Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -16,31 +16,52 @@ 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 |
-
|
| 20 |
-
> `agent3_dt.joblib` and `agent3_dt.meta.json`.
|
| 21 |
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
| File | Contents |
|
| 25 |
|---|---|
|
| 26 |
-
|
|
| 27 |
-
| `
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
-
## Input
|
| 30 |
|
| 31 |
-
Four features
|
| 32 |
|
| 33 |
```
|
| 34 |
-
a1_high_risk
|
| 35 |
-
a2_vulnerable
|
| 36 |
-
a1_risk_ratio
|
| 37 |
-
a2_risk_ratio
|
| 38 |
```
|
| 39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
## Output
|
| 41 |
|
| 42 |
-
|
| 43 |
-
`
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
## Policy
|
| 46 |
|
|
@@ -57,17 +78,20 @@ 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("
|
| 66 |
|
| 67 |
-
a1_high = a1_score >=
|
| 68 |
-
a2_high = a2_score >=
|
| 69 |
x = [[float(a1_high), float(a2_high),
|
| 70 |
-
a1_score /
|
| 71 |
|
| 72 |
proba = b["model"].predict_proba(x)[0]
|
| 73 |
verdict = b["classes"][int(proba.argmax())]
|
|
|
|
| 16 |
consensus + fail-safe policy. Part of a three-agent DevSecOps auditor
|
| 17 |
(metrics → transformer → gatekeeper).
|
| 18 |
|
| 19 |
+
## Model details
|
|
|
|
| 20 |
|
| 21 |
+
| Field | Value |
|
|
|
|
|
|
|
| 22 |
|---|---|
|
| 23 |
+
| Version | 1.0.0 (initial release) |
|
| 24 |
+
| Artifact | `model.joblib` (bundle: `model`, `feature_cols`, `classes`, `policy`, `source_thresholds`) |
|
| 25 |
+
| Task | 3-class tabular classification |
|
| 26 |
+
| Classes | `FAIL`, `PASS`, `REVIEW` |
|
| 27 |
+
| Own threshold | None — uses the upstream agent thresholds |
|
| 28 |
+
| Upstream thresholds | Agent 1 **0.3183**, Agent 2 **0.45** |
|
| 29 |
+
| Training data | Synthetic policy grid, seed 42 |
|
| 30 |
+
| Dataset version | Generator v1 — 45 ratios × 45 ratios × 4 quadrants = 8,100 rows |
|
| 31 |
+
|
| 32 |
+
`metadata.json` mirrors the policy, features, grid, and thresholds.
|
| 33 |
|
| 34 |
+
## Input features
|
| 35 |
|
| 36 |
+
Four numeric features, **in this exact order**:
|
| 37 |
|
| 38 |
```
|
| 39 |
+
1. a1_high_risk # 1.0 if agent1_score >= 0.3183
|
| 40 |
+
2. a2_vulnerable # 1.0 if agent2_score >= 0.45
|
| 41 |
+
3. a1_risk_ratio # agent1_score / 0.3183
|
| 42 |
+
4. a2_risk_ratio # agent2_score / 0.45
|
| 43 |
```
|
| 44 |
|
| 45 |
+
## Preprocessing
|
| 46 |
+
|
| 47 |
+
- Compute the two flags with the upstream thresholds above.
|
| 48 |
+
- Compute the two ratios as `score / threshold` (1.0 = exactly at the boundary).
|
| 49 |
+
- Keep the column order shown above; it must match `b["feature_cols"]`.
|
| 50 |
+
- On any error, missing value, or invalid score, output `REVIEW` directly.
|
| 51 |
+
|
| 52 |
## Output
|
| 53 |
|
| 54 |
+
Raw: `predict_proba(x)` returns probabilities for `["FAIL", "PASS", "REVIEW"]`
|
| 55 |
+
(the bundle's `classes` order). Take `argmax` for the verdict and `max` for
|
| 56 |
+
confidence.
|
| 57 |
+
|
| 58 |
+
Reference wrapper output:
|
| 59 |
+
|
| 60 |
+
```json
|
| 61 |
+
{"verdict": "FAIL", "confidence": 1.0, "status": "ok", "reason": "both agents flagged risk"}
|
| 62 |
+
```
|
| 63 |
+
|
| 64 |
+
`confidence` is the tree's class probability, **not** calibrated.
|
| 65 |
|
| 66 |
## Policy
|
| 67 |
|
|
|
|
| 78 |
post-check that forces `REVIEW` on any tree/policy mismatch — keep it if you
|
| 79 |
deploy this model.
|
| 80 |
|
| 81 |
+
Note: the saved `source_thresholds` fallback still lists Agent 2 as `0.51` from
|
| 82 |
+
an earlier run. Pass the current thresholds explicitly (`0.3183` / `0.45`).
|
| 83 |
+
|
| 84 |
## Usage
|
| 85 |
|
| 86 |
```python
|
| 87 |
import joblib
|
| 88 |
|
| 89 |
+
b = joblib.load("model.joblib")
|
| 90 |
|
| 91 |
+
a1_high = a1_score >= 0.3183
|
| 92 |
+
a2_high = a2_score >= 0.45
|
| 93 |
x = [[float(a1_high), float(a2_high),
|
| 94 |
+
a1_score / 0.3183, a2_score / 0.45]]
|
| 95 |
|
| 96 |
proba = b["model"].predict_proba(x)[0]
|
| 97 |
verdict = b["classes"][int(proba.argmax())]
|