I am digging deeper, it's what rabbits do. Claude's numbers check out against the code: a reply is 5 vs 0.66 for a like (about 7.5x), a copy link is 20 (about 30x), "not interested" is about -52 (about 79 likes) and a mute is -58.8 (about 89). The mutual follow boost, the 2-hour test slot for small accounts and the default settings caveat are all right too. There are two things I'd adjust: • "Worth 30 likes" doesn't mean one share equals 30 likes. The weights multiply the model's prediction of how likely each viewer is to do each thing. So a post people are likely to share ranks like one that gets 30 times more likes, but you copying a link once doesn't add 30 likes' worth to that post. That's why sharing your own posts doesn't move them. • Saying hashtags do nothing goes a little too far. I found no ranking weight for them. They do feed X's topic models and spam detection, though, so stuffing posts with them could count against you. Claude's other points match what I found. Links have no penalty in the code, and "answer your repliers" fits the mutual boost, though that part is advice drawn from the code rather than something it states.

It switched to 12 parallel workers. Full runs went from timing out at 15 min to 5–7 min. Then it sampled CPU and found the box never passed half load, because a few slow tests leave the other workers idle at the end. Fixing that next.