Update README.md

README.md

@@ -28,6 +28,14 @@ Dataset** product. It contains roughly **~1.3% of the full dataset** at
 identical schema, threat-actor-tier distribution, and statistical fingerprint,
 so you can evaluate fit before licensing the full product.
 
+> 🤖 **Trained baseline available:**
+> [**xpertsystems/cyb006-baseline-classifier**](https://huggingface.co/xpertsystems/cyb006-baseline-classifier)
+> — XGBoost + PyTorch MLP for **3-class user-risk-tier classification**
+> (insider-threat scoring use case), stratified split, multi-seed evaluation
+> (ROC-AUC 0.812 ± 0.048). **Includes a structural-leakage diagnostic on
+> the threat-actor detection task** that buyers planning ATO / threat-actor
+> ML work should read first.
+
 | File                       | Rows (sample) | Rows (full)   | Description                                  |
 |----------------------------|---------------|---------------|----------------------------------------------|
 | `identity_topology.csv`    | ~150           | ~3,200        | Identity domain registry                     |
@@ -58,6 +66,33 @@ machine** across diverse identity infrastructures, with:
 - **Conditional Access (CA) policy enforcement** modeling — ZTNA block
   strength tunable per architecture
 
+## Trained Baseline Available
+
+A working baseline classifier trained on this sample is published at
+**[xpertsystems/cyb006-baseline-classifier](https://huggingface.co/xpertsystems/cyb006-baseline-classifier)**.
+
+| Component | Detail |
+|---|---|
+| Primary task | **3-class user_risk_tier classification** (insider-threat scoring) |
+| Diagnostic | Audit of threat-actor detection on this sample (see `leakage_diagnostic.json`) |
+| Models | XGBoost (`model_xgb.json`) + PyTorch MLP (`model_mlp.safetensors`) |
+| Features | 34 per-user features (aggregates + non-leaky session aggregates + engineered) |
+| Split | **Stratified by user_risk_tier** — user-level task, n=200 |
+| Validation | Single seed + multi-seed aggregate across 10 seeds |
+| Demo | `inference_example.ipynb` — end-to-end copy-paste |
+| Headline metrics | XGBoost: accuracy 0.700 ± 0.082, macro ROC-AUC 0.812 ± 0.048 (multi-seed) |
+
+**Important diagnostic finding for buyers planning threat-actor detection
+work:** the model card documents that this sample's threat-actor-vs-legitimate
+session populations have **non-overlapping anomaly score distributions**
+across at least six feature groups (velocity, timestamp, credential attempt
+count, login outcome, geo country, device trust). As a result, a plain
+XGBoost achieves 100% test accuracy on threat-actor binary detection that
+does not reflect real-world detection difficulty. The baseline model
+targets `user_risk_tier` instead, which is a legitimate ML task on the
+sample. See the model card's [Leakage diagnostic](https://huggingface.co/xpertsystems/cyb006-baseline-classifier#leakage-diagnostic)
+section for the full audit and recommendations.
+
 ## Calibrated Benchmark Targets
 
 The full product is calibrated to **12 benchmark validation tests** drawn
@@ -161,15 +196,16 @@ identity domain schemas respectively.
 
 ## Suggested Use Cases
 
-- Training **account takeover (ATO) detection** models
-- **Threat-actor tier classification** — 4-class with realistic class imbalance
+- Training **insider threat scoring** models —
+  [worked example available](https://huggingface.co/xpertsystems/cyb006-baseline-classifier)
+- **Account takeover (ATO) detection** model development (see leakage diagnostic in the baseline model card before training)
+- **Threat-actor tier classification** — 4-class with realistic class imbalance (see leakage diagnostic before training)
 - **Impossible travel detection** — geo-velocity feature engineering
 - **MFA bypass detection** — distinguish FIDO2 anomalies from push fatigue
 - **Lateral movement detection** — session-graph traversal patterns
 - **Golden Ticket / Pass-the-Hash** detection benchmarking
 - **UEBA precision/recall tuning** with calibrated false-positive baselines
 - **Conditional Access policy effectiveness** simulation
-- **Insider threat scoring** with composite behavioral indicators
 - **Zero Trust posture validation** — ZTNA block rate analysis
 
 ## Loading the Data
@@ -186,7 +222,7 @@ domains  = pd.read_csv("identity_topology.csv")
 enriched = sessions.merge(users, on="user_id", how="left",
                           suffixes=("", "_user"))
 
-# Threat-actor tier classification target (4-class)
+# Threat-actor tier classification target (4-class) — see leakage diagnostic
 y_tier = sessions["threat_actor_capability_tier"]
 
 # Binary account-takeover detection target
@@ -196,6 +232,10 @@ y_ato = users["account_takeover_flag"]
 y_it = sessions["impossible_travel_flag"]
 ```
 
+For a worked end-to-end example with user-risk-tier classification,
+stratified splitting, and feature engineering, see the inference notebook
+in the [baseline classifier repo](https://huggingface.co/xpertsystems/cyb006-baseline-classifier/blob/main/inference_example.ipynb).
+
 ## License
 
 This **sample** is released under **CC-BY-NC-4.0** (free for non-commercial