Update README.md

README.md

@@ -28,6 +28,18 @@ Trajectory Dataset** product. It contains roughly **~4% of the full dataset**
 at identical schema, attacker-tier distribution, and statistical fingerprint,
 so you can evaluate fit before licensing the full product.
 
+> 🤖 **Trained baseline + leakage diagnostic available:**
+> [**xpertsystems/cyb011-baseline-classifier**](https://huggingface.co/xpertsystems/cyb011-baseline-classifier)
+> — XGBoost + PyTorch MLP for **7-class adversarial attack phase
+> classification** (the dataset's headline target), group-aware split
+> by `campaign_id`, multi-seed evaluation (acc 0.867 ± 0.010, ROC-AUC
+> 0.977 ± 0.002). **Includes a comprehensive `leakage_diagnostic.json`**
+> documenting 6 oracle paths discovered across the dataset's targets,
+> 4 README-suggested headline targets that are unlearnable on the
+> sample after honest leak removal, and the missing `nation_state`
+> attacker tier. Buyers planning adversarial ML research should read
+> the diagnostic first.
+
 | File                          | Rows (sample) | Rows (full)   | Description                                  |
 |-------------------------------|---------------|---------------|----------------------------------------------|
 | `network_topology.csv`        | ~200           | ~2,800        | Network segment / defender registry          |
@@ -39,7 +51,7 @@ so you can evaluate fit before licensing the full product.
 
 CYB011 simulates end-to-end **adversarial AI evasion attack campaigns**
 against ML-based security detection systems, modeled as a **6-phase
-adversarial state machine**:
+adversarial state machine** (data has 7 phases — adds `idle_dwell`):
 
 - **6 adversarial phases**: reconnaissance → feature_space_probe →
   perturbation_craft → evasion_attempt → feedback_adaptation →
@@ -65,6 +77,58 @@ adversarial state machine**:
 - **MLOps security signals** — gradient access patterns, feature-space
   probing, lateral pivoting between models
 
+## Trained Baseline + Leakage Audit Available
+
+A working baseline classifier + comprehensive leakage diagnostic is
+published at
+**[xpertsystems/cyb011-baseline-classifier](https://huggingface.co/xpertsystems/cyb011-baseline-classifier)**.
+
+| Component | Detail |
+|---|---|
+| Primary task | **7-class `attack_phase` classification** (the dataset's headline target) |
+| Secondary artifact | **`leakage_diagnostic.json`** — 6 oracle paths + 4 unlearnable targets + missing tier note |
+| Models | XGBoost (`model_xgb.json`) + PyTorch MLP (`model_mlp.safetensors`) |
+| Features | 37 (after one-hot encoding); pipeline included as `feature_engineering.py` |
+| Split | **Group-aware** (GroupShuffleSplit on `campaign_id`) — 200 campaigns, ~30 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 0.867 ± 0.010, macro ROC-AUC 0.977 ± 0.002 (multi-seed) |
+
+**Important findings for buyers planning CYB011 ML work** (full detail
+in
+[`leakage_diagnostic.json`](https://huggingface.co/xpertsystems/cyb011-baseline-classifier/blob/main/leakage_diagnostic.json)):
+
+**Missing `nation_state` attacker tier:** README lists 4 tiers; sample
+contains only 3 (script_kiddie 50%, opportunistic 40%, APT 10%).
+Nation_state events are entirely absent. Models trained on this
+sample cannot generalize to nation_state actors.
+
+**6 oracle paths documented** across the dataset's targets:
+
+**Phase target oracles (3 paths — must be dropped):**
+1. `detection_outcome` (`!= suppressed_alert` → 100% `evasion_attempt`)
+2. `detector_confidence_score` (threshold-derived from `detection_outcome`)
+3. `evasion_budget_consumed` (`== 0` → 100% one of 3 early phases)
+
+**Other documented leaks (for transparency):**
+4. `stealth_score` near-deterministic per `attacker_capability_tier` (inflates per-campaign tier prediction from honest ~0.50 to 0.94)
+5. Topology fingerprint — 7 segment-level features uniquely identify each `defender_architecture` (makes 8-class defender prediction trivially 100%, collapses to 0.13 when fingerprint dropped)
+6. `timestep` partial oracle for 3 phases — **KEPT in the published model as legitimate campaign-progress observable**
+
+**7 phases in data, README claims 6:** The data adds `idle_dwell` as
+a phase (17.5% of events). The published baseline trains on all 7.
+
+**4 README-suggested headline targets unlearnable after honest leak
+removal:**
+- `campaign_success_flag` binary (acc 0.51 vs majority 0.61)
+- `campaign_type` 8-class (acc 0.11 vs majority 0.17)
+- `coordinated_attack_flag` binary (acc 0.83 vs majority 0.90, only 20 positives)
+- `defender_architecture` 8-class (collapses when topology fingerprint dropped)
+
+**Only viable headline target:** `attack_phase` 7-class — acc 0.867,
+ROC-AUC 0.977 with group-aware split. All 7 classes earn nonzero F1
+(range 0.49-1.00).
+
 ## Calibrated Benchmark Targets
 
