Holland bucket-brigade credit assignment

⚠️ Curated Limitations & Negative Results

• The supplier-strength display on p. 2 ends with (1-c*r(C)/n(C,t))*s(C,t), which is generally inconsistent with its preceding additive expression s(C1,t)+b(C,t)/n(C,t) and the supplier-payment prose. The additive statement is used here and the conflicting printed expression is preserved explicitly. No official erratum, silently repaired full specification or author-implementation behavior is asserted.
• Specificity is a raw count, and describing c as small does not ensure c*r(C)<=1. For the isolated debit, nonnegative strength is preserved under nonnegative starting strength and 0<=c*r(C)<=1; adding nonnegative payments preserves that elementary boundary. These are explicit algebraic conditions, not a source-wide invariant for arbitrary specificity, signed rewards or simultaneous transactions. The equal-share formula also requires n(C,t)>0; the selected description does not resolve every zero-supplier case.
• The article describes the mechanism rather than complete executable pseudocode. It does not fully specify winner cardinality, ties, initial strengths, exploration, no-supplier behavior, overlapping or self-supplier transactions, or a complete order for concurrent debits and credits. These choices cannot be reconstructed as one mandatory implementation from the stated local equations.
• Credit recipients are message suppliers whose outputs satisfy a consumer’s conditions. This is not automatically credit to every temporally preceding rule, a generic eligibility trace, or proof that every paid supplier was counterfactually necessary for the environmental outcome. Restricted temporal-difference analogies do not identify the whole rule-credit procedure with TD, Q-learning or Sarsa.
• The parasite and default/exception hierarchy discussions are conditional illustrations, not general convergence, monotone-improvement or extinction theorems. A more specific exception can still bid less than a stronger default because the bid depends on both factors. The source leaves simultaneous parasite and useful stage-setting activity outside its simple illustration.
• The bridge example requires suitable persistent activation, a message link from the initiator, later payoff and subsequent reuse. It does not guarantee one-shot learning for arbitrary rule sets, immediate propagation to every earlier rule, or automatic discovery of the bridge. Genetic rule discovery and construction of further hierarchy components are additional mechanisms.
• The declared evidence is the inspected seven-page 1985 chapter. The historical discovery source cites a 1986 Holland chapter that was not obtained in this research, and a later book draft’s simplified Sarsa comparison points to unread Wilson 1994 details. Those later forms are not exact aliases of the selected version. No first-use priority, universal benchmark advantage, full later-version equivalence, author-code inspection or experiment reproduction is claimed.