diff --git a/docs/contributing/README.md b/docs/contributing/README.md
index e6155b33a..63921fba6 100644
--- a/docs/contributing/README.md
+++ b/docs/contributing/README.md
@@ -106,6 +106,151 @@ sudo mv ./bin/holodeck /usr/local/bin/holodeck
 - Update existing tests when modifying features
 - Run the full test suite with `make test`
 
+## E2E Testing
+
+Holodeck's end-to-end tests run on real AWS infrastructure. They are organized
+into two tiers that control when tests execute in CI.
+
+### E2E Test Structure
+
+**Smoke tier (pre-merge)** — `.github/workflows/e2e-smoke.yaml`
+
+Runs on every PR push. Covers two label filters:
+
+- `default && !rpm` — standard single-node environment without RPM distros
+- `cluster && minimal` — smallest valid multinode cluster
+
+Each job takes roughly 20 minutes, giving fast feedback before merge.
+
+**Full tier (post-merge)** — `.github/workflows/e2e.yaml`
+
+Runs only when a commit lands on `main` or a `release-*` branch.
+Covers 13 label filters plus a separate arm64 job and an
+integration-test job that exercises holodeck as a GitHub Action.
+
+| Label filter | What it covers |
+|---|---|
+| `legacy` | Kubernetes using a legacy version |
+| `dra` | Dynamic Resource Allocation enabled |
+| `kernel` | Kernel features / custom kernel |
+| `ctk-git` | Container Toolkit installed from git source |
+| `k8s-git` | Kubernetes built from git (kubeadm) |
+| `k8s-kind-git` | Kubernetes built from git (KIND) |
+| `k8s-latest` | Kubernetes tracking master branch |
+| `cluster && gpu && !minimal && !ha && !dedicated` | Standard GPU cluster |
+| `cluster && dedicated` | Cluster with dedicated CPU control-plane |
+| `cluster && ha` | HA cluster (3 control-plane nodes) |
+| `rpm-rocky` | Rocky Linux 9 — multiple container runtimes |
+| `rpm-al2023` | Amazon Linux 2023 — multiple container runtimes |
+| `rpm-fedora` | Fedora 42 — multiple container runtimes |
+
+The `arm64` job is a separate workflow job (not a matrix entry) that only
+runs on `main`. It uses `--label-filter='arm64'` — a test must carry
+`Label("arm64")` to be selected.
+
+### Label Taxonomy
+
+Tests are tagged with Ginkgo `Label()` annotations. Each test can carry
+multiple labels; CI selects tests using boolean filter expressions.
+
+**Single-node labels** (defined in `tests/aws_test.go`):
+
+| Label | Description |
+|---|---|
+| `default` | Basic AWS environment, default configuration |
+| `legacy` | Legacy Kubernetes version |
+| `dra` | Dynamic Resource Allocation |
+| `kernel` | Custom kernel features |
+| `ctk-git` | CTK from git source |
+| `k8s-git` | Kubernetes from git (kubeadm) |
+| `k8s-kind-git` | Kubernetes from git (KIND) |
+| `k8s-latest` | Kubernetes master branch |
+| `rpm` | Any RPM-based distribution |
+| `rpm-rocky` | Rocky Linux 9 |
+| `rpm-al2023` | Amazon Linux 2023 |
+| `rpm-fedora` | Fedora 42 |
+| `post-merge` | Excluded from smoke tier; full tier only |
+
+**Cluster labels** (defined in `tests/aws_cluster_test.go`):
+
+| Label | Description |
+|---|---|
+| `cluster` | Multinode cluster test |
+| `multinode` | Two or more nodes |
+| `gpu` | GPU worker nodes |
+| `minimal` | Smallest valid configuration (1 CP + 1 worker) |
+| `dedicated` | Dedicated CPU control-plane node |
+| `ha` | High-availability control plane (3 nodes) |
+| `rpm` | RPM-based cluster OS |
+| `rpm-rocky` | Rocky Linux 9 cluster |
+| `rpm-al2023` | Amazon Linux 2023 cluster |
+| `post-merge` | Excluded from smoke tier; full tier only |
+
+The `post-merge` label is the mechanism that keeps a test out of the smoke
+tier. The smoke workflow's label filter `"default && !rpm"` already excludes
+RPM tests, but adding `post-merge` makes the intent explicit and ensures the
+test is skipped by any future smoke filter that might otherwise match it.
+
+### How to Add New Tests
+
+1. **Single-node test** — add an `Entry(...)` to the `DescribeTable` in
+   `tests/aws_test.go`.
+1. **Cluster test** — add an `Entry(...)` to `tests/aws_cluster_test.go`.
+1. Create the corresponding fixture file under `tests/data/`.
+1. Assign Ginkgo labels with `Label("label1", "label2", ...)` as the last
+   argument of the `Entry`.
+1. If the test is an edge case, platform-specific variant, or is expensive
+   (> 30 min), add `"post-merge"` to its label list so it runs only in the
+   full tier.
+
+Example:
+
+```go
+Entry("My New Feature Test", testConfig{
+    name:        "my-feature-test",
+    filePath:    filepath.Join(packagePath, "data", "test_aws_my_feature.yml"),
+    description: "Tests my new feature end-to-end",
+}, Label("default", "my-feature")),
+```
+
+### Which Tier to Target
+
+| Use smoke tier (no `post-merge`) | Use full tier (`post-merge`) |
+|---|---|
+| Core functionality every PR should validate | Edge cases and less-common paths |
+| Fast tests (< 25 min) | Slow tests (> 30 min) |
+| Platform-agnostic defaults | Platform-specific variants (RPM distros, arm64) |
+| Minimal cluster configurations | Full-scale, HA, or dedicated cluster topologies |
+
+If in doubt, start with `post-merge` and promote the label out of the full
+tier once the test has demonstrated stability.
+
+### Running E2E Tests Locally
+
+Use the Ginkgo label filter to select which tests to run:
+
+```bash
+# Run only the smoke-equivalent tests
+make -f tests/Makefile test GINKGO_ARGS="--label-filter='default && !rpm'"
+
+# Run a specific label
+make -f tests/Makefile test GINKGO_ARGS="--label-filter='cluster && minimal'"
+
+# Run all RPM tests for Rocky 9
+make -f tests/Makefile test GINKGO_ARGS="--label-filter='rpm-rocky'"
+```
+
+Required environment variables:
+
+```bash
+export AWS_ACCESS_KEY_ID=<your-key-id>
+export AWS_SECRET_ACCESS_KEY=<your-secret>
+export E2E_SSH_KEY=<path-to-ssh-private-key>
+```
+
+> **Important:** Always validate E2E tests locally before pushing. CI E2E
+> runs provision real GPU instances on AWS, and unnecessary runs are expensive.
+
 ## Documentation
 
 - Update relevant documentation when adding features