Forecasting citation needs labelled citation events observed over time. Ours is 200 traces old. So rather than ship a score we can't defend, we're publishing the design first: the candidate features, the validation protocol, and the exact corpus threshold we'll run it at. Preregistering the test is the only way the result means anything afterwards — including if the answer turns out to be that the signal is weak.
Three things make a forecaster trustworthy: the data, the features, and the test. Here's all three, committed in advance, while there is still nothing to protect.
A forecaster needs labelled citation events observed over time. What exists today is the committed panel: 200 traces across two engines, 610 Claude citations and 878 Perplexity citations over one shared 100-query set. That is a seed, not a training set. The monthly panel run is what turns it into one.
Intent class, semantic distance between query and page, source authority, freshness, and schema completeness. Naming the candidate features in advance is the point: a feature set chosen after seeing which ones happen to work is how a model gets talked into a result it did not earn.
Held-out and time-aware: train on earlier months, score later ones, never the reverse. We commit to that protocol here, before any model exists, so the evaluation cannot be redesigned once we have seen how it lands. If the first honest run says the signal is weak, that is what gets published.
There is nothing to hand you today. Leave your details and you'll hear from us when the corpus clears the threshold and the first honest evaluation is published, whatever it says.