Characterize the role of real prior model-generated context

Determine whether the quantity of real, model-generated prior context in a session affects how qwen3:8b, llama3.1:8b, and mistral:7b respond to rare tool-call failures, independently of the immediate pre-failure token pattern.

Background

The guaranteed-failure recovery design uses synthetic filler to avoid the computational cost of running tens of thousands of live trials at the rarest failure rates. Although the design preserves a block of literal repeated success tokens immediately before the failure, it places the model in a shorter session containing fewer genuinely model-generated trials than the true-random design.

Consequently, the recovery design cannot determine whether the amount of real prior self-generated context, as distinct from the literal content immediately preceding the failure, changes the model’s response. Resolving this question would require a design that varies or matches the quantity of real prior model-generated context.

References

Whether the sheer quantity of real prior self-generated context (as opposed to its content) affects how a model responds to a failure is a question this design does not resolve, because it was not built to resolve it; it was built to make the rate-of-failure and the immediate pre-failure context comparable, and it succeeds at that narrower goal.

— Explanatory Engagement Under Rare Anomalous Failure: Asymptotic Rarity in Model Behavior (or: The Asymptotic AI)  (2608.13063 - Mao, 13 Aug 2026) in Section 3.7, Phase A.1: Guaranteed-Failure Recovery Design

All figures in Section~\ref{sec:results} were taken before the cap was applied. The per-model and per-surface rates should therefore be read as a lower bound with unknown per-model bias, since large-context models were penalised more than small ones. Re-reading the corpus under a fixed cap is the first item of future work.

— What Stops a Small Language Model From Driving a Database Agent  (2609.21341 - Bozoglan et al., 18 Sep 2026) in Section 7.1, “Unbounded serving context is a memory confound, and we hit it”