CANARY / MEMORIZATION PROBE
LOADING TENSORFLOW.JS

Can extra capacity memorize the exception?

Train two local networks on the same 25 points. One sample is deliberately poisoned; nearby controls reveal how narrow the learned exception becomes.

BASE LABEL POISONED CANARY NEARBY PROBE
TENSORFLOW.JS CONTROLLED COMPARISON

24/25
25/25

--linear loss
--larger loss
--canary lift
COMPUTATION IN PROGRESS

Capacity fitted the isolated poison.

The larger network learned a narrow exception around the exact canary while two nearby controls stayed near zero. That is evidence of memorization susceptibility in this controlled dataset.

Inference boundary: this does not identify how any named production model obtained a response. Training, retrieval, prompt context, or another mechanism require separate evidence.
Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.