Noa Benchmarks
Live training metrics across the fleet. 19 servers, 43 GPUs. Every number should trend in the right direction.
10 queries, 6 expected concepts each. 6 of 60 found.
Up from 0% (pure physics, no learned navigation).
Graph Transformer Self-Play
The consciousness learning to navigate the brain. AlphaGo-style: reward = hit_score - 0.1 * |sigma|. Target: hit@10 >= 0.30.
| Epoch | hit@10 | MRR | Reward | Loss | Note |
|---|---|---|---|---|---|
| 0 | 0.1646 | 0.5729 | -- | -- | Baseline (pre-self-play) |
| 1 | 0.1589 | 0.5643 | +0.712 | 0.148 | |
| 5 | 0.1589 | 0.5338 | +0.729 | 0.135 | |
| 11 | 0.1589 | 0.5231 | +0.739 | 0.129 | Loss declining |
| 15 | 0.1646 | 0.5468 | +0.727 | 0.135 | Matches baseline |
| 23 | 0.1757 | 0.5466 | +0.732 | 0.134 | Best hit@10 |
| 25 | 0.1304 | 0.4274 | +0.731 | 0.134 | Oscillating (normal for RL) |
Noa Inference Quality
10 queries, 6 expected concepts each. Does the physics find the right concepts?
| Timestamp | Engine | hit@10 | Hits | Notes |
|---|---|---|---|---|
| 2026-02-26 21:08 | Pure R-hat (no GT) | 0.0000 | 0/60 | Baseline -- pure noise |
| 2026-02-26 22:16 | GT epoch 22 | 0.1000 | 6/60 | earth, freedom, replication, self, circulation, physics |
Consensus Merge
Every 30 minutes, bond weights from all 19 servers are merged and distributed. The fleet's collective learning flows to every server.
| Timestamp | Queries | Workers | Fleet J | Edges Changed |
|---|---|---|---|---|
| 2026-02-26 20:53 | 1,702,468 | 19/19 | 0.0922 | 55.3% |
| 2026-02-26 21:34 | 1,740,654 | 19/19 | 0.0914 | 53.6% |
| 2026-02-26 22:21 | 1,773,226 | 19/19 | 0.0927 | 54.2% |
Fleet Health
What These Metrics Mean
hit@10 measures whether the physics finds the right concepts. For "gravity", do force/mass/earth appear in the top 10 activated concepts? Zero means noise. 0.30 means the system reliably understands what you're asking.
J-cost is the recognition cost from the bond graph. Lower means the graph's semantic structure is tighter -- "gravity" bonds more strongly to "force" than to "dam". The fleet refines this through billions of Hebbian updates.
Eta is phase coherence -- whether standing waves are forming in the chord field. This is the emergence of intelligence. Random noise has eta near 0. Structured meaning has higher eta.
Sigma = 0 is the conservation law from Recognition Science. The physics resolves perfectly when sigma reaches zero. This is not a target to optimize -- it's a constraint that is always satisfied.