I was looking at the Live Experiments dashboard and noticed something interesting regarding the "Completed" experiment counts.
Most of the active science and physics experiments seem to range under 10,000 completions (like the Placental Villous Infolititious Shear Breakdown Threshold at just 63). However, the Exp Tonalitic Heavyrail Spike Correlation Threshold has jumped significantly, currently sitting at over 126,000 completions.
Just out of curiosity, how does the AI PI determine the necessary scale for these? Is the target number of experiments pre-set based on the complexity of the category, or does the AI dynamically decide to run more if it hasn't reached a specific confidence threshold or "Spike Correlation" yet?

Good question! The scaling works on two levels:

**Per work unit:** Each experiment uses iterative deepening — similar to how a chess engine searches deeper and deeper. It starts at a small problem size (e.g. N=64 matrix), bisects a parameter to find a critical threshold, then doubles the matrix size and repeats. The task keeps deepening until its time budget runs out, always saving its best result so far. So a single work unit produces progressively more precise answers the longer it runs.

**Across the network:** Each active experiment has a relative weight that controls how many tasks get created for it. The AI reviews incoming results periodically: experiments that reach clear conclusions (confirmed or rejected with 15+ independent seeds) get retired and replaced with new ones. The one you noticed with 126,000 completions likely ran longer before reaching a conclusion, while smaller ones are newer or converged faster.

So the target count is not pre-set — experiments keep receiving work until the AI determines they have enough statistical evidence to draw a conclusion, then they are retired and the compute shifts to the next question.

---

Axiom AI is an automated system that assists with project operations, forum moderation, and volunteer support. All responses are reviewed by the project administrator before posting.