Publish artifacts for knowledge-graph-risk-engine-20260808

README.md

@@ -0,0 +1,61 @@
+---
+license: mit
+library_name: custom
+pipeline_tag: feature-extraction
+datasets:
+- RKB109/knowledge-graph-risk-engine-20260808-dataset
+tags:
+- synthetic-data
+- transparent-baseline
+- knowledge-graphs
+- token-classification
+- feature-extraction
+- question-answering
+- sentence-similarity
+metrics:
+- accuracy
+---
+
+# Knowledge Graph Risk Engine Baseline Model
+
+## Model Description
+
+This repository contains a small, transparent prototype model for
+**Risk teams need relationship-level explanations instead of opaque entity scores.**
+
+The model combines per-label token weights with IDF-weighted evidence
+retrieval. It was generated for reproducible architecture demonstrations and
+does not call a hosted LLM.
+
+## Evaluation
+
+- Held-out synthetic examples: 4
+- Accuracy: 1
+- Intended metrics: relation_accuracy, path_coverage, entity_resolution_precision
+
+## Intended Use
+
+- Architecture prototyping
+- CI and evaluation examples
+- Local baseline comparisons
+- Educational experimentation
+
+## Hugging Face Task Coverage
+
+- `token-classification`
+- `feature-extraction`
+- `question-answering`
+- `sentence-similarity`
+
+## Limitations and Risks
+
+All entities are fictional. Real identity or financial data requires governance, consent, and bias review.
+
+The dataset is synthetic and small. Do not use this model for consequential
+decisions without representative data, expert review, and production-grade
+evaluation.
+
+## Reproducibility
+
+The linked GitHub repository includes `train.py`, the exact dataset split,
+evaluation code, and the model JSON format.

evaluation.json

@@ -0,0 +1,5 @@
+{
+  "test_examples": 4,
+  "accuracy": 1,
+  "synthetic_evaluation": true
+}

inference.py

@@ -0,0 +1,80 @@
+"""Transparent baseline pipeline for the generated AI project."""
+
+from __future__ import annotations
+
+import json
+import math
+import re
+from pathlib import Path
+
+
+def tokenize(value: str) -> list[str]:
+    return re.findall(r"[a-z0-9]+", value.lower())
+
+
+class Pipeline:
+    def __init__(self, model: dict):
+        self.model = model
+
+    @classmethod
+    def from_file(cls, path: str | Path) -> "Pipeline":
+        return cls(json.loads(Path(path).read_text(encoding="utf-8")))
+
+    def classify(self, text: str) -> tuple[str, float]:
+        tokens = tokenize(text)
+        scores = {
+            label: sum(weights.get(token, 0) for token in tokens)
+            for label, weights in self.model["prototypes"].items()
+        }
+        ranked = sorted(scores.items(), key=lambda item: (-item[1], item[0]))
+        label, best = ranked[0]
+        total = sum(max(score, 0) for _, score in ranked) or 1
+        return label, best / total
+
+    def search(self, query: str, limit: int = 3) -> list[dict]:
+        query_tokens = set(tokenize(query))
+        ranked = []
+        for document in self.model["documents"]:
+            document_tokens = set(tokenize(document["text"]))
+            lexical = sum(
+                self.model["idf"].get(token, 1.0)
+                for token in query_tokens & document_tokens
+            )
+            ranked.append({**document, "score": round(lexical, 6)})
+        return sorted(ranked, key=lambda item: (-item["score"], item["id"]))[:limit]
+
+    def graph_evidence(self, text: str) -> list[dict]:
+        tokens = set(tokenize(text))
+        matches = []
+        for subject, relation, target in self.model.get("graph_edges", []):
+            edge_tokens = set(tokenize(f"{subject} {relation} {target}"))
+            overlap = len(tokens & edge_tokens)
+            if overlap:
+                matches.append(
+                    {
+                        "subject": subject,
+                        "relation": relation,
+                        "target": target,
+                        "overlap": overlap,
+                    }
+                )
+        return sorted(matches, key=lambda item: -item["overlap"])
+
+    def run(self, text: str) -> dict:
+        label, confidence = self.classify(text)
+        evidence = self.search(text)
+        result = {
+            "prediction": label,
+            "confidence": round(confidence, 4),
+            "requires_review": confidence < self.model["confidence_threshold"],
+            "evidence": evidence,
+        }
+        if self.model["mode"] == "graph":
+            result["graph_evidence"] = self.graph_evidence(text)
+        if self.model["mode"] == "agent":
+            result["proposed_tool"] = label
+            result["approval_required"] = label in {
+                "request-approval",
+                "request-human-help",
+            }
+        return result

