Six instruments, running live.

A fully-connected network with two tanh hidden layers and a sigmoid output, trained on mini-batches with momentum against binary cross-entropy. The background is the model’s decision surface, re-evaluated on a 44×44 grid every frame — you are watching the weights actually move. Click anywhere on the plot to add your own training point and see the boundary bend to accommodate it.

Centroids start on randomly chosen data points — deliberately naive, so you can watch real convergence rather than a lucky initialisation. Each iteration assigns every point to its nearest centroid, then moves each centroid to the mean of its members. The run stops when no point changes cluster. Inertia is the within-cluster sum of squares.

SHA-256 is implemented directly in this page — message padding, the 64-round compression function, all of it — rather than calling crypto.subtle, so you can verify it is genuinely doing the work. The grid is the 256 output bits of your input’s digest. Flip a single bit of the input and roughly half of those output bits change: that is the strict avalanche criterion, and it is what makes a hash usable for integrity.

Naive strength meters count character classes. This one starts from length × log₂(pool), then subtracts bits for the patterns an attacker would actually exploit — dictionary hits, character runs, repeats, keyboard walks and date-like digits — and converts the surviving guess count into wall-clock time under four attack models. Nothing you type is transmitted or stored.

Encrypt something, then throw the key away and press Break it. For Caesar, all 26 shifts are scored with a chi-squared test against English letter frequencies. For Vigenère, the key length is first estimated from the index of coincidence — English text sits near 0.067, random text near 0.038 — and then every column is solved as an independent Caesar. The histogram shows observed frequencies against expected English. Frequency analysis needs volume — shorten the text much below a paragraph and watch the key-length estimate start guessing, which is exactly the real limitation of the technique.

The event stream below is simulated — but the detector scoring it is not. Each event is reduced to a feature combining how bursty its source currently is with how rare that (source, event-type) pair has been so far. A running mean and variance, maintained with Welford’s online algorithm, turn that feature into a z-score. Anything past the threshold is flagged. Press Inject attack to send a burst from a hostile host and watch the score break out of the baseline.