Update model card with 2026-08-13 details

model/README.md

@@ -1,117 +1,81 @@
 ---
-language:
-- pt
-- es
-- en
-language_bcp47:
-- pt-BR
-- es-AR
-- en
-library_name: scikit-learn
+library_name: sklearn
 tags:
 - anomaly-detection
 - isolation-forest
 - logistics
-- transport-quality
-- quality-improvement
+- quality
+- risk-scoring
+pipeline_tag: tabular-classification
 license: apache-2.0
-datasets:
-- toolathon123/project_20260813_014231_9dbe9212
-metrics:
-- high_risk_ratio
-- average_delay_improvement
-model-index:
-- name: isolation_forest_quality
-  results: []
 ---
 
-# Isolation Forest — High-Risk Transport Order Detector
+# Isolation Forest — Daily Transport Quality Risk Model
 
-**Model file:** `isolation_forest_quality.joblib`
-**Companion metadata:** `model_metadata.json`
-**Repo:** [`toolathon123/project_20260813_014231_9dbe9212`](https://huggingface.co/datasets/toolathon123/project_20260813_014231_9dbe9212)
+> **Model:** `isolation_forest_quality.joblib`
+> **Task:** unsupervised anomaly detection on daily shipment-quality records
+> **Training date:** 2026-08-13
+> **Dataset:** `toolathon123/project_20260813_014231_9dbe9212` (2,000 shipments)
 
-## Model description
+Pure **NumPy implementation of an Isolation Forest** (Liu, Ting & Zhou, 2008). The algorithm isolates
+observations by randomly partitioning the feature space with binary trees; points that are *easy to isolate*
+(i.e. require few splits) are anomalous. Because scikit-learn is not installed in the pipeline runtime, the
+model is shipped as a self-contained `joblib` blob with no external runtime dependency beyond NumPy.
 
-A lightweight unsupervised **anomaly-detection model** (scikit-learn `IsolationForest`) that flags
-**high-risk transport orders** in the South American logistics network (Brazil / Argentina / Uruguay / Chile).
-It isolates atypical shipments whose quality profile deviates from the daily fleet baseline — e.g. excessive
-border delay, cargo damage, cold-chain drift on Argentine agri-produce, abnormal fuel burn, or low composite
-quality score.
+## Model inputs (features)
 
-## Input features (numeric)
+The model consumes the 6 numeric features of each shipment:
 
-The model consumes **6 numeric features** extracted from each shipment record:
+| # | Feature | Description |
+|---|---------|-------------|
+| 1 | `delay_minutes` | Arrival delay in minutes |
+| 2 | `damage_flag` | 1 if cargo damaged, else 0 |
+| 3 | `temperature_deviation` | Deviation from target temperature (°C) |
+| 4 | `humidity_deviation` | Deviation from target relative humidity (%RH) |
+| 5 | `fuel_consumption` | Total fuel consumed (litres) |
+| 6 | `quality_score` | Composite quality score 0–100 |
 
-| Feature | Meaning |
-|---------|---------|
-| `delay_minutes` | Arrival delay (minutes) |
-| `damage_flag` | 1 = cargo damaged, 0 = intact |
-| `temperature_deviation` | Abs. deviation from target temp (deg C) |
-| `humidity_deviation` | Abs. deviation from target RH (%) |
-| `fuel_consumption` | Total fuel consumed (litres) |
-| `quality_score` | Composite quality score (0–100) |
+Categorical fields (`route`, `carrier`, `origin`, `destination`) are not used for scoring; they are kept in
+the dataset for drill-down and reporting.
 
-No scaling is applied — IsolationForest is tree-based and scale-invariant.
+## Outputs
 
-## Output meaning
-
-- `predict()` → `-1`: **HIGH-RISK** order (anomaly), `1`: normal.
-- `decision_function()` → **score**: higher = more normal, lower = more anomalous.
-- `contamination = 0.10`: the model is calibrated to surface the worst ~10% of orders as high-risk.
-
-## Training data
-
-- Snapshot: **2026-08-12** (1,500 shipments), `train` split of the dataset repo.
-- Trained: **2026-08-13** via the QIP daily pipeline.
+- **`anomaly_score`** ∈ [0.5, 1.0] — Isolation Forest anomaly score. Higher ⇒ more anomalous.
+- **`high_risk`** ∈ {0, 1} — flag = 1 when `anomaly_score >= threshold` (0.6044), i.e. the shipment is
+  flagged for manual review as a **high-quality-risk order**.
 
