openflukeEXPLORATORY RUNS · AAI / TIDE / LUCY
Welvet experiments.
4 experiment series · 90,334 model cells run. Cameral LR sweeps, Test53 dayroute, a volatile Linode soak, and a Pi 3B 1GB layer farm — each with a plain-language writeup and links to the OpenFluke file server.
WATCH FIRST
The Cameral Graph Experiment.
A short walkthrough of Experiment 01 — Lucy live-fit, cameral sandwiches, and why target-propagation modes keep serving while they learn. Watch this before opening the panels if you want the shape of the series.
Model counts are completed Tide cells (dtype × mode × arch configs) from the archived boards. PDFs and JSON open as direct file links on each experiment page — the file-host folder root is not a browsable directory.
EXPERIMENT 01 · LUCY · GPT-CHAR · CAMERAL
Cameral learning-rate sweeps
Thirteen Tide Lucy reports for a tiny Shakespeare character model: same live GPT stack, different cam counts and learning rates, on desktop, Mac M4/M5, and Acer Nitro machines. Mid-high learning rates plus sandwich train modes beat polite SGD on Lucy Score.
Open the experiment 02 / 04 EXPERIMENTSDOCUMENTEDEXPERIMENT 02 · TEST53 · DAYROUTE · TRAIN MODES
Test53 training modes
Test53 dayroute farms on Mac M4/M5: dense × train mode × dtype × “funny” learning rates, with and without cameral-3. Tween / Split families crush classical BP on Lucy Score; mid-to-high LRs are a feature on discrete storage, not a bug.
Open the experiment 03 / 04 EXPERIMENTSDOCUMENTEDEXPERIMENT 03 · TIDE · OCEAN · LINODE SOAK
Volatile online farm
Three ~$5 Linode boxes ran Tide test51 at full CPU for days/weeks in a noisy public-cloud environment. The point was volatility and longevity — and nothing broke. Ocean consolidated the mid-run Lucy leaders.
Open the experiment 04 / 04 EXPERIMENTSDOCUMENTEDEXPERIMENT 04 · EDGE · RASPBERRY PI · 1GB
Pi 3B · 1GB RAM layer farm
Twenty-two Welvet layer sprints on Raspberry Pi 3B-class hosts with ~1GB RAM. Dense/CNN/norms flew; heavy MHA/LSTM/Mamba still finished cells. Matched mode-vs-SGD boards: non-SGD beats SGD on Acc ~97–98% of the time.
Open the experiment