The Proving Grounds

A lab you can walk through: the Overworld is the world, the Verdict is the judge, and certification is the serious bench underneath both. Evidence-backed, same harness that gates our own deploys. How certification works →

Walk the lab

The world

The Overworld

Walk six real districts — a tile world built from the same engine the fleet runs on. Nothing staged.

The judge

The Verdict

Judge a query against two retrieved passages, rapid-fire. The engine's live ranking is revealed after your call.

The bench

Get Certified

The serious bench underneath the games: a paid, reproducible benchmark run with a published, linkable scorecard.

Evaluation tiers

Basic

Chunk + bi-encoder retrieve

Intermediate

Hybrid sparse+dense + rerank

Advanced

Multi-hop, decomposition, HyDE

Benchmark suite

Exam	Tier	nDCG@10
BEIR MS MARCO slice	basic	0.54
Multi-hop bridge docs	advanced	0.41
HyDE query augmentation	advanced	0.48
Hybrid BM25 + dense	intermediate	0.57
Rotating eval slice (spec 0005)	basic	0.61
arXiv embedding/retrieval MCQ	intermediate	—

Reference leaderboard · fixtures

Reference fixtures from the eval-public panel. Live per-workspace metrics appear here once a model is trained and the API is reachable (SYNTH_API_KEY set).

Top · Plain InfoNCE

erank 3.8

nDCG@10 (reference data)0.39 · deploy gate 0.35 ✓

text-embedding-3-small

erank 11.2

nDCG@10 (reference data)0.48 · deploy gate 0.35 ✓

e5-large-v2

erank 14.6

nDCG@10 (reference data)0.51 · deploy gate 0.35 ✓

BGE-M3

erank 18.4

nDCG@10 (reference data)0.54 · deploy gate 0.35 ✓

Qwen3-Embedding-0.6B

erank 22.1

nDCG@10 (reference data)0.56 · deploy gate 0.35 ✓

Baseline comparison on dumbmodel.com →