The loop

Every topic in a config has a weight you set: how often it should come up relative to the others. The loop adjusts how that weight actually plays out, based on how each topic has really performed for that specific account.

How the number is made

multiplier = topic's average engagement rate ÷ account's average engagement rate

Engagement rate is (likes + replies + reposts) ÷ impressions, not raw like counts. A topic that gets fewer likes but reaches far fewer people can still be your best-performing one relative to its own reach. It's scored per account, because what counts as good engagement for one account says nothing about another.

Actually used (the number that really controls how often a topic is picked) is simply your weight multiplied by this multiplier. The loop never edits your column; it only ever scales it.

Why it can't run away

Two guardrails keep this from spiraling:

• The multiplier is locked between 0.5× and 2.0×. A topic can never be silenced entirely by a bad run, and one lucky post can never take over the whole rotation.
• Below 3 posts, a topic's multiplier stays neutral and says so explicitly in the UI: there isn't enough data yet to trust it either way.

What it's deliberately not allowed to do

The loop scores which topics get written about more. It does not, and will not, touch how things are written. Optimizing prose style directly against engagement has a documented failure mode: a model tuned that way tends to make its own output more extreme to chase attention (arXiv:2402.06627, ICML 2024). So voice only ever changes from a human edit or rejection in the review queue, never from a like count.

Competitors

The same page also tracks competitor accounts: X handles you're not managing but want visibility into. Add a handle, and Pegacraft scrapes its recent posts daily using one of your own connected sessions, surfacing average engagement so you have an outside reference point, not just your own history, when judging whether a topic is actually working.