Realized vs unrealized
Why the board ranks on closed positions, and what that costs.
Almost every wallet leaderboard in this category ranks on unrealized PnL. Ours does not, and this is the decision that most changes what the board says.
The problem with paper
A wallet is up 400% on a position it still holds. It sits at the top of the leaderboard. People copy it.
That wallet has won nothing. Its gains exist only as long as somebody is willing to buy at the current price, and the moment it tries to convert them into money it becomes the seller into whatever bid remains. A leaderboard ranked on unrealized PnL is, quite reliably, a ranked list of the next exit liquidity.
The gains are also trivially manufactured. A thinly traded token can be walked up by its own holder at almost no cost while gas is subsidized. Unrealized PnL on such a position is a number the wallet chose for itself.
Closed positions with their cost basis are harder to compute and they are the only number that survives the exit.
What realized means here, exactly
A wallet's realized PnL only ever moves against inventory the engine watched being bought.
That sentence has a consequence people hit immediately. If a wallet sells a token the engine never saw it acquire, that sale is recorded as activity and it is not booked as profit. Not as a gain, not as a loss, not as zero. It stays out of the number.
At first glance that looks like an obvious undercount. It is, and it is the right direction to be wrong in.
The alternative is assuming a cost basis for a position you never saw opened, which usually means assuming zero. Assume zero and every sell of an unwatched token becomes pure profit, which turns anyone who transferred tokens in from another wallet into a genius, and turns airdrop farmers into the best traders on the chain. The undercount is obvious and boring. The overcount is silent and flattering, which makes it far more dangerous on a screen people use to decide who to follow.
The rules underneath
Four of them, and each is pinned by a test that was proven to bite before it was trusted.
Cost is a weighted average. Buy the same token three times at three prices and the basis is the average weighted by size. Positions are held per token.
Proceeds are prorated to matched size. Sell more than the engine watched you buy and only the matched portion realizes. The rest is activity. Paying out the full proceeds against a partial basis is the specific bug that would inflate every partial exit on the board.
An unmatched sell is never profit. Covered above, and it is the promise the whole ledger rests on.
Unknown stays null. Unrealized PnL is not computed at all, because marking an open position needs a live price for every token in it including ones minted an hour ago. A dash there means unknown.
How this was verified
The engine ships thirty four tests that run on the production box rather than only on a laptop, and they cover the ledger's promises rather than its implementation.
The part that makes them worth anything is that each one was checked against a deliberately broken engine first. Paying out proceeds unprorated, booking an unmatched sell as profit, and dropping the concentration rule from cluster detection each fail exactly one test, and the right one.
A test that has only ever passed is not evidence. It is a comment with a green tick next to it.
What this costs us
Honesty about the method includes honesty about its price.
The board undercounts. Wallets that were trading before the engine started watching carry less realized profit than they earned. This falls as the ledger accumulates, and the ledger now survives restarts, but a wallet's first appearance is always its worst representation.
It cannot rank the patient. A wallet that bought once and has held for three months has realized nothing and does not appear at all. That is not the same as a wallet doing nothing, and the board cannot currently tell you the difference.
Windows matter and are labelled. The board reports over the window the ledger actually holds, and it says which window that is rather than implying all of history.
Where the number appears
The wallet board on the terminal ranks by it in three modes. The Board covers those.
Autopsy computes the same class of number for a single address on demand, walking that wallet's history rather than reading the engine's window, which is why the two can differ for a wallet that traded before the engine was looking.
There is no NFT equivalent yet. When it lands it will be sharper than the token version rather than weaker, because a tokenId identifies exactly which item was bought and sold and no cost basis has to be averaged.