diff --git a/effect/architecture.md b/effect/architecture.md
index 7094c03..c19ce23 100644
--- a/effect/architecture.md
+++ b/effect/architecture.md
@@ -79,7 +79,7 @@ The `Run` method wraps the user-defined method (`Junctions()` or `RunInternal`)
 
 The `EffectRunner` coordinates all registered effect providers. It builds its provider list at construction time by querying `IEffectProviderFactory` instances filtered through the `IEffectRegistry`.
 
-It exposes three operations that fan out to every active provider:
+It exposes three operations that fan out to every active provider, awaiting each provider sequentially:
 - **`Track(model)`** — Register a new model for tracking (e.g., add `Metadata` to the EF change tracker)
 - **`Update(model)`** — Notify providers of an in-memory mutation (e.g., re-serialize parameters after output is set)
 - **`SaveChanges(ct)`** — Persist all accumulated changes across all providers
@@ -173,7 +173,7 @@ Junctions execute inside the "Execute Train Chain" box. Each mutation to the tra
 
 ### DataContext
 
-`DataContext<TDbContext>` extends EF Core's `DbContext` and implements both `IDataContext` and `IEffectProvider`. It maps `Track` → `Add`, `Update` → EF `Update`, and `SaveChanges` → `SaveChangesAsync`, delegating persistence to EF Core's change tracker. It also provides transaction support via `BeginTransaction`, `CommitTransaction`, and `RollbackTransaction`.
+`DataContext<TDbContext>` extends EF Core's `DbContext` and implements both `IDataContext` and `IEffectProvider`. It maps `Track` and `Update` to direct entity state assignment (Added for new entities, Modified for existing), and `SaveChanges` to `SaveChangesAsync`. State is set on the target entity only — navigation properties are not traversed, which prevents unnecessary UPDATE statements when entities cross DI scope boundaries (e.g., metadata loaded by `LoadMetadataJunction` passed into a child train's context). It also provides transaction support via `BeginTransaction`, `CommitTransaction`, and `RollbackTransaction`.
 
 **DbSets:**
 
diff --git a/sdk-reference/scheduler-api/add-trax-job-runner.md b/sdk-reference/scheduler-api/add-trax-job-runner.md
index 8130a0d..26b0c89 100644
--- a/sdk-reference/scheduler-api/add-trax-job-runner.md
+++ b/sdk-reference/scheduler-api/add-trax-job-runner.md
@@ -167,7 +167,7 @@ Maps a `POST` endpoint at the specified route that:
 3. Creates a new DI scope and resolves `IJobRunnerTrain`
 4. Calls `Run(new RunJobRequest(metadataId, input))`
 5. Returns `200 OK` with a `RemoteJobResponse` containing the metadata ID on success
-6. On error: returns `200 OK` with `RemoteJobResponse.IsError = true` and structured error fields (`ErrorMessage`, `ExceptionType`, `StackTrace`)
+6. On error: logs the exception, then returns `200 OK` with `RemoteJobResponse.IsError = true` and structured error fields (`ErrorMessage`, `ExceptionType`, `StackTrace`)
 
 ### UseTraxRunEndpoint
 
diff --git a/sdk-reference/scheduler-api/trax-lambda-function.md b/sdk-reference/scheduler-api/trax-lambda-function.md
index 1a67d34..ac36b74 100644
--- a/sdk-reference/scheduler-api/trax-lambda-function.md
+++ b/sdk-reference/scheduler-api/trax-lambda-function.md
@@ -164,9 +164,9 @@ To minimize cold start time:
 
 ## Error Handling
 
-For `Execute` requests, exceptions are caught and returned as a `RemoteJobResponse` with structured error fields (`IsError`, `ErrorMessage`, `ExceptionType`, `StackTrace`). The `LambdaJobSubmitter` on the scheduler side does not read this response (fire-and-forget), but the error is persisted in the Lambda function's metadata.
+For `Execute` requests, exceptions are logged and returned as a `RemoteJobResponse` with structured error fields (`IsError`, `ErrorMessage`, `ExceptionType`, `StackTrace`). Errors that occur within the train itself are also persisted to the `Metadata` table by `ServiceTrain.Run`. However, pre-train errors (e.g., deserialization failures) only appear in the log output. The `LambdaJobSubmitter` on the scheduler side does not read the response (fire-and-forget).
 
-For `Run` requests, `ITraxRequestHandler.RunTrainAsync` returns a `RemoteRunResponse` that may contain structured error fields. The `LambdaRunExecutor` on the scheduler side reads the response and reconstructs a `TrainException` with the full error context.
+For `Run` requests, exceptions are logged before being rethrown. `ITraxRequestHandler.RunTrainAsync` returns a `RemoteRunResponse` that may contain structured error fields. The `LambdaRunExecutor` on the scheduler side reads the response and reconstructs a `TrainException` with the full error context.
 
 ## See Also