See what Veritate captures.
Sample dataEvery time Veritate generates a byte, it records a full MRI of the model's internal state. This is a demonstration of the interpretability data captured during generation: per-token confidence, next-byte candidates, feed-forward neuron activity, the logit lens forming a prediction across all 24 layers, attention, the residual stream, a decision trace, and a hallucination grade. Press Generate to cycle through sample prompts and watch every panel populate and animate.
Model MRI
A demonstration of the interpretability data Veritate captures during generation, replaying sample generations so every readout stays populated. The scrubber, hover, and neuron modal are all interactive.
output + timeline raw generated bytes; scrub backward to see the brain at any earlier moment
The technical byte view: the model's raw output. Each generated byte is a frame. Drag the slider and every panel below reflects the brain's state at that instant. Click live to snap back to the most recent token. Bytes that aren't printable show as a hex glyph.
letter timing ms per generated byte. scrub the slider above to inspect any frame.
Each bar is one generated byte. Bar height is wall-clock ms for that forward pass. The cursor marks the current frame.
telemetry surprise, uncertainty, latency over time
surprise (red) is the byte we sampled. uncertainty (blue) is the model's prediction before we sampled. latency (gray) is your CPU.
panoramic ffn brain every neuron, every layer, current frame
One row per transformer layer (top = "sensory", bottom = "output"). Each row max-pools all FFN neurons into buckets. Brightness = activation magnitude. Hover a cell to highlight it; click to open its neuron.
logit lens what each layer would predict if the network stopped there
Each row = what the model would predict if it stopped at that layer. Gold row = first layer whose top-1 matches the byte that was actually sampled.
memorization fingerprint training stories closest to current firing pattern
For the strongest-firing neurons right now, which training stories activated those same neurons during a probe pass. The model's nearest training memories.
candidates top 12 next-byte predictions
The probability distribution before sampling. Tight = the model is sure. Flat = it's guessing.
residual stream depth how much information has accumulated by each layer
L2 norm of the residual after each layer. Growth = layer added context.
per-layer contribution which layer is actually working for this byte
How much each layer changed the residual for this byte. Big bar = layer did real work.
decision trace direct logit attribution - who voted for which byte
top contributors to picked byte -
top contributors to expected byte -
The left list explains why the byte that got sampled was sampled. The right list explains the byte the model expected (its argmax). Click any neuron to see what it's specialized for.
candidate decision trace why each top-K candidate byte got the votes it did
Pick any of the top-K candidate bytes the model considered. Each shows the same DLA decomposition, but for that specific byte.
confidence four-component calibrated probability
Per-token calibrated probability the sampled byte matches a held-out target. Composite of logit margin, entropy, lens consistency, and residual stability.
hallucination auto-scored after every generation
The verdict and its one-line reason lead; the answer is shown with each word underlined by grounding and colored by confidence. The key below explains the colors.
top firing neurons identifiers of the strongest 8 neurons per layer
The 8 neurons firing hardest right now per layer. Hover to highlight; click any cell to see the training stories that activated that neuron hardest.
per-layer decisiveness at each layer, did the model commit to one byte or hedge?
Tall bars = the layer made up its mind; short bars = the layer was hedging. Bars are tinted by region. The number is max_abs / mean_abs of that layer's logit-delta vector.
live co-activation which neurons fire together as bytes stream
Pairs of FFN neurons that fired strongly on the same generated bytes during this generation. Resets when you start a new generation. Click either neuron to inspect it.
information flow which input positions did attention pull from?
Aggregating across all layers and heads, brightness shows how much total attention each input position received. Stacks into multiple rows as the sequence grows.
byte to character reference what each number means
Veritate sees text as bytes. Each byte is a number 0 to 255. This is ASCII: 65 is "A", 32 is space, 10 is newline. The current byte cell is highlighted purple.
Want this level of visibility into your models?
Veritate turns generation into something you can inspect, byte by byte. If you want interpretability like this for your own models, talk to us.
Talk to us