Update README.md

README.md

@@ -16,7 +16,7 @@ tags:
   - zero-day
 pretty_name: CYB009 — Synthetic Vulnerability Intelligence (Sample)
 size_categories:
-  - 10K<n<100K
+  - 100K<n<1M
 ---
 
 # CYB009 — Synthetic Vulnerability Intelligence Dataset (Sample)
@@ -29,6 +29,27 @@ dataset rows** (but generated from ~40% the org/asset count) at identical
 schema, CVSS distribution, and statistical fingerprint, so you can
 evaluate fit before licensing the full product.
 
+> 🤖 **Trained baseline + comprehensive leakage audit available:**
+> [**xpertsystems/cyb009-baseline-classifier**](https://huggingface.co/xpertsystems/cyb009-baseline-classifier)
+> — XGBoost + PyTorch MLP for **8-class vulnerability classification**
+> (acc 0.244 ± 0.023, ROC-AUC 0.687 ± 0.014). **The primary artifact
+> is `leakage_diagnostic.json`** — the XpertSystems catalog's most
+> comprehensive structural-leakage audit, documenting 8 oracle paths
+> and 6 README-suggested headline targets that are unlearnable on the
+> sample after honest leak removal. Buyers planning CYB009 ML work
+> should read the diagnostic first.
+
+> ⚠️ **Important: most README-suggested ML targets are not viable on
+> this sample.** The baseline's leakage diagnostic documents that
+> `exploit_maturity_final`, `exploitation_occurred_flag`,
+> `zero_day_flag`, `cisa_kev_flag`,
+> `supply_chain_propagation_flag`, `false_positive_flag`, and the
+> per-timestep `lifecycle_phase` / `patch_status` / `remediation_status`
+> targets all have structural label-feature determinism that makes
+> them either trivially solvable via oracle features or unlearnable
+> after honest leak removal. The dataset is still useful for
+> evaluation, but ML training requires careful target selection.
+
 *Note: This sample is larger than other CYB SKU samples (~45 MB total).
 CYB009 has subset-conditional benchmarks (CISA KEV listing rate, supply
 chain propagation) that need a reasonable vulnerability population to
@@ -68,6 +89,50 @@ CISA KEV modeling, covering:
 - **Compensating controls and risk acceptance** outcomes
 - **Internet-exposed asset modeling** — 38% exposure baseline
 
+## Trained Baseline + Leakage Audit Available
+
+A working baseline classifier + comprehensive leakage diagnostic is
+published at
+**[xpertsystems/cyb009-baseline-classifier](https://huggingface.co/xpertsystems/cyb009-baseline-classifier)**.
+
+| Component | Detail |
+|---|---|
+| **Primary artifact** | **`leakage_diagnostic.json`** — 8 oracle paths + 6 unlearnable targets documented |
+| Secondary artifact | 8-class `vulnerability_class` baseline (XGBoost + PyTorch MLP) |
+| Models | `model_xgb.json` + `model_mlp.safetensors` |
+| Features | 57 (after one-hot encoding); pipeline included as `feature_engineering.py` |
+| Split | Stratified random (per-vulnerability) |
+| Validation | Single seed + multi-seed aggregate across 10 seeds |
+| Demo | `inference_example.ipynb` — end-to-end copy-paste |
+| Headline metrics | XGBoost: accuracy 0.244 ± 0.023, macro ROC-AUC 0.687 ± 0.014 (multi-seed) — the catalog's weakest baseline by design |
+
+**Findings for buyers planning CYB009 ML work** (full detail in
+[`leakage_diagnostic.json`](https://huggingface.co/xpertsystems/cyb009-baseline-classifier/blob/main/leakage_diagnostic.json)):
+
+**8 oracle paths discovered on the sample:**
+
+1. `cvss_temporal_score_final / cvss_base_score` ratio is near-deterministic per `exploit_maturity_final` tier (CVSS v3.1 multipliers 0.91/0.94/0.97/1.00)
+2. `time_to_exploit_days` (-1 sentinel) is a perfect oracle for `exploitation_occurred_flag`
+3. `time_to_remediate_days` (120 sentinel) is a perfect oracle for `remediation_success_flag` and `sla_compliance_flag`
+4. `severity_class` is a 100% mechanical function of `cvss_base_score` (CVSS v3.1 boundaries)
+5. Five `lifecycle_phase` values pin `remediation_status` deterministically (`residual_risk_review` → 100% `remediated`, etc.)
+6. `patch_status = deployed` → 100% `remediated`; four other values → 99% `in_remediation`
+7. `risk_score_composite` is computed from flag fields (indirect oracle)
+8. `patch_lag_days` is suspected to have similar sentinel structure (precaution)
+
+**6 README-suggested headline targets unlearnable after honest leak removal:**
+- `exploit_maturity_final` 4-class (acc 0.31 vs majority 0.36)
+- `exploitation_occurred_flag` binary (acc 0.86 vs majority 0.92)
+- `zero_day_flag` binary (acc 0.95 vs majority 0.97)
+- `cisa_kev_flag` binary (only 14 positives in sample)
+- `supply_chain_propagation_flag` binary (only 20 positives)
+- `false_positive_flag` binary (acc 0.87 vs majority 0.92)
+
+**Only viable headline target:** `vulnerability_class` 8-class — acc
+0.244, ROC-AUC 0.687 vs majority 0.176. The catalog's weakest baseline,
+shipped as a reference and as proof that vulnerability_class is the
+only README-suggested target that learns honestly on the sample.
+
 ## Calibrated Benchmark Targets
 
