1 · The unit
The input is an absolute percentage-point change in a predicted action probability. Entering 1 means, for example, 2% → 3%, not a 1% relative increase.
X value model · sensitivity instrument
The Weight Room
Change one predicted action probability. See its direct contribution to X’s public value-model score, with every other output held still.
Frozen source record
Defaults mirrored September 25, 2026
xai-org/x-algorithm · 4c5cfe8
points
Calculation
direct score change = weight × 0.0100
Instrument
Read carefully
This is local, first-order sensitivity—not an impression forecast, an action-count equivalence, or proof that a creator can cause the selected probability change. A negative input means the predicted probability falls.
Positive-weight heads
bars scale within groupNegative-weight heads
bars scale within groupIncluded zero-weight head
Calculated independently at +1.0 percentage point.
Downstream multipliers, filters, diversity, experiments, and candidate selection are outside this instrument.
Bench notes
What this dial can—and cannot—say.
Reading —
Use this to compare the mechanical score effect of the same modeled probability change across selected actions. It can help prioritize what to investigate. It cannot tell you which creative choice will move a probability, whether a post will win a feed slot, or how many impressions follow.
For one included model head i, the instrument reports Δsᵢ = wᵢΔpᵢ while all other predicted outputs are fixed. It does not propagate the perturbation through reranking, eligibility, author multipliers, diversity constraints, top-K selection, or the distribution of competing candidates.
The machine has many levers, but this bench isolates one at a time. The length of a lever tells you how far the local gear turns—not whether your hand can reach it, nor whether the movement survives the rest of the machine.
2 · The source
The displayed weights are selected public defaults in home-mixer/params/param.rs at the frozen commit. X notes that experiments can override defaults.
3 · The boundary
Weights multiply model outputs, not observed engagement counts. “One report cancels 468 likes” is therefore not a valid reading. Scores are personalized and only one stage in distribution.