 The full product is calibrated to 12 benchmark validation tests drawn from
@@ -146,10 +210,16 @@ log and segment/defender registry schemas respectively.
 
 - Training **adversarial example detectors** — distinguish clean vs
   perturbed inputs from feature-space telemetry
+- **Attack phase classification** (the baseline ships this) — predict
+  the 7-phase position of a trajectory event —
+  [worked example available](https://huggingface.co/xpertsystems/cyb011-baseline-classifier)
 - **Attacker tier attribution** — 4-class classification of evasion
-  campaigns by capability tier
+  campaigns by capability tier (see leakage diagnostic — nation_state
+  tier MISSING from sample; per-campaign prediction inflated by
+  stealth_score leakage)
 - **Defender architecture vulnerability assessment** — predict which
-  defender architectures are most evadable
+  defender architectures are most evadable (see leakage diagnostic —
+  trivially leaky via topology fingerprint; unlearnable when dropped)
 - **L∞ / L2 perturbation budget detection** — calibrate ε-thresholds
 - **Query budget exhaustion attacks** — model black-box query patterns
 - **Concept drift poisoning detection** — distinguish natural drift
@@ -173,24 +243,29 @@ summaries    = pd.read_csv("campaign_summary.csv")
 events       = pd.read_csv("campaign_events.csv")
 topology     = pd.read_csv("network_topology.csv")
 
-# Join trajectory data with campaign-level labels
-enriched = trajectories.merge(summaries, on=["campaign_id", "attacker_id"],
-                              how="left", suffixes=("", "_summary"))
-enriched = enriched.merge(topology, on="segment_id", how="left")
+# Join trajectory data with topology (segment-level features)
+enriched = trajectories.merge(topology, left_on="target_segment_id",
+                              right_on="segment_id", how="left")
 
-# 4-class attacker tier target
-y_tier = summaries["attacker_tier"]
+# 7-class attack_phase target (the baseline ships this)
+y_phase = trajectories["attack_phase"]
 
-# Binary evasion success target
-y_evasion = summaries["evasion_success_flag"]
+# Multi-class attacker tier (3 values in sample; see leakage diagnostic)
+y_tier = trajectories["attacker_capability_tier"]
 
-# Multi-class defender architecture target
-y_defender = topology["defender_architecture"]
+# Binary evasion success target (see leakage diagnostic — unlearnable)
+y_evasion = summaries["campaign_success_flag"]
 
-# Binary concept drift / poisoning detection
-y_poisoned = summaries["concept_drift_injected_flag"]
+# Multi-class defender architecture target (see leakage diagnostic —
+# trivially leaky via topology fingerprint)
+y_defender = topology["defender_architecture"]
 ```
 
+For a worked end-to-end example with `attack_phase` 7-class
+classification, group-aware splitting, feature engineering, and the
+full 6-oracle-path leakage audit, see the
+[baseline classifier repo](https://huggingface.co/xpertsystems/cyb011-baseline-classifier).
+
 ## License
 
 This **sample** is released under **CC-BY-NC-4.0** (free for non-commercial