diff --git a/.github/workflows/coverity.yml b/.github/workflows/coverity.yml
index 86c4e66..22e4d37 100644
--- a/.github/workflows/coverity.yml
+++ b/.github/workflows/coverity.yml
@@ -50,7 +50,9 @@ jobs:
           architecture: x64
 
       - name: Install build dependencies
-        run: pip install meson-python ninja cmake cython "numpy>=2" mkl-devel
+        # Cython is pinned here only (not in pyproject.toml) to keep the generated
+        # code stable between scans, so Coverity CIDs and their triage survive
+        run: pip install meson-python ninja cmake "cython==3.3.0" "numpy>=2" mkl-devel
 
       - name: Download Coverity Build Tool
         timeout-minutes: 15
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 7319acd..0346ac3 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Added
 
 ### Changed
+* Pinned Cython in the Coverity Scan workflow so generated code stays stable between scans, and added `coverity/README.md` documenting the known Cython-boilerplate false positives and the scan review checklist [gh-164](https://github.com/IntelPython/mkl_random/pull/164)
 
 ### Fixed
 * Fixed `uniform` to return a Python `float` for scalar bounds with `size=None` instead of a 0-d array [gh-167](https://github.com/IntelPython/mkl_random/pull/167)
diff --git a/coverity/README.md b/coverity/README.md
new file mode 100644
index 0000000..8c77638
--- /dev/null
+++ b/coverity/README.md
@@ -0,0 +1,82 @@
+# Triaging Coverity Scan findings
+
+Static analysis runs on [Coverity Scan](https://scan.coverity.com) via
+[`.github/workflows/coverity.yml`](../.github/workflows/coverity.yml) (weekly + on
+demand). Analysis runs on Black Duck's servers; triage is done in the Scan web UI.
+
+Most findings are false positives in Cython-**generated** `mklrand.cpp`, not in
+code we maintain. This guide records why, and how to keep triage from resetting.
+
+## Where findings come from
+
+- **Generated** `mklrand.cpp` (from `mklrand.pyx`): `__Pyx_*` / `__pyx_pw_*` /
+  `__pyx_tp_*` helpers and wrappers are boilerplate — findings here are ~always
+  false positives (see the table below). `__pyx_pf_*` functions are the C
+  translation of our `.pyx` bodies; a real `.pyx` bug could surface here, though
+  so far all have been false positives too (Coverity can't see Python-level
+  invariants like non-negative dtype sizes or fixed-length tuples).
+- **Hand-written** C++ under `mkl_random/src/`. Most likely place for a real bug.
+
+## Keeping triage durable: the Cython pin
+
+A Cython *version* bump regenerates `mklrand.cpp` wholesale, which churns the
+Coverity CIDs and silently drops their triage — the same boilerplate then returns
+under new CIDs. So **Cython is pinned in `coverity.yml`** (not `pyproject.toml`,
+so shipped wheels are unaffected). The pin works only because the build runs with
+`--no-build-isolation`; bumping it means re-triaging the boilerplate.
+
+## Reducing the noise: a Project Component
+
+**Project Settings → Components** buckets defects by a path regex. Define one to
+group (not hide) the generated file so it can be filtered out of view — path-based,
+so it survives Cython bumps:
+
+- **Name:** `Cython-generated`  **Path regex:** `.*/mklrand\.cpython.*`
+
+This matches only the generated file (hand-written sources are under
+`mkl_random/src/`, no `mklrand` substring). Group only — do **not** mark it
+*ignored*, as that also drops the `__pyx_pf_*` bodies (see [declined](#evaluated-and-declined)).
+
+## Review checklist
+
+Don't blanket-ignore the generated file — prioritise instead:
+
+1. **Findings under `mkl_random/src/`** — review every one.
+2. **High/Medium findings in `__pyx_pf_*`** — verify against the `.pyx`; if it's a
+   Python-level invariant Coverity can't see, mark `False Positive` with a reason.
+3. **Boilerplate families below** — bulk-triage `False Positive` / `Ignore`.
+
+## Known false-positive families
+
+Match on **checker + mechanism**, not CID (CIDs reset on a Cython bump or engine
+upgrade). Helper names below are from Cython 3.3.0 and vary between versions.
+
+| Family | Checker | Why it's a false positive |
+| --- | --- | --- |
+| `__pyx_tp_traverse_*`, `__Pyx_CyFunction_traverse` | DEADCODE | `__Pyx_call_type_traverse` is a constant-`0` macro on standard CPython, so `if (e) return e;` is dead; live only in the Limited-API build. |
+| `__Pyx_PyCode_New`, `__Pyx_CallSlotAsVectorcallUnpackDict` | DEADCODE | `__Pyx_PyTuple_SET_ITEM` expands to void `PyTuple_SET_ITEM` yielding `0`, so `if (... != 0)` is dead; live only in the Limited-API build. |
+| `__Pyx_AddTraceback` | DEADCODE | `c_line` is `0` unless the optional `CYTHON_CLINE_IN_TRACEBACK` feature is on, so the `-c_line` branch is dead by default. |
+| `__Pyx_ParseKeywordDict` | DEADCODE | In the CPython < 3.13 branch `found` is only 0/1, so `if (found < 0)` is dead; that guard serves the ≥ 3.13 `PyDict_GetItemRef` path. |
+| `__Pyx_PyLong_As_*` (per C type) | DEADCODE | Integer-conversion helper templates; dead branches are compile-time-selected version/overflow guards. |
+| `__pyx_pf_*` temp cleanup (e.g. `choice`, `multivariate_normal`) | UNUSED_VALUE | `__pyx_t_N = 0;` nulls a temporary after its ref is transferred, guarding the error path against a double-DECREF; dead only on the straight-line path. |
+| `__pyx_pw_*` keyword wrappers (`rand`, `randn`, …) | CHECKED_RETURN | `PyDict_Size` result is checked on the *next* line (`if (... < 0) __PYX_ERR(...)`); the statistical heuristic misfires because it's one line from the call. |
+
+`sanity_check_for_cython` (DEADCODE) is the same disposition but comes from
+meson's compiler-probe unit, not `mklrand.cpp`.
+
+**Verified individually — OUT_OF_BOUNDS in `_seed_impl` (High):** the unpack
+`brng_token, stream_id = _parse_brng_argument(brng)` (`mklrand.pyx:1571`) generates
+`PyTuple_GET_ITEM(sequence, 1)`, guarded by `if (unlikely(size != 2)) __PYX_ERR(...)`,
+and `_parse_brng_argument` always returns a fixed 2-tuple (`mklrand.pyx:1533`).
+Coverity just doesn't tie the index back to the guard. No overflow possible.
+
+## Evaluated and declined
+
+- **Modeling files** correct the behavior of *called* functions; our FPs are
+  intraprocedural (dead branches, macros, `#if`), which models can't reach. The
+  only fit is the `PyDict_Size` CHECKED_RETURN case — not worth the maintenance.
+- **Dropping the generated unit** (hard exclude), e.g. after `cov-build`:
+  ```bash
+  cov-manage-emit --dir cov-int --tu-pattern "file('.*mklrand.*\\.cpp')" delete
+  ```
+  Also drops the `__pyx_pf_*` bodies, so it's disabled in favour of the checklist.