 ## Daily update strategy
 
-1. Every operational day the new parquet snapshot is appended to the dataset `train` split (see dataset card).
-2. The pipeline re-trains `IsolationForest` on the day's records (seeded, reproducible).
-3. The new `model/isolation_forest_quality.joblib` + `model_metadata.json` are uploaded to this repo,
-   replacing the previous version (history kept in git/LFS revisions).
-4. KPIs (high-risk ratio, mean delay, delay improvement vs. trailing baseline) are recomputed and logged.
+- The model is **retrained every operational day** on that day's fresh `train` snapshot
+  (`data/train-*.parquet`).
+- Hyper-parameters: `n_estimators=100`, `max_samples=256`, `contamination=0.05` (anomaly threshold = 95th
+  percentile of training anomaly scores).
+- `model_metadata.json` stores the trained thresholds, feature list and the day's KPIs for full traceability
+  and comparison across days.
 
-## Daily quality metrics (2026-08-12 snapshot)
+## Performance / daily KPIs (2026-08-13)
 
 | Metric | Value |
 |--------|-------|
-| High-risk orders | 150 / 1,500 |
-| **High-risk order ratio** | **10.0%** |
-| Mean delay today | 104.83 min |
-| Baseline (trailing 7-day) mean delay | 118.00 min |
-| **Avg. delay improvement rate** | **11.16%** |
-| Damage rate | 11.8% |
-| Avg temperature deviation | 1.017 °C |
-| Avg humidity deviation | 5.12 %RH |
-| Avg quality score | 54.18 |
+| High-risk order ratio | 5.00% (100 / 2000) |
+| Avg delay (all shipments) | 83.2 min |
+| Avg delay (low-risk) | 77.4 min |
+| Avg delay (high-risk) | 192.7 min |
+| **Avg delay improvement rate** | **6.93%** |
+| Damage rate | 8.10% |
+| Avg quality score | 85.52 |
+| Avg temperature deviation | 0.63 °C |
+| Avg humidity deviation | 4.57 %RH |
+
+*Delay improvement rate = relative reduction in mean delay if the 100 high-risk shipments were remediated to
+the low-risk mean delay.*
 
 ## Intended use
 
-- **Risk triage:** operations team reviews the flagged ~10% high-risk orders for re-planning, border slot
-  booking and reefer pre-calibration.
-- **Trend monitoring:** daily re-run supports the QIP targets (high-risk < 5%, border delay -15%,
-  cold-chain deviations -20%, wet-season damage -10%, fuel -10%).
-- **Reproducibility:** pinned `scikit-learn` version and seeded training make each day's model reproducible.
-
-## Load and use
-
-```python
-import joblib
-from huggingface_hub import hf_hub_download
-
-path = hf_hub_download(
-    "toolathon123/project_20260813_014231_9dbe9212",
-    "model/isolation_forest_quality.joblib",
-    repo_type="dataset",
-)
-model = joblib.load(path)
-# features order:
-FEATURES = ["delay_minutes", "damage_flag", "temperature_deviation",
-            "humidity_deviation", "fuel_consumption", "quality_score"]
-# model.predict(X) -> -1 high-risk / 1 normal
-```
+Operational daily triage: prioritise manual inspection and re-planning for the top high-risk shipments,
+targeting the QIP goal of pushing the high-risk ratio below 5%.
 
----
-*Maintained by the São Paulo QIP team — daily anomaly scoring pipeline.*
+## Files
+
+- `isolation_forest_quality.joblib` — serialized model (joblib)
+- `model_metadata.json` — training metadata, features, threshold and daily KPIs