The score that used to lie by accident

CreatorOS's Health Score is supposed to tell you, honestly, how your channel is trending. For a while, it didn't. On channels with very little view history — brand new channels, or an unusually quiet week — it would still confidently report "Excellent" or "At Risk." In one case it displayed a growth rate of literal Infinity percent.

The bug was almost embarrassingly simple once we found it. The score is partly built from a week-over-week percentage growth calculation: this week's average views divided by last week's, minus one. If last week's average was zero — completely plausible for a new or dormant channel — that division either produces `NaN` (0 divided by 0) or `Infinity` (anything divided by 0). Either one can silently trip a scoring branch that was never designed to see a number like that.

Why this mattered more than a normal bug

CreatorOS's whole pitch is that it won't pretend a guess is a fact. Every insight in the app — Health Score, Growth Coach, and now this blog's own topic pages — carries a label: Fact, Pattern, Recommendation, or Hypothesis, plus a confidence level of High, Medium, or Exploratory. A tool that says that about itself and then quietly shows "Infinity% growth" as a confident verdict isn't just buggy, it's undermining the one thing it's supposed to be trustworthy about.

The fix was a shared minimum-average-views floor, used by both the Health Score and Growth Coach, that gates the percentage calculation entirely. Below that floor, there simply isn't enough real view volume for a percentage to mean anything — so instead of a number pulled from noise, both features now report an honest "Insufficient Data" state. We added that as a first-class value on the Health Score's label field rather than shoehorning it into an existing category, because it's a genuinely different kind of statement: not "your channel is struggling," but "we don't have enough signal yet to tell you anything real."

It's a small fix in terms of lines of code. But it's the kind of bug that, left alone, quietly breaks trust one confident-looking wrong number at a time — which for a product built on "honest AI" as a core differentiator, is about as bad as bugs get.

Want to see what we’ve actually built?