Update README.md

README.md

@@ -28,6 +28,18 @@ Log Dataset** product. It contains roughly **~10% of the full dataset** at
 identical schema, MITRE ATT&CK technique coverage, and statistical
 fingerprint, so you can evaluate fit before licensing the full product.
 
+> 🤖 **Trained baseline + leakage diagnostic available:**
+> [**xpertsystems/cyb010-baseline-classifier**](https://huggingface.co/xpertsystems/cyb010-baseline-classifier)
+> — XGBoost + PyTorch MLP for **5-class attack lifecycle phase
+> classification** (the dataset's headline target), group-aware split
+> by `incident_id`, multi-seed evaluation (acc 0.936 ± 0.007, ROC-AUC
+> 0.988 ± 0.001 — tightest AUC std in the catalog). **Includes a
+> comprehensive `leakage_diagnostic.json`** documenting 11 oracle
+> paths discovered across the dataset's targets and 2 README-suggested
+> headline targets that are unlearnable on the sample after honest
+> leak removal. Buyers planning SIEM ML work should read the
+> diagnostic first.
+
 | File                       | Rows (sample) | Rows (full)   | Description                                  |
 |----------------------------|---------------|---------------|----------------------------------------------|
 | `host_inventory.csv`       | ~400           | ~3,200        | Enterprise host inventory                    |
@@ -60,6 +72,60 @@ lifecycle state machine** across realistic detection environments, with:
 - **IOC seeding density** — calibrated indicator-of-compromise injection
   for threat intel detection benchmarking
 
+## Trained Baseline + Leakage Audit Available
+
+A working baseline classifier + comprehensive leakage diagnostic is
+published at
+**[xpertsystems/cyb010-baseline-classifier](https://huggingface.co/xpertsystems/cyb010-baseline-classifier)**.
+
+| Component | Detail |
+|---|---|
+| Primary task | **5-class `attack_lifecycle_phase` classification** (the dataset's headline target) |
+| Secondary artifact | **`leakage_diagnostic.json`** — 11 oracle paths + 2 unlearnable targets |
+| Models | XGBoost (`model_xgb.json`) + PyTorch MLP (`model_mlp.safetensors`) |
+| Features | 87 (after one-hot encoding); pipeline included as `feature_engineering.py` |
+| Split | **Group-aware** (GroupShuffleSplit on `incident_id`) — 500 incidents, ~75 in test fold |
+| Validation | Single seed + multi-seed aggregate across 10 seeds |
+| Demo | `inference_example.ipynb` — end-to-end copy-paste |
+| Headline metrics | XGBoost: accuracy 0.936 ± 0.007, macro ROC-AUC 0.988 ± 0.001 (multi-seed) |
+
+**Important findings for buyers planning CYB010 ML work** (full detail
+in
+[`leakage_diagnostic.json`](https://huggingface.co/xpertsystems/cyb010-baseline-classifier/blob/main/leakage_diagnostic.json)):
+
+**11 oracle paths documented across two task families:**
+
+**Phase target oracles (6 paths)** — drop these when training your own
+phase classifier:
+
+1. `mitre_tactic == "benign"` → 100% `benign_background` phase
+2. `mitre_technique_id` → `mitre_tactic` (perfect ATT&CK-by-design oracle)
+3. `label_malicious == False` → 100% `benign_background`
+4. `threat_actor_id == "NONE"` → 100% benign
+5. `threat_actor_profile == "benign_user"` → 100% benign
+6. `event_type` (many values phase-specific; e.g. `c2_beacon_outbound` → 100% exfil)
+
+**Alert TP target oracles (7 paths)** — `label_true_positive` on
+`alert_records.csv` is 100% accurate with any single one of these
+intact:
+
+1. `alert_category == "false_positive_noise"` → 100% FP
+2. `label_false_positive` (mirror target)
+3. `time_to_detect_seconds == 0` → 100% FP (sentinel)
+4. `correlated_chain_length == 1` → near-100% FP (sentinel)
+5. `analyst_triage_priority ∈ {P1,P2,P3}` → 100% TP
+6. `suppression_reason == NaN` → 100% TP
+7. `alert_rule_name` (rule names encode answer)
+
+**2 README-suggested headline targets unlearnable after honest leak
+removal:**
+- `threat_actor_profile` 4-class malicious-only (acc 0.55 vs majority 0.61)
+- `event_class` 12-class (acc 0.35 vs majority 0.42)
+
+**Viable secondary task:** `label_true_positive` binary on alerts —
+acc 0.80, AUC 0.89 after dropping all 7 oracle columns. Documented in
+the diagnostic.
+
 ## Calibrated Benchmark Targets
 
 The full product is calibrated to 6 benchmark validation tests drawn from
@@ -143,9 +209,16 @@ Sample benchmark results:
 
 ## Suggested Use Cases
 
-- Training **SIEM alert triage** models — predict true_positive vs false_positive
+- Training **attack lifecycle phase classification** models (the
+  baseline ships this) —
+  [worked example available](https://huggingface.co/xpertsystems/cyb010-baseline-classifier)
+- Training **SIEM alert triage** models — predict true_positive vs
+  false_positive (see leakage diagnostic — 7 oracle columns must be
+  dropped; honest acc 0.80 / AUC 0.89)
 - **MITRE ATT&CK technique classification** from raw log lines
 - **Threat actor attribution** — 5-class with realistic class imbalance
+  (see leakage diagnostic — 4-class malicious-only is unlearnable;
+  5-class works only because benign separation is trivial)
 - **Multi-format log parser training** — 8 SIEM vendor formats in one corpus
 - **Dwell time forecasting** under varying defender posture
 - **Lateral movement detection** from event sequences
@@ -169,21 +242,31 @@ enriched = events.merge(hosts, on="host_id", how="left",
                         suffixes=("", "_host"))
 
 # Join alerts back to source event and incident
-alerts_full = alerts.merge(events, left_on="triggering_event_id",
+alerts_full = alerts.merge(events, left_on="correlated_event_ids",
                            right_on="event_id", how="left",
                            suffixes=("_alert", "_event"))
 
-# Multi-class threat actor profile target
+# 5-class attack lifecycle phase target (the baseline ships this)
+y_phase = events["attack_lifecycle_phase"]
+
+# Multi-class threat actor profile target (5-class with benign;
+# see leakage diagnostic — 4-class malicious-only is unlearnable)
 y_actor = events["threat_actor_profile"]
 
 # Binary false-positive prediction target
+# (see leakage diagnostic — 7 oracle columns must be dropped)
 y_fp = alerts["label_false_positive"]
 
 # Multi-class MITRE technique target (filter to malicious events)
-malicious = events[events["mitre_technique_id"] != ""]
+malicious = events[events["label_malicious"] == True]
 y_technique = malicious["mitre_technique_id"]
 ```
 
+For a worked end-to-end example with `attack_lifecycle_phase` 5-class
+classification, group-aware splitting, feature engineering, and the
+full 11-oracle-path leakage audit, see the
+[baseline classifier repo](https://huggingface.co/xpertsystems/cyb010-baseline-classifier).
+
 ## License
 
 This **sample** is released under **CC-BY-NC-4.0** (free for non-commercial