MhamedRayan commited on
Commit
ad80727
·
verified ·
1 Parent(s): 8d5e810

Upload folder using huggingface_hub

Browse files
Files changed (3) hide show
  1. README.md +79 -0
  2. metadata.json +29 -0
  3. model.joblib +3 -0
README.md ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ library_name: scikit-learn
4
+ pipeline_tag: tabular-classification
5
+ tags:
6
+ - cybersecurity
7
+ - gatekeeper
8
+ - decision-tree
9
+ - scikit-learn
10
+ - joblib
11
+ ---
12
+
13
+ # Agent 3 — Verdict Gatekeeper (Decision Tree)
14
+
15
+ Decision tree that turns two agent scores into a final verdict using a
16
+ consensus + fail-safe policy. Part of a three-agent DevSecOps auditor
17
+ (metrics → transformer → gatekeeper).
18
+
19
+ > For a Hugging Face repo, upload this file as `README.md` alongside
20
+ > `agent3_dt.joblib` and `agent3_dt.meta.json`.
21
+
22
+ ## Files
23
+
24
+ | File | Contents |
25
+ |---|---|
26
+ | `agent3_dt.joblib` | `model`, `feature_cols`, `classes`, `policy`, `source_thresholds` |
27
+ | `agent3_dt.meta.json` | Policy, features, grid size, agreement, thresholds |
28
+
29
+ ## Input
30
+
31
+ Four features per sample:
32
+
33
+ ```
34
+ a1_high_risk # 1.0 if agent1_score >= agent1_threshold
35
+ a2_vulnerable # 1.0 if agent2_score >= agent2_threshold
36
+ a1_risk_ratio # agent1_score / agent1_threshold
37
+ a2_risk_ratio # agent2_score / agent2_threshold
38
+ ```
39
+
40
+ ## Output
41
+
42
+ One of `PASS`, `REVIEW`, `FAIL`, with the tree's class probability as
43
+ `confidence`. That confidence is **not** calibrated.
44
+
45
+ ## Policy
46
+
47
+ | Agent 1 | Agent 2 | Verdict |
48
+ |---|---|---|
49
+ | low risk | not vulnerable | `PASS` |
50
+ | high risk | vulnerable | `FAIL` |
51
+ | high risk | not vulnerable | `REVIEW` |
52
+ | low risk | vulnerable | `REVIEW` |
53
+ | any error / missing / invalid input | | `REVIEW` |
54
+
55
+ The tree is trained on a synthetic grid labelled by this policy, so it encodes
56
+ the policy rather than learning from ground truth. The reference wrapper adds a
57
+ post-check that forces `REVIEW` on any tree/policy mismatch — keep it if you
58
+ deploy this model.
59
+
60
+ ## Usage
61
+
62
+ ```python
63
+ import joblib
64
+
65
+ b = joblib.load("agent3_dt.joblib")
66
+
67
+ a1_high = a1_score >= a1_threshold
68
+ a2_high = a2_score >= a2_threshold
69
+ x = [[float(a1_high), float(a2_high),
70
+ a1_score / a1_threshold, a2_score / a2_threshold]]
71
+
72
+ proba = b["model"].predict_proba(x)[0]
73
+ verdict = b["classes"][int(proba.argmax())]
74
+ confidence = float(proba.max())
75
+ ```
76
+
77
+ ## License
78
+
79
+ Not specified. Update the `license` field before publishing.
metadata.json ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "policy": "consensus_fail_safe",
3
+ "features": [
4
+ "a1_high_risk",
5
+ "a2_vulnerable",
6
+ "a1_risk_ratio",
7
+ "a2_risk_ratio"
8
+ ],
9
+ "classes": [
10
+ "FAIL",
11
+ "PASS",
12
+ "REVIEW"
13
+ ],
14
+ "max_depth": 3,
15
+ "seed": 42,
16
+ "grid": {
17
+ "ratio_min": 0.0,
18
+ "ratio_max": 2.0,
19
+ "ratios": 45,
20
+ "rows": 8100
21
+ },
22
+ "policy_agreement": 1.0,
23
+ "train_accuracy": 1.0,
24
+ "source_thresholds": {
25
+ "agent1": 0.3182926377275692,
26
+ "agent2": 0.5099999999999998
27
+ },
28
+ "note": "Synthetic policy grid; the tree encodes the consensus policy, not ground-truth verdicts."
29
+ }
model.joblib ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f5508cbc56b41660f0fbc2e6eea7f9caf9612ae511084b956fd2c17eb179b083
3
+ size 2244