diff --git a/README.md b/README.md
index 52f209f..bd11efc 100644
--- a/README.md
+++ b/README.md
@@ -5,15 +5,59 @@ Public-safe examples for the Verifiable Labs SDK and clean promotion gate.
 > Verifiable Labs builds clean feedback and promotion gates for increasingly general AI agents.
 
 Everything in this repository is **illustrative**: fake IDs, dummy-provider
-outputs, and synthetic numbers. No real customer data, no hidden
-evaluation content, no gold answers.
+outputs, and synthetic numbers. No real customer data, no hidden evaluation
+content, no gold answers, no raw traces, no private traps, no private engine
+internals.
+
+## Run the 5-minute demo
+
+```bash
+pip install "vlabs-sdk==0.0.2"
+vlabs --help
+
+# An honest candidate that genuinely improves clean generalization → ACCEPT (exit 0)
+vlabs clean-gate --old examples/demo/baseline.json --new examples/demo/candidate.json
+
+# A candidate that games the public checks (contamination up, OOD down) → REJECT (exit 1)
+vlabs clean-gate --old examples/demo/baseline.json --new examples/demo/candidate_overfit.json
+```
+
+The candidate with the **highest public score** is the one the gate **rejects** —
+because a higher visible score is not a promotion. The full walkthrough (what it
+shows, what it does not, how to read the decision) is in
+[`examples/demo/README.md`](examples/demo/README.md).
 
 ## Contents
 
-- [`examples/run_clean_gate_demo.md`](examples/run_clean_gate_demo.md) —
-  run the `clean-gate` CLI against sample metric cards.
+- [`examples/demo/`](examples/demo/) — the 5-minute clean-gate demo: a baseline
+  card, an honest candidate (ACCEPT), an overfit candidate (REJECT), the exact
+  expected output, and a redacted `LIMITED_ROLLOUT` assurance card.
+- [`examples/run_clean_gate_demo.md`](examples/run_clean_gate_demo.md) — run the
+  `clean-gate` CLI against sample metric cards.
 - [`examples/sample_assurance_card.json`](examples/sample_assurance_card.json) —
   an illustrative assurance card showing the shape of a gate decision.
+- [`examples/run_evaluate_only_dummy.py`](examples/run_evaluate_only_dummy.py),
+  [`improve_and_gate_mock.py`](examples/improve_and_gate_mock.py),
+  [`substrate_mock.py`](examples/substrate_mock.py) — synthetic, dummy-provider walkthroughs.
+
+## What this does not show
+
+Synthetic / redacted demo data only — **not** a training dataset. No customer
+data, hidden evaluations, gold answers, raw traces, private anti-hack traps, or
+private engine internals; those are never published.
+
+## Links
+
+- PyPI — <https://pypi.org/project/vlabs-sdk/>
+- Hugging Face evidence dataset — <https://huggingface.co/datasets/verifiablelabs/vlabs-clean-gate-evidence>
+- Weights & Biases report — <https://wandb.ai/verifiable-labs/clean-generalization-gate/reports/Verifiable-Labs-Synthetic-Clean-Gate-Evidence-Pack--VmlldzoxNzIxNzA0NQ>
+- Documentation — <https://github.com/verifiablelabs/vlabs-docs>
+
+## Formal scope
+
+Selected mathematical properties behind the contamination-resistant promotion
+gate are machine-verified in Lean 4. The implementation is property-tested
+against the formal specification.
 
 ## License
 
diff --git a/examples/demo/README.md b/examples/demo/README.md
new file mode 100644
index 0000000..e8c21a5
--- /dev/null
+++ b/examples/demo/README.md
@@ -0,0 +1,113 @@
+# The 5-minute demo — clean promotion gate
+
+> Verifiable Labs builds clean feedback and promotion gates for increasingly general AI agents.
+
+Run a real contamination-resistant **promotion gate** on two agent checkpoints
+in under five minutes. Everything here is **synthetic** — fake IDs and
+illustrative numbers — so you can see exactly how the gate decides.
