Which centipawn calculation algorithm is Lichess currently using?

Lichess page about accuracy and centipawns calculate win rate% using:
Win% = 50 + 50 * (2 / (1 + exp(-0.00368208 * centipawns)) - 1)
"
https://lichess.org/page/accuracy
"

While official stockfish document presents a much different and more steep curve, where 1.0 pawn advantage being a 0.5 (that is 50%) win probability ,and almost 0.5 draw probability, and almost no lose at all.
"
https://lichess.org/faq#acpl
https://official-stockfish.github.io/docs/stockfish-wiki/Stockfish-FAQ.html
"

Since both pages seem to be long not up-to-date with rushing generations of Stockfish, I really want to make clear which algorithm is still valid now.

What defination of centipawn is Lichess using now (to calculate ACPL, accuracy, win rate etc.), and does it have the same meaning with the centipawn analysis result given by SF16, SF17 or Fairy-SF14 (as they are mostly default for "board editor/analysis" panel)?

Stockfish's curve is generated from Stockfish playing against Stockfish, while the Lichess curve is from Lichess games. I'd assume both are still largely valid, just for different purposes.

@PurpleInferno said in #2:

Stockfish's curve is generated from Stockfish playing against Stockfish, while the Lichess curve is from Lichess games. I'd assume both are still largely valid, just for different purposes.

But the chances will be quite small that these centipawns coincidence to be the same? Or are they not the same centipawn?

Stockfish outputs centipawns, so the centipawns themselves are the same. They just mean different things for human win probability compared to Stockfish win probability.

Thank you, that does make sense, that centipawns have different interpretation