Does a behavioural signal survive repeated self-distillation?
Controlled experiments on behavioural signal decay across recursive self-distillation.
Finding
In the toy tier the behavioural signal decays with a measurable half-life. The local-LLM tier came out negative, which marks where the result stops.
In practice
A behavioural result from a toy setting may not carry over to a real model. Test it where you deploy.
Why this question matters
If a system repeatedly learns from its own outputs, what survives? A behavioural signal gives that question something specific to track.
What I built
A toy-tier experiment and a separate local-LLM investigation, with distinct evidence boundaries.
A design choice
Before running a recursive chain on a language model, I set a pre-registered gate: one hop had to transfer the signal against neutral and shuffled controls. It did not pass, so the chain was never run. Running it anyway would have measured the decay of a signal that never arrived. The toy signal did transfer in one hop, which is why its half-life stands and the LLM tier is reported as a boundary.
Where the evidence stops
The toy finding is not a recursive LLM transfer law.