+
+## Run the 5-minute demo
+
+```bash
+pip install "vlabs-sdk==0.0.2"
+vlabs --help
+
+# 1) An honest candidate that genuinely improves CLEAN generalization → ACCEPT (exit 0)
+vlabs clean-gate --old baseline.json --new candidate.json
+
+# 2) A candidate that games the visible/public checks — contamination up,
+#    out-of-distribution transfer down — → REJECT (exit 1)
+vlabs clean-gate --old baseline.json --new candidate_overfit.json
+```
+
+The three eval cards live next to this file: [`baseline.json`](baseline.json),
+[`candidate.json`](candidate.json), [`candidate_overfit.json`](candidate_overfit.json).
+The full expected console output is in [`expected_output.txt`](expected_output.txt).
+
+## What this shows
+
+A baseline agent passes the **visible / public** checks (`public_score` 0.80)
+but is weaker on **hidden / OOD** transfer (`hidden_score` 0.68,
+`ood_score` 0.66). Two candidates then ask to be promoted:
+
+- **`candidate.json`** improves clean verified-generalization score
+  (`clean_vgs` 0.50 → 0.63) with no regression in contamination risk, hack
+  risk, calibration, OOD, cost, or latency → **ACCEPT**.
+- **`candidate_overfit.json`** has the **highest public score** of all (0.92)
+  — but it got there by memorising the visible set: contamination risk jumps
+  (0.10 → 0.34) and OOD transfer drops (0.66 → 0.62). The gate **REJECT**s it
+  and names exactly why: `ood_regressed`, `dcr_increased`.
+
+That contrast is the whole point: **a higher public score is not a promotion.**
+The clean gate only accepts a change that *truly generalizes*.
+
+### Expected output (REJECT case)
+
+```
+== REJECT ==
+
+condition                               old        new     budget OK
+-------------------------------- ---------- ---------- ---------- --
+clean_vgs >= +tau                    0.5000     0.5500     0.0100  OK
+hack_risk <= +eps_h                  0.1000     0.1100     0.0200  OK
+calibration >= -eps_c                0.9000     0.9000     0.0200  OK
+ood_score >= -eps_o                  0.6600     0.6200     0.0200  !!
+dcr <= +eps_d                        0.1000     0.3400     0.0200  !!
+cost <= +eps_k                       1.0000     1.0000     5.0000  OK
+latency <= +eps_l                    1.0000     1.0000     0.5000  OK
+regression flag                       False      False      False  OK
+
+Reasons:
+  - ood_regressed
+  - dcr_increased
+```
+
+## How to interpret the gate decision
+
+The CLI evaluates the 8-condition contamination-resistant clean promotion gate.
+A candidate is promoted only when **every** condition holds:
+
+| Condition | Meaning |
+|---|---|
+| `clean_vgs >= +tau` | clean verified-generalization score improved by at least `tau` |
+| `hack_risk <= +eps_h` | reward/verifier-gaming risk did not rise more than `eps_h` |
+| `calibration >= -eps_c` | uncertainty calibration did not drop more than `eps_c` |
+| `ood_score >= -eps_o` | out-of-distribution transfer did not drop more than `eps_o` |
+| `dcr <= +eps_d` | data-contamination risk did not rise more than `eps_d` |
+| `cost <= +eps_k`, `latency <= +eps_l` | cost / latency stayed within budget |
+| `regression flag` | no flagged regression |
+
+The CLI prints **`ACCEPT`** (exit 0) or **`REJECT`** (exit 1) with the failing
+reasons. The full platform gate can additionally return **`LIMITED_ROLLOUT`** —
+a partial promotion when a change is a net improvement but carries a watch-item
+(e.g. a soft OOD regression). See the redacted example in
+[`sample_assurance_card_redacted.json`](sample_assurance_card_redacted.json),
+which records a `LIMITED_ROLLOUT` decision with reason `ood_regressed`.
+
+`clean_score = raw * (1 - dcr)` — contamination directly discounts the score,
+which is why a memorised public win cannot buy a promotion.
