Reading probabilistic rows after ATT without laundering them
Probabilistic modelling did not vanish when ATT arrived. It changed costume. The reporting failure is not that models exist. It is that modelled rows quietly sit in the same total as observed clicks.
After ATT, many MMPs still emit rows that are inferred rather than observed. Device graphs, IP-plus-UA hints, and vendor “probabilistic” switches produce install claims that look identical to a deterministic row in a CSV. If your weekly total does not separate them, you are laundering uncertainty into a number that finance will treat as counted.
A style rule, not a philosophy
In class we require three fields on every modelled row: method name, confidence or tier if the vendor supplies one, and a boolean observed=false. The weekly memo then prints two totals: observed and modelled. You may discuss the modelled figure. You may not add it into the signed install count unless the source-of-truth policy explicitly allows a modelled band — and even then it stays a separate line.
What this is not
This is not a claim that probabilistic methods are useless. For some Android catalogues they still catch traffic that would otherwise sit in unknown. It is also not a tutorial in turning those methods back on. We do not take MMP admin access. If your vendor’s toggle is poorly documented, that belongs in the source map as a risk, not as a growth hack.
Privacy Sandbox on a different clock
Android Privacy Sandbox attribution, where it appears, reports on its own delay. Mixing those delayed rows into the same week as MMP click-attribution without a label recreates the SKAN Monday problem on another OS. Put the reporting clock in the cell. If the clock is unknown, the row is not ready for the ledger.
Module 03 of The Attribution Ledger is built around this labelling drill. Bring an export that still has a probabilistic flag, even if you think you switched it off. Students are often surprised.