f850
--- /dev/null
+++ b/docs/adr/0004-configurable-chunked-odata-batch.md
@@ -0,0 +1,57 @@
+# ADR-0004: Configurable, chunked OData `$batch` writes
+
+- **Status:** Accepted
+- **Date:** 2026-07-02
+
+## Context
+
+Batch writes (`CreateBatchCommand<T>` / `UpdateBatchCommand<T>` / `DeleteBatchCommand<T>`) previously
+went through `PanoramicData.OData.Client`'s changeset. That path has two limits IntegratoR consumers hit:
+
+- **No composite / anonymous-key support.** PanoramicData ToStrings the key on every write path
+  (re-verified empirically on 10.0.84), which is why single-entity composite-key writes already use a
+  raw-`HttpClient` bypass in `ODataClientAdapter` — see [ADR-0003](0003-odata-expression-translator-fork.md)
+  for the sibling "the upstream library doesn't model our key/wire shape" pattern.
+- **Batch was implicitly all-or-nothing with no knobs.** No failure-mode choice, no large-dataset
+  handling, and no per-item outcome for import error analysis. D365 F&O caps a single `$batch` at
+  **200 operations** (owner-verified against live JFI; the Microsoft "service protection" doc's 5000 is
+  a different limit), so any real bulk import already has to be split.
+
+Consumers importing long datasets need two things the old path could not express: **choose rollback vs.
+continue-on-error**, and **get a structured list of which rows failed and why**.
+
+## Decision
+
+Route **every** batch write through one hand-rolled `multipart/mixed` OData `$batch`
+(`ODataBatchRequestBuilder` / `ODataBatchResponseParser`, OASIS OData v4.01 §11.7.7) — uniform across
+create/update/delete and **every key shape**, including cross-company / no-`DataAreaId` / single-key
+global entities. PanoramicData's changeset is dropped for writes.
+
+- **`BatchFailureMode`** selects behaviour per call (`BatchOptions`) or per config
+  (`ODataSettings.Batch.FailureMode`):
+  - `Atomic` (**default**) — one all-or-nothing OData changeset per chunk.
+  - `ContinueOnError` — operations applied independently, failures collected.
+- **Auto-chunking** at `ODataSettings.Batch.MaxOperationsPerChunk` (**default 150**, validator-bounded
+  **1..200** by D365's ceiling). `StopOnFirstFailedChunk` (default true) halts further chunks in `Atomic`.
+- **`Result<BatchOutcome>`** (breaking) — success only if every item commits; a failure carries a
+  `BatchIntegrationError` whose `BatchOutcome` lists per-item `Index` / `ChunkIndex` / `Key` / status /
+  D365 error, retrievable via `result.GetError()`.
+- The `$batch` POST rides the named `"ODataClient"` Polly policy (429/5xx) but is **not** wrapped in the
+  operation retry — a processed batch must never be blindly re-POSTed.
+
+## Consequences
+
+- **Breaking (v3.0.0).** `IBatchService` / `IODataBatchService`, `ODataService<T>`, the three generic
+  batch commands + handlers, the F&O `*sCommand` handlers, and `IODataClientAdapter.Batch*Async` all
+  change to `Result<BatchOutcome>`; the `ODataService<T>` ctor gains an optional
+  `IOptions<ODataSettings>?` (binary break for direct constructors, DI consumers unaffected). Signalled
+  via `GitVersion.yml next-version: 3.0.0`.
+- **Atomicity is per-chunk, not per-dataset.** A dataset larger than the chunk size spans several
+  changesets; each commits or rolls back independently. This is inherent to D365's 200-op cap, not a
+  choice — documented in `wiki/Known-Limitations.md`. Keep imports idempotent on the composite key.
+- **`ContinueOnError` is per-item today.** The single-`$batch` transport (`Prefer:
+  odata.continue-on-error`) is undocumented for D365 F&O; enabling it is deferred behind a live-JFI
+  `$batch` smoke run that also confirms changeset atomicity end-to-end.
+- One uniform write path replaces two (PanoramicData changeset + raw bypass), at the cost of owning the
+  multipart emit/parse. Pinned by `ODataBatchWireTests`, `ODataClientAdapterAtomicBatchTests`, and
+  `ODataServiceBatchChunkingTests`.
diff --git a/docs/adr/README.md b/docs/adr/README.md
index 00a901d..08e0651 100644
--- a/docs/adr/README.md
+++ b/docs/adr/README.md
@@ -37,3 +37,4 @@ one and mark the old `Superseded by ADR-NNNN`.
 | [0001](0001-generic-handler-and-validator-registration.md) | Cross-assembly generic handler & validator registration | Accepted |
 | [0002](0002-durable-functions-result-converter-wiring.md) | Durable Functions `Result<T>` converter wiring | Accepted |
 | [0003](0003-odata-expression-translator-fork.md) | Forked OData expression translator for `[JsonPropertyName]` | Accepted |
+| [0004](0004-configurable-chunked-odata-batch.md) | Configurable, chunked OData `$batch` writes | Accepted |