The published code rewards predicted actions with very different force—and punishes predicted negative feedback heavily. Here is what creators can responsibly infer from that.
Published default weights
A like’s default weight is 0.5. Copy-link is 20.0. For the same change in predicted probability, that is a 40-fold difference.
View source on GitHubSkip re-reading the repository before every post. Get the signal hierarchy, practical implications, risk checklist, and system map in two private PDFs.
Evidence → implication
Each recommendation keeps the published fact separate from interpretation. None is a guaranteed growth tactic.
Copy-link has a 20.0 default weight; share via DM and quote each have 5.0; favorite has 0.5.
Open linesUseful references, surprising evidence, reusable tools, and clear ideas are more aligned with the published hierarchy than asking for likes.
Reply is 5.0. For an eligible mutual-follow original candidate, the 15.0 boost makes its effective reply weight 20.0.
Open linesA specific question, defensible position, or useful unfinished edge can invite substance. Manufactured outrage also raises negative-feedback risk.
Follow-author has a 4.0 default weight—8 times favorite’s 0.5 for the same predicted-probability change.
Open linesA coherent body of work and a clear promise of future value matter more than a one-off like request.
Repeated-author multipliers begin 1, 0.625, 0.4375, 0.34375 under the published defaults.
Open linesClosely clustered candidates from the same author can compete against a progressive discount. More posts do not translate linearly into more distribution.
The maximum post age constant is 48 hours; older candidates leave the published path before ranking.
Open linesEvergreen quality can still be valuable to people, but this open For You candidate path does not keep scoring the same post indefinitely.
Report, mute, not-interested, and block carry default weights of -234, -58.8, -43.2, and -31.2.
Open linesMisleading, repetitive, hostile, or bait-heavy work may win shallow reactions while increasing the probability of much heavier negative contributions.
Evidence boundary
Absence from this reviewed slice is not proof that a factor never matters anywhere. It is a reason not to present folklore as published code.
Before publishing