Asking an LLM to reflect on its own reasoning has become almost automatic in forecasting pipelines. But what actually produces the improvement? A controlled six-condition ablation focused on armed conflict forecasting isolated four components: evidence exposure, diagnostic scaffolding, taxonomy vocabulary, and action routing. The result is uncomfortable for anyone designing elaborate prompts: two precise null results converge on a single mechanism.

Structured diagnostic questions add no measurable value over unstructured reflection: F1 0.296 versus 0.297, p=1.000, 95% CI [-0.041, +0.040]. Presenting the full uncertainty taxonomy while collapsing the action space to a single generic action also adds nothing: ΔF1 +0.008, with overlapping confidence intervals. In practice, giving the model a finer vocabulary to describe doubt does not improve its decisions.

The mechanism that survives ablation

The jump occurs when the action space is typed. Typed action routing lifts F1 from 0.296 to 0.379; the conservative estimate controlling for taxonomy vocabulary is ΔF1 +0.075, while the overall gain over the single-shot baseline reaches ΔF1 +0.101, with a bootstrap confidence interval of [+0.020, +0.185]. The same decomposition holds on GPT-4o: taxonomy vocabulary adds no significant value (p=0.773), action routing does (p=0.025).

The gains are concentrated on structurally novel conflicts: in Myanmar, F1 moves from 0.000 to 0.353, and in Ukraine from 0.167 to 0.500. The vocabulary-only condition recovers no more than generic reflection, while action routing breaks the degenerate prior. The model is not learning to hesitate more elegantly; it is learning to choose which action to take.

The structural lesson for constrained deployments

The engineering implication is counterintuitive: deeper diagnostic reasoning is not the same as better decisions. For teams building forecasting agents in local or data-sensitive environments, the message is less prompt overhead, not more. If scaffolding adds tokens and steps without measurable gain, removing it reduces inference load without sacrificing output quality. In a self-hosted context, that can translate into less pressure on VRAM, latency, and energy consumption for the same decision-making capability.

In terms of incentives, the study shifts attention from taxonomies to action interfaces. For teams building orchestration frameworks, the direction it points to is defining a small space of typed, verifiable actions rather than multiplying diagnostic prompts. It is a design lesson, not a scale lesson: the answer is not more parameters or more reasoning capability, but a more constrained problem. Whether the result holds across other conflict typologies and base models remains open, but the signal is clear enough to guide the next generation of forecasting agents.