--- license: other library_name: scikit-learn pipeline_tag: tabular-classification tags: - cybersecurity - gatekeeper - decision-tree - scikit-learn - joblib --- # Agent 3 — Verdict Gatekeeper (Decision Tree) Decision tree that turns two agent scores into a final verdict using a consensus + fail-safe policy. Part of a three-agent DevSecOps auditor (metrics → transformer → gatekeeper). ## Model details | Field | Value | |---|---| | Version | 1.0.0 (initial release) | | Artifact | `model.joblib` (bundle: `model`, `feature_cols`, `classes`, `policy`, `source_thresholds`) | | Task | 3-class tabular classification | | Classes | `FAIL`, `PASS`, `REVIEW` | | Own threshold | None — uses the upstream agent thresholds | | Upstream thresholds | Agent 1 **0.3183**, Agent 2 **0.45** | | Training data | Synthetic policy grid, seed 42 | | Dataset version | Generator v1 — 45 ratios × 45 ratios × 4 quadrants = 8,100 rows | `metadata.json` mirrors the policy, features, grid, and thresholds. ## Input features Four numeric features, **in this exact order**: ``` 1. a1_high_risk # 1.0 if agent1_score >= 0.3183 2. a2_vulnerable # 1.0 if agent2_score >= 0.45 3. a1_risk_ratio # agent1_score / 0.3183 4. a2_risk_ratio # agent2_score / 0.45 ``` ## Preprocessing - Compute the two flags with the upstream thresholds above. - Compute the two ratios as `score / threshold` (1.0 = exactly at the boundary). - Keep the column order shown above; it must match `b["feature_cols"]`. - On any error, missing value, or invalid score, output `REVIEW` directly. ## Output Raw: `predict_proba(x)` returns probabilities for `["FAIL", "PASS", "REVIEW"]` (the bundle's `classes` order). Take `argmax` for the verdict and `max` for confidence. Reference wrapper output: ```json {"verdict": "FAIL", "confidence": 1.0, "status": "ok", "reason": "both agents flagged risk"} ``` `confidence` is the tree's class probability, **not** calibrated. ## Policy | Agent 1 | Agent 2 | Verdict | |---|---|---| | low risk | not vulnerable | `PASS` | | high risk | vulnerable | `FAIL` | | high risk | not vulnerable | `REVIEW` | | low risk | vulnerable | `REVIEW` | | any error / missing / invalid input | | `REVIEW` | The tree is trained on a synthetic grid labelled by this policy, so it encodes the policy rather than learning from ground truth. The reference wrapper adds a post-check that forces `REVIEW` on any tree/policy mismatch — keep it if you deploy this model. Note: the saved `source_thresholds` fallback still lists Agent 2 as `0.51` from an earlier run. Pass the current thresholds explicitly (`0.3183` / `0.45`). ## Usage ```python import joblib b = joblib.load("model.joblib") a1_high = a1_score >= 0.3183 a2_high = a2_score >= 0.45 x = [[float(a1_high), float(a2_high), a1_score / 0.3183, a2_score / 0.45]] proba = b["model"].predict_proba(x)[0] verdict = b["classes"][int(proba.argmax())] confidence = float(proba.max()) ``` ## License Not specified. Update the `license` field before publishing.