 The full product is calibrated to 12 benchmark validation tests drawn from
@@ -150,21 +215,23 @@ event log and asset registry schemas respectively.
 
 ## Suggested Use Cases
 
+- Training **vulnerability classification** models (the baseline ships this) —
+  [worked example available](https://huggingface.co/xpertsystems/cyb009-baseline-classifier)
 - Training **vulnerability triage** models — predict CVSS/EPSS-prioritized
   remediation order
 - **Zero-day prediction** — feature engineering from pre-disclosure
-  telemetry
-- **CISA KEV listing prediction** — early-warning for emergency patching
-- **Supply chain compromise detection** — SBOM signal modeling
+  telemetry (see leakage diagnostic — unlearnable on the sample)
+- **CISA KEV listing prediction** — early-warning for emergency patching (see leakage diagnostic — too few positives in the sample)
+- **Supply chain compromise detection** — SBOM signal modeling (see leakage diagnostic — too few positives in the sample)
 - **Patch deployment ETA forecasting** — per-class patch development
   duration prediction
-- **SLA breach prediction** — early-warning for at-risk vulnerabilities
+- **SLA breach prediction** — early-warning for at-risk vulnerabilities (see leakage diagnostic — unlearnable on the sample)
 - **Asset criticality classification** from inventory features
-- **EPSS calibration validation** — empirical vs predicted exploitation
+- **EPSS calibration validation** — empirical vs predicted exploitation (see leakage diagnostic — exploit_maturity_final structurally encoded)
 - **Compensating control effectiveness** modeling
 - **Risk acceptance decision** modeling — predict which vulns get
   accepted vs remediated
-- **Lifecycle phase transition prediction** — multi-class sequence modeling
+- **Lifecycle phase transition prediction** — multi-class sequence modeling (see leakage diagnostic — state-machine determinism)
 
 ## Loading the Data
 
@@ -181,19 +248,24 @@ enriched = records.merge(vulns, on=["vuln_id", "asset_id"], how="left",
                          suffixes=("", "_summary"))
 enriched = enriched.merge(assets, on="asset_id", how="left")
 
-# Binary exploitation-in-wild target
-y_exploited = vulns["exploited_in_wild"]
+# 8-class vulnerability classification target (the baseline ships this)
+y_class = vulns["vulnerability_class"]
 
-# Binary CISA KEV listing target (rare event ~6.5%)
-y_kev = vulns["cisa_kev_listed"]
+# Binary exploitation-in-wild target (see leakage diagnostic — unlearnable on sample)
+y_exploited = vulns["exploitation_occurred_flag"]
 
-# Multi-class vulnerability classification
-y_class = vulns["vuln_class"]
+# Binary CISA KEV listing target (rare event — only 14 positives in sample)
+y_kev = vulns["cisa_kev_flag"]
 
-# Binary SLA breach prediction
-y_sla = records["sla_breached_flag"]
+# Binary SLA breach prediction (see leakage diagnostic — unlearnable)
+y_sla = records["sla_compliance_flag"]  # Note: data uses compliance flag (True=compliant), not breach flag
 ```
 
+For a worked end-to-end example with vulnerability_class 8-class
+classification, stratified splitting, feature engineering, and the
+full 8-oracle-path leakage audit, see the
+[baseline classifier repo](https://huggingface.co/xpertsystems/cyb009-baseline-classifier).
+
 ## License
 
 This **sample** is released under **CC-BY-NC-4.0** (free for non-commercial