GRAPH.md

# FACTPROP graph checkpoint

`factprop_graph_v1.pkl` is the unchanged original checkpoint, 125,274,720 bytes.

SHA-256: `437a434260edbb019c85575d65b4775cb2461f145e17966ee3acc8e4625ce7c8`

## Structure

The pickle contains a dictionary. `payload["graph"]` is a NetworkX **DiGraph**, with 100,015 nodes and 432,562 stored directed edges. Additional keys are `state`, `seed_entities`, `validator_state`, and `scheduler_stats`, retained as original construction metadata.

Node identifiers are entity-label strings. Node attributes are `qid` and `qid_status`; QID mappings may be missing or shared by multiple nodes.

| Edge field | Meaning |
| --- | --- |
| `relation` | Relation identifier |
| `question` | Associated natural-language question |
| `surface` | Natural-language statement |
| `evidence` | Stored supporting text |
| `confidence` | Construction-time confidence value |
| `group` | Construction category |
| `is_inverse` | Whether the edge is an inverse traversal edge |

There are 357,205 forward edges and 75,357 inverse edges. Popularity counts incoming edges with `is_inverse is not True`. Do not use total in-degree without filtering inverse edges when comparing to the browser index. Field presence does not guarantee that all values are non-empty or independently correct.

## Load and inspect

Install `networkx` and `huggingface_hub`, then run the accompanying `load_graph.py`. It downloads the graph, checks its checksum, loads the dictionary, and prints graph counts. Python pickle loading can execute code, so only load a checkpoint you trust.

```python
from load_graph import load_graph

graph = load_graph()
entity = "Apple Inc."
score = sum(1 for _, _, attrs in graph.in_edges(entity, data=True)
            if attrs.get("is_inverse") is not True)
assert score == 467
```

No graph nodes, mappings, edges, or checkpoint metadata were changed for this release. Mapping revisions must be maintained separately.