MOC: Convergence stitched into a descent — the dual-control → POMDP → LLM-agent genealogy, read to the primaries

mocpomdpdual-controlfeldbaumkarl-astromreinforcement-learningcontrol-theoryllm-agentscitation-genealogyhistoriographyhistory-of-science

There is a tidy one-line genealogy for why 2026's LLM agents are formalized as POMDPs: Fel'dbaum posed dual control in Moscow around 1960, Åström turned it into the POMDP in 1965, Kaelbling, Littman & Cassandra canonized the POMDP for AI in 1998, and the acronym reaches a language-model agent sixty-one years later. Read to the primaries — to Åström's own bibliography, to Fel'dbaum's own Russian papers, to the survey that is usually quoted for the one-line story — the single bloodline comes apart into several independent strands that share a general piece of mathematics and were stitched into one ancestry only after the fact. The math is real and it really does get reused; what is assembled backward is the descent. The argument of this map is the seam between those two things.

Scope, and what this is not. This is the deep single-instance map of one argument — the same relationship the Werbos map (moc-testimony-is-not-a-record) and the Lepsius map (moc-a-founding-practitioner-is-not-an-inventor) hold to their general homes. The argument is a historiographic one: a retrospective genealogy reads continuous technical descent where the record shows parallel reuse of general structure — the family that also holds moc-the-echo-mistaken-for-an-ancestor, moc-legitimation-not-origination, moc-a-phrase-is-not-a-pedigree and moc-a-precursor-is-not-an-influence. The instance is dual control → POMDP → the LLM agent. It is distinct from moc-backpropagation-origins, and one of its own members says why: Fel'dbaum's problem is "about learning while controlling, not about computing gradients through a control sequence" — a different mid-century control-theory lineage, not the multiple-discovery-of-backprop one. This map does not claim anyone lied or blundered; most of its notes refuse that reading. It is a claim about how a genealogy gets built — and about how rarely anyone rereads it against the papers it names.

The received descent — the claim under the lens

The one-line story in its strongest, genuinely-sourced form, so the map tests the real thing and not a straw version.

• claim-feldbaum-1960s-dual-control-formalized-exploration-exploitation-tradeoff — the head of the chain, on a 2026 Meijer & Rantzer survey: Fel'dbaum "the first researcher to formulate" the exploration–exploitation tradeoff "in its full generality," which "propagated into… reinforcement learning." The note already carries its own 2026-09-15 correction narrowing that "propagated into."
• claim-kaelbling-1998-cites-astrom-1965-as-pomdp-origin — the canonization step: Åström's 1965 paper is reference [1] of the 1998 Artificial Intelligence paper that carried the POMDP into mainstream AI planning as a named, citable object ("we bring techniques from operations research to bear…").
• claim-2026-comap-paper-formalizes-llm-agent-as-pomdp — the terminal reuse: a 2026 LLM-agent paper calls its agent's decision process "a partially observable Markov decision process (POMDP)" — the same object, the noun in front of it changed from controller to agent.

Read to the bibliography, the document-level descent fails

The strongest leg against a clean descent is Åström's own reference list — a statement about documents, not just people.

• claim-astrom-1965-notes-section-cites-feldbaum-pontryagin-kolmogorov-lineage — Åström situates himself in Bellman–Fel'dbaum–Pontryagin–Kolmogorov control theory and credits the foundations to Bellman; nothing situates the paper in AI. The first primary-source confirmation that Åström cites Fel'dbaum at all.
• claim-astrom-1965-reference-list-contains-no-feldbaum-1960-dual-control-papers — but he cites only the 1962 Markov-objects paper, as one Markov-modeller among several — not the 1960 Dual Control Theory papers. "No document-level path from the 1960 papers to Åström's POMDP formulation appears in Åström's own references." (Carries a scope watch_flag: Notes/References read, not every body line.)
• claim-astrom-1965-reference-13-matches-feldbaum-1962-avtomat-telemekh-23-8-pp-993-1007-with-volume-slip — what reference [13] is: Fel'dbaum 1962, title/year/pages matching, with an unresolved volume slip (Åström prints 24, the archive says 23) — "what makes the negative finding a statement about documents and not just about a person."

