# When the Next Question Has a History

At a call-level interface, answers arrive separately. In an adaptive collection,
each answer may change which question is worth asking next.

A collector begins with a working map of the capabilities it wants. It sends
probes, inspects what came back, and updates the map. If one tool task succeeds
and a nearby variant fails, the next query can hold the task fixed while
varying permissions, context, or tool order.

A fixed batch can target the same capabilities. Adaptation requires the later
choice to depend on the earlier result. That dependency is the proposed source
of additional value.

The selected outputs become training data only when a learning procedure uses
them in a model update. The set becomes a curriculum only if training orders,
weights, or schedules its examples.

I can inspect one answer and still miss the adaptive signal. It lives in why
the next question was chosen.

The [All-In discussion of Kimi
K3](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=493s) moved among output
training, [benchmarking](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=931s),
and [copied weights](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1042s).
Those are adjacent objects. None establishes how later queries were selected.

Anthropic's [February
report](https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks)
attributes more than 3.4 million Claude exchanges to Moonshot through hundreds
of accounts it calls fraudulent and multiple access pathways. Across the
report's three cases, Anthropic says scale, coordinated accounts, narrow
capability targets, and evasion can distinguish extraction when one prompt
appears benign. Those indicators support Anthropic's campaign classification.
The report does not disclose an adaptive selection policy or measure
selection's contribution to a student.

One controlled test would match teacher-query budget, student initialization,
base data, training compute, and evaluation suite. An adaptive selector would
face a prespecified baseline across repeated runs and preregistered
capabilities. If the adaptive arm showed no reproducible gain, the experiment
would find no additional value from selection under those conditions.

An answer can stand alone.

When selection is adaptive, the next question carries the history of the set.