Ocean does not retrain — it polls /api/board. Lucy Score = Throughput × Availability × Acc / 10,000. Acc is learning. Availability is the live loop (SGD that blocks serve dies here). LPD condenses that live-fit into RAM. Honesty / test48 is where FastProxy and MeshTween show up. Remote Pis POST /api/register then Start all.
Mode names — [T]=Tween · [S]=Split · [FP]=FastProxy · [L]=Linear · [HP]=HeadProxy
Per-layer best-Score cells (Score > 0) vote for train mode, dtype, and cameral arch. Suggested default is the mode/dtype/arch that wins the most Lucy axes below. Rank goldilocks by Acc/Thru/Avail density, not Score/MiB.
Ocean-wide best cell on each metric. First three axes are the consciousness pillars (hard Acc, Throughput, Availability). Lucy Score is the Welvet SoftAcc product. acc_thru is hard Acc × throughput.
| axis | value | tide | mode | dtype | arch | cell |
|---|
Run and train at the same time, in a small box. Pure Acc is not enough — SGD can look accurate while it blocks serve. Score = Thru × Avail × Acc. Q = Acc/Thru/Avail vs learner peaks. Gold = all three ≥80% at ≤20% Acc-champ RAM. Honesty/test48 is the train-mode truth (FastProxy / MeshTween / Sparse vs SGD).
| mode | n | best Acc | smallest KiB | fastest Thru | cell |
|---|
| band | tide | cell | Acc% | Thru% | Avail% | Q% | RAM% | LPD | KiB |
|---|
Same Lucy axes, but the champion is picked inside each tide. Open a layer name to see its live dashboard.
| tide | lr | mode | dtype | arch | score | soft | acc | thru | adapt | done |
|---|
| tide | cell | score | soft | acc | avail |
|---|
Train-mode truth: Acc is argmax. SoftAcc is serve-confidence, not Score. Avail is the duty clock. Axes come from finished cells across tides — not a frozen mode list.