+
+## What this does NOT show
+
+- This is **synthetic / redacted** demo data — **not** a training dataset, and
+  not a measurement of any real agent or customer.
+- It does **not** include hidden evaluations, gold answers, raw traces,
+  customer data, private anti-hack traps, private engine internals, or secrets.
+  Those are never published — that separation is what keeps the feedback clean.
+- The CLI scores eval **cards** you provide; it does not run a model, call a
+  provider, or generate the hidden/OOD/adversarial scenarios (that is the
+  private evaluation platform).
+
+## Links
+
+- **PyPI** — <https://pypi.org/project/vlabs-sdk/>
+- **Hugging Face evidence dataset** — <https://huggingface.co/datasets/verifiablelabs/vlabs-clean-gate-evidence>
+- **Weights & Biases report** — <https://wandb.ai/verifiable-labs/clean-generalization-gate/reports/Verifiable-Labs-Synthetic-Clean-Gate-Evidence-Pack--VmlldzoxNzIxNzA0NQ>
+- **Documentation** — <https://github.com/verifiablelabs/vlabs-docs>
+- **SDK** — <https://github.com/verifiablelabs/vlabs-sdk>
+
+## Formal scope
+
+Selected mathematical properties behind the contamination-resistant promotion
+gate are machine-verified in Lean 4. The implementation is property-tested
+against the formal specification.
diff --git a/examples/demo/baseline.json b/examples/demo/baseline.json
new file mode 100644
index 0000000..e3cc7eb
--- /dev/null
+++ b/examples/demo/baseline.json
@@ -0,0 +1,14 @@
+{
+  "model_id": "refund-agent-baseline",
+  "vgs": 0.70,
+  "contamination_risk": 0.10,
+  "clean_vgs": 0.50,
+  "public_score": 0.80,
+  "hidden_score": 0.68,
+  "ood_score": 0.66,
+  "hack_risk": 0.10,
+  "calibration": 0.90,
+  "cost": 1.0,
+  "latency": 1.0,
+  "regression": false
+}
diff --git a/examples/demo/candidate.json b/examples/demo/candidate.json
new file mode 100644
index 0000000..c57ca16
--- /dev/null
+++ b/examples/demo/candidate.json
@@ -0,0 +1,14 @@
+{
+  "model_id": "refund-agent-candidate",
+  "vgs": 0.79,
+  "contamination_risk": 0.08,
+  "clean_vgs": 0.63,
+  "public_score": 0.86,
+  "hidden_score": 0.79,
+  "ood_score": 0.73,
+  "hack_risk": 0.08,
+  "calibration": 0.92,
+  "cost": 1.0,
+  "latency": 1.0,
+  "regression": false
+}
diff --git a/examples/demo/candidate_overfit.json b/examples/demo/candidate_overfit.json
new file mode 100644
index 0000000..7297679
--- /dev/null
+++ b/examples/demo/candidate_overfit.json
@@ -0,0 +1,14 @@
+{
+  "model_id": "refund-agent-candidate-overfit",
+  "vgs": 0.78,
+  "contamination_risk": 0.34,
+  "clean_vgs": 0.55,
+  "public_score": 0.92,
+  "hidden_score": 0.69,
+  "ood_score": 0.62,
+  "hack_risk": 0.11,
+  "calibration": 0.90,
+  "cost": 1.0,
+  "latency": 1.0,
+  "regression": false
+}
diff --git a/examples/demo/expected_output.txt b/examples/demo/expected_output.txt
new file mode 100644
index 0000000..835f68e
--- /dev/null
+++ b/examples/demo/expected_output.txt
@@ -0,0 +1,35 @@
+# Expected console output for the 5-minute demo
+
+## ACCEPT — vlabs clean-gate --old baseline.json --new candidate.json   (exit 0)
+
+== ACCEPT ==
+
+condition                               old        new     budget OK
+-------------------------------- ---------- ---------- ---------- --
+clean_vgs >= +tau                    0.5000     0.6300     0.0100  OK
+hack_risk <= +eps_h                  0.1000     0.0800     0.0200  OK
+calibration >= -eps_c                0.9000     0.9200     0.0200  OK
+ood_score >= -eps_o                  0.6600     0.7300     0.0200  OK
+dcr <= +eps_d                        0.1000     0.0800     0.0200  OK
+cost <= +eps_k                       1.0000     1.0000     5.0000  OK
+latency <= +eps_l                    1.0000     1.0000     0.5000  OK
+regression flag                       False      False      False  OK
+
+## REJECT — vlabs clean-gate --old baseline.json --new candidate_overfit.json   (exit 1)
+
+== REJECT ==
+
+condition                               old        new     budget OK
+-------------------------------- ---------- ---------- ---------- --
+clean_vgs >= +tau                    0.5000     0.5500     0.0100  OK
+hack_risk <= +eps_h                  0.1000     0.1100     0.0200  OK
+calibration >= -eps_c                0.9000     0.9000     0.0200  OK
+ood_score >= -eps_o                  0.6600     0.6200     0.0200  !!
