Conjecture 6.3: Skepticism Annealing

EXPLORED Skepticism is most beneficial early in learning and should be annealed (reduced) over time.

Statement

In a system of belief agents with misaligned desires, the skeptical agent (whose desire partially contradicts evidence) provides the greatest benefit during early adaptation and should be gradually removed via an annealing schedule as the system converges.

Status: Explored

Explored in the exploratory desire-as-regulariser runs. In those runs the sustained skeptic records the lowest calibration error at each horizon tested and the annealed skeptic sits between the sustained and absent cases. The runs cover a few horizons and one belief model, so the conjecture remains open.

Evidence Summary

The experiment exp_skeptical_annealing.sx tested three strategies:

Full skeptic (constant misalignment): lowest calibration error at each horizon tested
Annealed skeptic (gradually reduced misalignment): intermediate performance
No skeptic (aligned desires only): highest calibration error

One reading of the annealed case is that its intermediate error comes from the diversity introduced during the transition rather than from the reduction itself. The companion reading is Conjecture 6.5 (sustained skepticism), also explored and open.

Relevant Experiments

exp_skeptical_annealing.sx — direct test of annealing vs sustained skepticism
exp_anima_deep.sx — deep belief improvement from skeptical desire
exp_anima_correlated.sx — desire regularisation across correlated streams

What This Means

The conjecture treats skepticism as a form of exploration to be reduced once the system has learned enough. The runs to date do not show a horizon at which reducing it helps, but they test one belief model at three horizons. Whether annealing is ever preferable, and what mechanism separates the three strategies, are open questions.