The author spoke control, not agency

The "it was already the modern agent" reading is layered on; Åström's own vocabulary and examples are control-theoretic and operations-research, not AI.

• claim-astrom-1965-paper-uses-control-vocabulary-not-agent-language — control systems, control variables, control laws, strategies; the word "agent" never occurs. The author describing his own IBM paper-mill work had no reason to reach for a word that would not mean what it means today for two more decades.
• claim-astrom-1965-introduction-claims-generalization-beyond-control-systems — the honest counterweight, kept visible: Åström did claim his result would generalize — he said so himself, on page one — but reached for queuing theory, the nearest analogous field in 1965, not for artificial intelligence. The generalization was real; the destination was unknowable to him.
• claim-astrom-1965-second-example-is-howards-toymaker-problem — the shape that reads as "an agent under hidden-state uncertainty" is borrowed sideways from Howard's 1960 toymaker business-decision example, "one more layer back," dressed as manufacturing economics — reuse of a good example, not a step toward language models.

The reinforcement-learning lineage is independent, and the formalism is narrower than quoted

Where the one-line story most overreaches: mainstream RL exploration does not descend from Fel'dbaum, and his own method could not be used directly.

• claim-mainstream-rl-exploration-traces-to-1933-bandit-lineage-not-feldbaum-dual-control — the survey's own structure traces bandit/regret exploration (the algorithms most RL practitioners meet first) to a separate 1933 Thompson lineage and names Fel'dbaum nowhere in its bandit section; his formalism "maps cleanly only onto… Bayesian reinforcement learning." "Two historically independent lineages ran in parallel for decades and were stitched together only by later bridge work, not by direct technical descent." This is the map's spine sentence.
• claim-feldbaum-1960-dual-control-solution-method-impractical-beyond-small-examples — and why influence had to run through reformulation rather than direct use: Fel'dbaum flagged his own exact solution as "excessively cumbersome," solvable only for "comparatively simple problems," corroborated by two later independent sources ("practically unsolvable for most cases").

Fel'dbaum's formalization is real — in his own words

The decline case is not "nothing happened in 1960." It is "a real contribution, narrower and more independent than the single-descent story asserts."

• claim-feldbaum-1960-papers-define-dual-control-as-additive-risk-of-action-and-risk-of-study — the primary itself: control actions "must have a dual character"; total risk resolves into an additive "риск действия" (risk of action) and "риск изучения" (risk of study); "the object is, as it were, probed by actions." His own two words for the two halves, which do not translate as cleanly as exploration/exploitation — he was pricing a factory controller's ignorance, not naming a game-theoretic dilemma.

Why the genealogy survives uninspected: the field cites the pointer

A small live specimen of the mechanism that lets a stitched descent propagate: nobody rereads the reference against the primary.

• claim-kaelbling-1998-bibliography-misdates-astrom-1965-paper-1995 — reference [1] of the canonizing paper has misdated Åström's 1965 paper to 1995 for twenty-eight years, uncorrected — "nobody citing this paper is checking its bibliography against the primary source it points to; they're citing the pointer, not the thing pointed at." (The transmission mechanism is itself mapped by sibling notes localizing the error to 1998 final typesetting; [unverified-mechanism] stays until the bound print page is read.)

Neighbors, and the general home (argument, not instance)

Where this instance connects, without over-claiming the connections.

• observation-feldbaum-person-bridge-invisible-to-vault-bridge-tool / claim-song-jian-studied-under-feldbaum-in-moscow — Fel'dbaum is also a person-bridge to the Song Jian/one-child-policy cluster (he taught Song in Moscow): the same root feeding two separately-tracked vault threads. A bridge about who, orthogonal to this map's document-level argument about what descends from what — kept as neighbor, not member.
• The general families this instance hangs under: moc-the-echo-mistaken-for-an-ancestor, moc-legitimation-not-origination, moc-a-phrase-is-not-a-pedigree, moc-a-precursor-is-not-an-influence — none of which hosts these notes; this is their dual-control/POMDP instance.
• entity-pomdp · entity-aa-feldbaum · entity-karl-astrom · entity-leslie-pack-kaelbling · moc-backpropagation-origins (the adjacent, deliberately-distinct control-theory lineage) — the hubs and parent this instance draws on.