Lost contact with the Evolvr server. Start it with ~/.venvs/default/bin/python ~/evolvr/dashboard/server.py. Retrying...
Inside Evolvr's head
Live, from the model itself. What it is doing, the same problems re-solved as it learns (each digit shaded by how sure it is, written right to left like it writes them), what it is studying, and the calls it makes about its own answers.
Connecting...
Watching the same problems
Its diary narrated by the program from real numbers
Studying right now
Exam
Teacher's notebook
How often it gets each kind of problem right (latest quizzes, weakest first)
Stubborn mistakes it keeps bringing back
Just fixed for good
Its calls on its own answers
Learning to read and write
His language brain, grown from scratch: random weights and his own word pieces. He reads children's stories and learns to guess the next word. Every so often he takes a test on stories he has never seen, then writes one of his own.
Writing right now
He's given the bold opening and writes the rest himself. Each word piece is shaded by how sure he was when he picked it (darker = surer). Hover or tap one to see the other words he was thinking of.
Everything he's written, newest first
Teaching himself: the council
Three copies of him, each grown from a different random start, learn addition up to the taught length. Then they write answers to longer problems nobody showed them. An answer is kept only if all three agree and it passes the rules he was given (length, last digit, casting out nines and elevens). Before going one digit longer, every member must score at least the gate on a hidden exam. The true answers are used only to grade the filter, never to teach him.
What happened, newest first
Report card
School: placement test, teaching, power training on weak spots, then the super-intense exam (100% on every kind, plus no forgetting earlier grades) before moving up a grade.
Knowledge map subject › topic › sub-topic › kind › finer sub-topics it split off where mistakes bunch up
Latest exam, weakest kinds first
Mind map: accuracy by number length, round by round
Each row is a round of self-improvement; each column is a number length in digits. Bracket = taught with real answers. Notch = how far it has taught itself. Everything right of the notch is pure generalization.
Learning curve
Training loss (log scale). Dashed lines mark the start of each self-improvement round.
Birth vs now
Accuracy by number length: right after seed training, and at the latest round.
Self-verification
Per round, on problems one digit past the frontier: how often its own answers were right, how many it kept after its checks, and how many of the kept ones were actually right. Hidden truth is used here only to score it.
Brain
What this run's mind is made of and how it is being raised.
Resources (live)
Changes apply within a few seconds, no restart.
Lifecycle
One continuous being. Its memory, mistake ledger and self-model carry through every stage.
The rungs of RSI
From ROADMAP.md. A rung counts only when a logged run shows it.
Raise a new mind
Start another run with different checks or settings. Max 2 at once; RAM is capped per run.