Myth ledger: "the structural credit-assignment problem comes from Minsky (1961)"

credit-assignmentstructural-credit-assignmentminskysuttonterminologyhistory-of-mlcitation-practice

The circulating claim. A 2024 PLOS Computational Biology paper states "the structural credit assignment problem arises when the causal structure between actions and subsequent outcomes is hidden from direct observation," and cites Minsky's 1961 "Steps Toward Artificial Intelligence" as its source.

What the primary actually says. Minsky's 1961 text, read directly against the MIT-hosted full primary (claim-minsky-1961-named-credit-assignment), frames credit assignment temporally: "If the run is successful, how can we assign credit for the success among the multitude of decisions?" — a sequential-decision problem. The word "structural" as a qualifier does not appear in that passage. Attaching "structural" to Minsky 1961 is therefore a retrospective relabel, not a quotation of his framing.

Status: contested. This is the same citation-practice pattern the vault already documents for "spatial credit assignment (Werbos 1974)" (claim-werbos-1974-no-credit-assignment-language): a later term bolted onto an earlier text that never used it. The paired temporal-vs-structural vocabulary was standard, citable field language no later than 2004 (Agogino & Tumer, "Unifying Temporal and Structural Credit Assignment Problems," AAMAS 2004, NASA Ames — Tier 1, "Multi-agent systems have the structural credit assignment problem of determining the contributions of a particular agent to a common task"). Field lore attributes the temporal/structural distinction to Richard Sutton's 1984 PhD dissertation, but that attribution is itself unverified against the primary (see watch-flag) — so "who first wrote 'structural credit assignment'" remains open even as the Minsky-1961 attribution is contestable.

Watch-flag / receipts. Verify against Sutton's 1984 dissertation whether the literal phrase appears there: question-verify-sutton-1984-structural-credit-assignment. The "structural" sense — distributing error across a network's weights — is what backpropagation mechanically solves; see moc-backpropagation-origins.

• 2026-09-01: the Sutton half of the watch-flag is now substantially narrowed, not yet closed. Sutton's own dissertation abstract, his own later retrospective summary, and his earliest personally-dated use of the literal phrase (a 2023 talk, framed as unsolved) all converge against a 1984 coinage — claim-sutton-1984-dissertation-abstract-omits-structural-credit-assignment, claim-sutton-retrospective-summary-names-temporal-credit-assignment-not-structural, claim-sutton-first-personal-use-of-structural-credit-assignment-dates-to-2023. The dissertation's 223-page body text remains unread (tooling-blocked), so this stays contested rather than debunked.