Add tips for debugging correctness issues. (#8763)
diff --git a/docs/developers/debugging/integration_correctness_issue_breakdown.md b/docs/developers/debugging/integration_correctness_issue_breakdown.md
new file mode 100644
index 0000000..72aa41b
--- /dev/null
+++ b/docs/developers/debugging/integration_correctness_issue_breakdown.md
@@ -0,0 +1,51 @@
+This doc describes tips about triaging correctness issue. Feel free to reach out
+to @hanhanW or ask questions on Discord if you need help or tips on triaging
+correctness issue.
+
+# Decouple the reproduce from integrations.
+
+## TF integration tests
+
+See [instructions for reproducing failures in TF/TFLite integration tests](https://github.com/hanhanW/iree/blob/main/docs/developers/debugging/tf_integrations_test_repro.md).
+
+For input data, they are not dumped within the flagfile. You can construct the
+function inputs by looking into `log.txt`. There is an [issue](https://github.com/google/iree/issues/8658)
+for tracking this.
+
+## iree-samples
+
+Follow [README](https://github.com/google/iree-samples#readme) to run the model.
+The MLIR files will be generated. You'll find the saved file from log. E.g.,
+
+```
+[ RUN ] MobilenetV2Int8Test.test_compile_tflite
+I0401 17:27:04.084272 140182373025024 test_util.py:119] Setting up for IREE
+I0401 17:27:04.085064 140182373025024 binaries.py:218] Invoke IREE Pipeline:
+ /tmp/iree-samples/iree-samples.venv/lib/python3.9/site-packages/iree/tools/tflite/iree-import-tflite
+ /tmp/iree-samples/tflitehub/tmp/mobilenet_v2_int8_test.py/model.tflite
+ --mlir-print-debuginfo
+ --save-temp-tfl-input=/tmp/iree-samples/tflitehub/tmp/mobilenet_v2_int8_test.py/tflite.mlir
+ --save-temp-iree-input=/tmp/iree-samples/tflitehub/tmp/mobilenet_v2_int8_test.py/tosa.mlir
+```
+
+Unfortunately, the artifacts are not dumped in the runs. There is an [issue](https://github.com/google/iree/issues/8756)
+for tracking this. A workaround can be found in the issue.
+
+# Narrow down the repro
+
+The model itself is big. IREE breaks a model into dispatches and launches the
+kernels. The inputs and outputs could be diverged starting from one of
+launches. To get a smaller reproduce, you can use [-iree-flow-trace-dispatch-tensors](https://github.com/google/iree/blob/main/docs/developers/developing_iree/developer_overview.md#iree-flow-trace-dispatch-tensors).
+You can compare the logs between builds/backends, and get the idea about which
+dispatch results in wrong outputs. The dumped inputs can be reused in a
+flagfile.
+
+Since we get the suspicious dispatch, we are able to create a test case based on
+the dispatch function. The dispatch function can be derived after the
+`OutlineDispatchRegions` pass. The function signatures have to be modified
+manually. You'll have to put `flow.dispatch.tensor.load` variables to function
+arguments, and replace `flow.dispatch.tensor.store` with `return` op.
+
+At this stage, the reproduce is narrowed down to a single dispatch function.
+
+Note: This only works when dispatch formation logics are identical between runs.
