Update README.md

README.md

@@ -28,6 +28,16 @@ Dataset** product. It contains roughly **~10% of the full dataset** at
 identical schema, MITRE ATT&CK tactic coverage, and statistical fingerprint,
 so you can evaluate fit before licensing the full product.
 
+> 🤖 **Trained baseline available:**
+> [**xpertsystems/cyb008-baseline-classifier**](https://huggingface.co/xpertsystems/cyb008-baseline-classifier)
+> — XGBoost + PyTorch MLP for **5-class SOC alert triage outcome
+> classification** (the README's first headline use case), stratified
+> split, multi-seed evaluation (ROC-AUC 0.955 ± 0.003). **Includes a
+> structural-leakage diagnostic** documenting three oracle columns
+> dropped from the feature set, and a separate unlearnable-target
+> finding for MITRE ATT&CK tactic classification. Buyers planning
+> SOC ML work should read the diagnostic first.
+
 | File                       | Rows (sample) | Rows (full)   | Description                                  |
 |----------------------------|---------------|---------------|----------------------------------------------|
 | `soc_topology.csv`         | ~25           | ~2,400        | SOC / analyst registry                       |
@@ -61,6 +71,45 @@ lifecycles across enterprise detection environments, with:
 - **Rule drift modeling** — periodic rule-noise amplification simulating
   detection-engineering signal decay
 
+## Trained Baseline Available
+
+A working baseline classifier trained on this sample is published at
+**[xpertsystems/cyb008-baseline-classifier](https://huggingface.co/xpertsystems/cyb008-baseline-classifier)**.
+
+| Component | Detail |
+|---|---|
+| Primary task | **5-class `resolution_outcome` classification** (SOC alert triage — the README's first headline use case) |
+| Diagnostic | Structural-leakage audit (3 oracle columns dropped) + unlearnable-target finding for `mitre_tactic` |
+| Models | XGBoost (`model_xgb.json`) + PyTorch MLP (`model_mlp.safetensors`) |
+| Features | 53 (after one-hot encoding); pipeline included as `feature_engineering.py` |
+| Split | **Stratified random** — no natural row-level group key (25 analysts, 5 SOCs, only 9% of alerts link to an incident) |
+| Validation | Single seed + multi-seed aggregate across 10 seeds |
+| Demo | `inference_example.ipynb` — end-to-end copy-paste |
+| Headline metrics | XGBoost: accuracy 0.777 ± 0.007, macro ROC-AUC 0.955 ± 0.003 (multi-seed); MLP slightly outperforms |
+
+**Important findings for buyers planning SOC ML work:**
+
+1. **Three structural oracles in the data** (`alert_lifecycle_phase`,
+   `automation_resolved`, `escalation_flag`) deterministically encode
+   the `resolution_outcome` label. With these columns present, a
+   plain XGBoost achieves 100% accuracy. The baseline excludes them
+   to demonstrate honest learning — and the documented honest result
+   (acc 0.78, AUC 0.96) is genuinely useful.
+
+2. **MITRE ATT&CK tactic classification is NOT learnable on this
+   sample.** The README lists tactic classification as a top use case,
+   but feature distributions are nearly identical across all 12
+   tactics. A trained model performs *below* majority baseline
+   (acc 0.08 vs 0.14). The baseline model card documents this
+   explicitly with a recommendation to the dataset author.
+
+3. **SLA breach prediction is also not learnable** (acc 0.68 vs
+   majority 0.82). Documented as out-of-scope.
+
+See the model card and `leakage_diagnostic.json` for the full audit
+and our recommendations to make these tasks viable in the next
+dataset version.
+
 ## Calibrated Benchmark Targets
 
 The full product is calibrated to 12 benchmark validation tests drawn from
@@ -136,8 +185,10 @@ and SOC registry schemas respectively.
 
 ## Suggested Use Cases
 
-- Training **alert triage** models — predict true_positive vs false_positive
-- **MITRE ATT&CK tactic classification** from alert features
+- Training **alert triage** models — predict TP vs FP, or full 5-class
+  resolution outcome (the baseline ships this) —
+  [worked example available](https://huggingface.co/xpertsystems/cyb008-baseline-classifier)
+- **MITRE ATT&CK tactic classification** from alert features (see baseline diagnostic — not learnable on this sample)
 - **SOAR playbook recommendation** — predict which alerts benefit from
   automation
 - **Alert prioritization** — calibrate triage scores against ground-truth
@@ -146,7 +197,7 @@ and SOC registry schemas respectively.
   workload
 - **Kill-chain detection** — group related alerts into multi-stage
   incidents
-- **SLA breach prediction** — early-warning systems for at-risk alerts
+- **SLA breach prediction** — early-warning systems (see baseline diagnostic — not learnable on this sample)
 - **Alert storm detection** — distinguish coordinated bursts from baseline
   volume
 - **False positive reduction** modeling — reduce 45% FP rate
@@ -166,23 +217,25 @@ topology  = pd.read_csv("soc_topology.csv")
 enriched = alerts.merge(topology, on=["soc_id", "analyst_id"], how="left",
                         suffixes=("", "_analyst"))
 
-# Binary true-positive classification target
+# 5-class triage outcome target (the README's first headline use case)
+y_outcome = alerts["resolution_outcome"]
+
+# Binary true-positive collapse (for binary triage)
 y_tp = alerts["resolution_outcome"].isin([
-    "true_positive_remediated",
-    "true_positive_escalated",
-    "incident_declared",
+    "true_positive_remediated", "true_positive_escalated",
 ]).astype(int)
 
-# Multi-class ATT&CK tactic classification target
+# Multi-class ATT&CK tactic classification target — see leakage diagnostic
 y_tactic = alerts["mitre_tactic"]
 
-# Binary SLA breach prediction target
+# Binary SLA breach prediction target — see leakage diagnostic
 y_sla = alerts["sla_breached_flag"]
-
-# Per-incident severity classification
-y_severity = incidents["incident_severity"]
 ```
 
+For a worked end-to-end example with 5-class triage classification,
+stratified splitting, and feature engineering, see the inference notebook
+in the [baseline classifier repo](https://huggingface.co/xpertsystems/cyb008-baseline-classifier/blob/main/inference_example.ipynb).
+
 ## License
 
 This **sample** is released under **CC-BY-NC-4.0** (free for non-commercial