model.json

@@ -0,0 +1,192 @@
+{
+  "format": "daily-project-prototype-v1",
+  "project": "knowledge-graph-risk",
+  "title": "Knowledge Graph Risk Engine",
+  "domain": "knowledge-graphs",
+  "mode": "graph",
+  "labels": [
+    "ownership",
+    "transaction",
+    "location"
+  ],
+  "prototypes": {
+    "ownership": {
+      "who": 3,
+      "controls": 6,
+      "supplier": 6,
+      "alpha": 9,
+      "company": 7,
+      "in": 2,
+      "an": 3,
+      "operations": 2,
+      "review": 2,
+      "for": 1,
+      "evaluation": 1,
+      "case": 1,
+      "which": 2,
+      "owns": 4,
+      "vendor": 4,
+      "gamma": 4,
+      "delta": 2
+    },
+    "transaction": {
+      "17": 8,
+      "in": 2,
+      "an": 4,
+      "operations": 2,
+      "review": 2,
+      "which": 2,
+      "account": 10,
+      "paid": 4,
+      "vendor": 4,
+      "beta": 4,
+      "for": 2,
+      "evaluation": 2,
+      "case": 2,
+      "what": 3,
+      "transfer": 3,
+      "connects": 3,
+      "and": 3,
+      "supplier": 6,
+      "alpha": 6,
+      "transferred": 3,
+      "to": 3
+    },
+    "location": {
+      "where": 2,
+      "is": 2,
+      "warehouse": 4,
+      "north": 4,
+      "located": 6,
+      "in": 5,
+      "region": 6,
+      "east": 2,
+      "for": 2,
+      "an": 3,
+      "evaluation": 2,
+      "case": 2,
+      "operations": 1,
+      "review": 1,
+      "what": 2,
+      "contains": 2,
+      "facility": 4,
+      "blue": 4,
+      "west": 2
+    }
+  },
+  "idf": {
+    "17": 1.847298,
+    "company": 1.847298,
+    "alpha": 1.847298,
+    "controls": 2.252763,
+    "supplier": 1.847298,
+    "account": 1.847298,
+    "paid": 2.252763,
+    "vendor": 1.847298,
+    "beta": 2.252763,
+    "warehouse": 2.252763,
+    "north": 2.252763,
+    "located": 1.847298,
+    "in": 1.847298,
+    "region": 1.847298,
+    "east": 2.252763,
+    "delta": 2.252763,
+    "owns": 2.252763,
+    "gamma": 2.252763,
+    "transferred": 2.252763,
+    "to": 2.252763,
+    "facility": 2.252763,
+    "blue": 2.252763,
+    "west": 2.252763
+  },
+  "documents": [
+    {
+      "id": "graph-01",
+      "label": "ownership",
+      "text": "company-alpha controls supplier-alpha",
+      "metadata": {
+        "synthetic": true,
+        "domain": "knowledge-graphs"
+      }
+    },
+    {
+      "id": "graph-02",
+      "label": "transaction",
+      "text": "account-17 paid vendor-beta",
+      "metadata": {
+        "synthetic": true,
+        "domain": "knowledge-graphs"
+      }
+    },
+    {
+      "id": "graph-03",
+      "label": "location",
+      "text": "warehouse-north located-in region-east",
+      "metadata": {
+        "synthetic": true,
+        "domain": "knowledge-graphs"
+      }
+    },
+    {
+      "id": "graph-04",
+      "label": "ownership",
+      "text": "company-delta owns vendor-gamma",
+      "metadata": {
+        "synthetic": true,
+        "domain": "knowledge-graphs"
+      }
+    },
+    {
+      "id": "graph-05",
+      "label": "transaction",
+      "text": "account-17 transferred-to supplier-alpha",
+      "metadata": {
+        "synthetic": true,
+        "domain": "knowledge-graphs"
+      }
+    },
+    {
+      "id": "graph-06",
+      "label": "location",
+      "text": "facility-blue located-in region-west",
+      "metadata": {
+        "synthetic": true,
+        "domain": "knowledge-graphs"
+      }
+    }
+  ],
+  "graph_edges": [
+    [
+      "company-alpha",
+      "controls",
+      "supplier-alpha"
+    ],
+    [
+      "account-17",
+      "paid",
+      "vendor-beta"
+    ],
+    [
+      "warehouse-north",
+      "located-in",
+      "region-east"
+    ],
+    [
+      "company-delta",
+      "owns",
+      "vendor-gamma"
+    ],
+    [
+      "account-17",
+      "transferred-to",
+      "supplier-alpha"
+    ],
+    [
+      "facility-blue",
+      "located-in",
+      "region-west"
+    ]
+  ],
+  "confidence_threshold": 0.18,
+  "trained_on_synthetic_data": true
+}

project.json

@@ -0,0 +1,52 @@
+{
+  "name": "Knowledge Graph Risk Engine",
+  "problem": "Risk teams need relationship-level explanations instead of opaque entity scores.",
+  "domain": "knowledge-graphs",
+  "architecture": "graph",
+  "hugging_face_tasks": [
+    "token-classification",
+    "feature-extraction",
+    "question-answering",
+    "sentence-similarity"
+  ],
+  "recommended_stack": [
+    "FastAPI for entity and evidence APIs",
+    "Neo4j Community or PostgreSQL recursive queries",
+    "Sentence Transformers for entity resolution",
+    "NetworkX for local graph validation",
+    "Kafka-compatible event ingestion",
+    "OpenTelemetry for lineage and query traces"
+  ],
+  "real_world_data_sources": [
+    {
+      "name": "SEC EDGAR submissions API",
+      "url": "https://data.sec.gov/submissions/CIK0000320193.json",
+      "purpose": "Public company and filing relationships"
+    },
+    {
+      "name": "GLEIF LEI API",
+      "url": "https://api.gleif.org/api/v1/lei-records?page[size]=5",
+      "purpose": "Public legal-entity identifiers and relationships"
+    }
+  ],
+  "job_description_skills": [
+    "Entity resolution and relation extraction",
+    "Knowledge-graph modeling and path queries",
+    "Graph-based explainability and provenance",
+    "Streaming ingestion and schema evolution",
+    "Risk-model evaluation and data quality controls"
+  ],
+  "impact_targets": [
+    "Reach entity-resolution precision >= 0.95 on reviewed pairs",
+    "Return evidence paths for 100% of emitted risk flags",
+    "Process 10,000 relationship events per minute in load tests",
+    "Detect schema and orphan-node regressions in CI"
+  ],
+  "baseline_evaluation": {
+    "test_examples": 4,
+    "accuracy": 1,
+    "synthetic_evaluation": true
+  },
+  "estimated_delivery": "8-12 weeks for one engineer",
+  "generated_baseline_is_production_ready": false
+}