The primary model or rule says long or short. The meta model is trained only on the binary question of whether that call was right, using features about context: volatility, liquidity, time of day, recent performance. Its output becomes a confidence score that drives size.
This separation helps because direction is the hardest thing to predict and confidence is easier. It also keeps the economic logic in the primary rule, so the machine learning is confined to a narrower, more auditable job.
The caveat is sample. The meta model only sees observations where the primary model fired, so a rule with 400 historical signals gives you a 400-row training set, which is far too small for anything flexible. Meta-labelling suits high-frequency-of-signal systems, not rare setups.
Related: label, triple-barrier-method, position-sizing, sample-size