cp500/infon-sheaf-gnn

Infon Sheaf hypergraph GNN

Tiny (~140 K params, 370 KB FP16) line-graph Sheaf GNN that scores reasoning chains. Two outputs:

• 3-class verdict: SUPPORTS / REFUTES / NOT_ENOUGH_INFO
• Per-edge H¹ structural-discrepancy score (used as an anomaly prior by the symbolic reasoner)

Quick start (JavaScript)

npm install @cp500/infon-sheaf-gnn onnxruntime-web

import { SheafGNNModel } from '@cp500/infon-sheaf-gnn';

const model = await SheafGNNModel.fromHub('cp500/infon-sheaf-gnn');
const r = await model.scoreChain([
  { kind: 'connective', polarity: 1, confidence: 0.92,
     touchesSource: true, connectsPrev: true },
  { kind: 'connective', polarity: 1, confidence: 0.85,
     touchesTarget: true, isLast: true, connectsPrev: true },
]);
console.log(r.verdict);            // 'SUPPORTS'
console.log(r.totalDiscrepancy);   // small (clean chain)

The JS client source is mirrored under js/ for self-contained installs.

Architecture

Line-graph Sheaf GNN: each edge becomes a node; message passing flows forward + backward through per-relation-kind restriction maps. The Sheaf design types messages by 3-class kind (connective, terminal, reportive), not by predicate slug — which is what makes the model transferable across relation vocabularies.

chain → encode 10-d edge features → StalkEncoder → Linear(10, 64)
                                                         |
                                ┌── 3 × SheafMessageLayer (64) ──┐
                                │  • per-kind restriction maps   │
                                │  • forward + backward messages │
                                │  • H¹ discrepancy per edge     │
                                └─────────────┬──────────────────┘
                                              ▼
                                      ChainVerdictHead (3)

Evaluation

Best-by-val-accuracy checkpoint, trained on the default synthetic chain corpus (n=2000, 5 chain kinds: clean / retracted / anomaly / disconnected / cyclic):

Metric	Value
Best val accuracy	0.998
Per-kind: clean	0.995
Per-kind: retracted	1.000
Per-kind: anomaly	1.000
Per-kind: disconnected	1.000
Per-kind: cyclic	1.000

Files

File	Size	Purpose
onnx/sheaf_gnn.onnx	~627 KB	FP32 graph
onnx/sheaf_gnn.fp16.onnx	~370 KB	FP16 graph (default for browser)
gnn.pt	small	PyTorch state dict + config
meta.json	tiny	JS client contract
relation_kinds.json	tiny	predicate→kind lookup mirror
js/	~30 KB	source of @cp500/infon-sheaf-gnn

License

Apache 2.0 for both weights and JS code.