AVRAI demos
HGCNN benchmarklexical graph v2RecBole comparison

HGCNN benchmark proof demo

This page shows the current simulated benchmark harness: 500 train rows, 1000 holdout/test rows, and 500 holdout queries with ideal answers. It compares AVRAI's HGCNN paths with five RecBole baselines using the same frozen fixture, then shows why the result is still shadow evidence instead of empirical promotion.

Public proof boundary
simulated sourceshadow evidenceno live actionpromotion blocked
Train rows
500
used for fitting
Holdout rows
1000
test surface
Holdout queries
500
ideal answer check
RecBole reports
5
external baselines
Real packets
0
required next
State
blocked
not_empirical_ready
Apples-to-apples score readback
How well the ideal answer is ranked in the top three, with higher rank rewarded.
#1AVRAI
HGCNN trained shadow
candidate ranker

The trained HGCNN candidate scorer put the ideal simulated answer first on the 500 holdout queries.

1.0
#2AVRAI
Actualizer-bound HGCNN
governed decision layer

The governed Actualizer readback kept the ideal answer in the top three while applying safety and decision gating.

0.7158
#3RecBole
KGIN
knowledge graph baseline

Best RecBole baseline on NDCG@3 for this exported simulated fixture.

0.6415
#4RecBole
LightGCN
graph collaborative filtering

Strong graph collaborative filtering baseline, but lower top-three coverage than the HGCNN paths.

0.6131
#5RecBole
KGNN-LS
knowledge-aware neural sampler

This model did not fit the fixture shape well in the external run.

0.0198
#6RecBole
BPR
pairwise ranking baseline

Classic pairwise baseline included as a low-complexity control.

0.0091
#7RecBole
KTUP
translation-based KG recommender

Translation-style KG recommendation underperformed on this simulated ranking task.

0.0032
Best RecBole baseline
KGIN

NDCG@3 0.6415

HGCNN candidate path
1.0

trained shadow NDCG@3

Actualizer-bound path
0.7158

governed readback NDCG@3

Benchmark harness graph
Click nodes to inspect the evidence refs

Full harness path from simulated packets to blocked decision receipt.

10 nodes / 10 edges
PacketsDatasetRecBoleHGCNNCompareDecision
Six proving pieces
1. Simulated packet shape
passed

Creates unique packet-shaped rows with ideal answers, then checks temporal holdout and no raw payload storage.

simulated-packet-shape-validation:hgcnn-lexical-actualizer:v1
2. Frozen fixture dataset
passed

Locks the 500 train, 1000 holdout, and 500 query split so every model sees the same task.

frozen-simulated-scale-500-1000-500-dataset:hgcnn-lexical-actualizer:v1
3. RecBole export/import
passed

Translates the fixture into RecBole run material and verifies the selected baseline reports.

recbole-run-import:hgcnn-lexical-actualizer:v1
4. HGCNN shadow run
passed

Runs the AVRAI HGCNN candidate scorer and Actualizer-bound readback without live authority.

trained-shadow-cross-reference:hgcnn-lex-graph-v2:v1
5. External baseline comparison
passed

Compares HGCNN and Actualizer metrics against BPR, LightGCN, KGNN-LS, KTUP, and KGIN.

recbole-hgcnn-cross-reference:simulated-scale-500-1000-500:v1
6. Promotion decision receipt
blocked

Carries the result as shadow evidence and blocks empirical promotion until real validator-clean packet rows exist.

decision-receipt:recbole-run-import:simulated-scale-500-1000-500:v1
Next evidence required
not_empirical_ready
external_recbole_cross_reference_passed_shadow_only
missing_validator_clean_real_packet_rows_for_same_scope
missing_full_empirical_promotion_gate
Required before promotion
validator-clean-m31-p10-64-exported-packet-rows:same-scope
promotion-decision-receipt:same-scope:hgcnn-lexical-actualizer