diff --git a/AGENTS.md b/AGENTS.md
index 9b1e26a987ef..ab2561e8a135 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -91,8 +91,9 @@ When adding tests:
   the cheapest level that actually reproduces it? Fixes to user-visible
   behavior need a test through a real entry point (`LLM`, OpenAI server); see
   [`tests/AGENTS.md`](tests/AGENTS.md).
-- **Reuse before create.** Extend existing test files, `conftest.py` fixtures, and
-  helpers; add a new file only when no nearby suite fits.
+- **Reuse before create.** First try adding a case to an existing parametrized
+  test, then extending or refactoring existing tests, `conftest.py` fixtures, and
+  helpers; add a new test or file only when no nearby suite fits.
 - **Test behavior with intent.** Assert observable outcomes through public APIs;
   state why in the name or docstring. Skip trivial wiring; flaky tests are worse
   than no tests.
diff --git a/tests/AGENTS.md b/tests/AGENTS.md
index 5cd967bb7f33..b8b73dc94a94 100644
--- a/tests/AGENTS.md
+++ b/tests/AGENTS.md
@@ -1,14 +1,22 @@
 # tests
 
-- **Test real objects on the lightest real path.** Construct vLLM objects with
-  their real constructors; do not stand them in with `SimpleNamespace` or
-  `MagicMock`, set private attributes, or patch methods to reach a state. Mock
-  only external services. Pick the cheapest setup that exercises the change: a
-  layer or kernel test does not need `vllm_runner`, and single-rank parallel
-  state is enough when parallelism is irrelevant. A fix to user-visible behavior
-  needs at least one regression test through `LLM` (`vllm_runner`) or
-  `RemoteOpenAIServer`, with real weights when loading is involved. A test that
-  only asserts `assert_called*` or "no exception" guards nothing.
+These are defaults, not absolutes: when one does not fit, deviate and say why in
+the PR. Never weaken a test to make it pass.
+
+- **Choose the level deliberately.** End-to-end tests through `LLM`
+  (`vllm_runner`) or `RemoteOpenAIServer` survive refactors but take seconds to
+  minutes; unit tests on pure functions iterate fast but cover only a slice, so
+  factor complex logic into pure functions where you can. A fix to
+  user-visible behavior needs at least one end-to-end regression test, with real
+  weights when loading is involved. A layer or kernel test does not need
+  `vllm_runner`; single-rank parallel state is enough when parallelism is
+  irrelevant.
+- **Prefer real objects.** Construct vLLM objects with their real constructors
+  rather than `SimpleNamespace` or `MagicMock`, setting private attributes, or
+  patching methods; such tests couple to the implementation. Mock external
+  services, or an object that is impractical to construct, and say why in the
+  test. A test that only asserts `assert_called*` or "no exception" guards
+  nothing.
 - **Show the failure before the fix.** When the bug is reproducible, run the new
   test on the base commit (rebuild if the fix touches `csrc/`) and confirm it
   fails for the reported reason, then passes with the fix. Paste both commands
@@ -20,15 +28,18 @@
   on main (`.buildkite/scripts/rerun-test.sh`,
   `docs/contributing/ci/failures.md`); the only exception is
   `xfail(strict=True)` linking an open issue for a bug on main.
-- **Deterministic and exact by construction.** Use greedy or seeded sampling.
-  Assert equality where the result is exact (a change that must not alter
-  output is compared against the code it replaces, not another reference), and
-  keep tolerances as tight as the math allows. vLLM is not batch invariant by default: batch size, chunked
-  prefill, prefix caching, backend, and TP size can flip greedy tokens, so
-  compare against a reference with `check_logprobs_close` /
-  `check_embeddings_close`, or require `VLLM_BATCH_INVARIANT=1` for exact
-  equality. Never synchronize with a fixed `sleep`; poll an explicit condition
-  with a timeout, and force race orderings with events.
+- **Deterministic and exact by construction.** Use greedy or seeded sampling;
+  when the behavior under test is sampling itself (temperature > 0), fix the
+  seed and assert a statistic against a stated bound (p-value, confidence
+  interval) rather than exact tokens. Assert equality where the result is exact
+  (a change that must not alter output is compared against the code it
+  replaces, not another reference), and keep tolerances as tight as the math
+  allows. vLLM is not batch invariant by default: batch size, chunked prefill,
+  prefix caching, backend, and TP size can flip greedy tokens, so compare
+  against a reference with `check_logprobs_close` / `check_embeddings_close`,
+  or require `VLLM_BATCH_INVARIANT=1` for exact equality. Never synchronize
+  with a fixed `sleep`; poll an explicit condition with a timeout, and force
+  race orderings with events.
 - **Make sure CI runs it.** A test no step in `.buildkite/test_areas/*.yaml`
   selects never runs, which looks like passing. Check that the step's
   `commands` and pytest markers (`core_model`, `slow_test`, `distributed`)