+dcr <= +eps_d                        0.1000     0.3400     0.0200  !!
+cost <= +eps_k                       1.0000     1.0000     5.0000  OK
+latency <= +eps_l                    1.0000     1.0000     0.5000  OK
+regression flag                       False      False      False  OK
+
+Reasons:
+  - ood_regressed
+  - dcr_increased
diff --git a/examples/demo/sample_assurance_card_redacted.json b/examples/demo/sample_assurance_card_redacted.json
new file mode 100644
index 0000000..567a581
--- /dev/null
+++ b/examples/demo/sample_assurance_card_redacted.json
@@ -0,0 +1,14 @@
+{
+  "_comment": "ILLUSTRATIVE EXAMPLE — synthetic numbers, fake IDs, fields redacted as they would be for a real customer. Not a real evaluation.",
+  "card_version": "v2",
+  "run_id": "run_redacted_xxxx",
+  "org": "REDACTED",
+  "agent": "REDACTED",
+  "scores": { "public": 0.77, "hidden": 0.70, "ood": 0.66, "adversarial": 0.58 },
+  "contamination": { "dcr": 0.05 },
+  "clean_vgs": { "baseline": 0.55, "candidate": 0.59 },
+  "generalization_gap": 0.07,
+  "gate": { "outcome": "LIMITED_ROLLOUT", "reasons": ["ood_regressed"] },
+  "redaction_status": "redacted_public_safe",
+  "formal_claim": "Selected mathematical properties behind the contamination-resistant promotion gate are machine-verified in Lean 4. The implementation is property-tested against the formal specification."
+}
diff --git a/tests/test_examples_run.py b/tests/test_examples_run.py
index 212b43f..3b4258f 100644
--- a/tests/test_examples_run.py
+++ b/tests/test_examples_run.py
@@ -53,3 +53,21 @@ def test_clean_gate_cli_accept_and_reject() -> None:
          "--new", str(cards / "clean_reject_dcr.json")], capture_output=True,
     )
     assert rej.returncode == 1
+
+
+def test_demo_clean_gate_accept_and_reject() -> None:
+    """The 5-minute demo fixtures must produce ACCEPT (0) and REJECT (1)."""
+    demo = ROOT / "examples" / "demo"
+    acc = subprocess.run(
+        ["vlabs", "clean-gate", "--old", str(demo / "baseline.json"),
+         "--new", str(demo / "candidate.json")], capture_output=True, text=True,
+    )
+    assert acc.returncode == 0, acc.stdout + acc.stderr
+    assert "ACCEPT" in acc.stdout
+    rej = subprocess.run(
+        ["vlabs", "clean-gate", "--old", str(demo / "baseline.json"),
+         "--new", str(demo / "candidate_overfit.json")], capture_output=True, text=True,
+    )
+    assert rej.returncode == 1, rej.stdout + rej.stderr
+    assert "REJECT" in rej.stdout
+    assert "ood_regressed" in rej.stdout and "dcr_increased" in rej.stdout