Game theory is a normative theory of optimal play given a stated game structure. It is not, and was not designed to be, a descriptive theory of actual human strategic behavior. The famous solution concepts (Nash bargaining, Rubinstein, iterated-PD cooperation, etc.) require importing assumptions — BATNA stability, discount factor, unknown end, axiomatic restrictions — that aren’t derived from the strategic situation. Those imports are behavioral, institutional, or psychological facts the model imports to make the math close. Where the imports don’t hold, the predictions don’t hold. Where the predictions don’t match observation (one-shot cooperation, real bargaining, the Axelrod TFT result), the imports are doing the work — and what actually drives behavior is norms guiding action, not utility maximization. The implication for the vault’s trade-theory work is that the transfer problem is hard not because we lack a clever algorithm, but because the bargaining situation has no game-theoretic answer without psychology-level imports.
Links: Nash Bargaining Problem, Multiplayer Coalition Problem, Bilateral Trade Valuation, Frontier Trade Theory, LLMs as Praxeological Actors, Planner-LM Composites, Hayek vs Mises: The Calculation Problem, Catan 47k Empirical, Coconut Island & Manufactured Option Space (a debate-sparked thesis that applies this note’s “norms guide action” point to manufactured-scarcity rhetoric)
Game theory produces clean solutions only when the strategic structure is augmented with imports the strategic structure doesn’t itself provide:
| Famous result | Strategic structure | Import doing the work |
|---|---|---|
| Nash bargaining solution | Two players, $100, agree on a split or get nothing | Cooperative axioms (symmetry, IIA) |
| Rubinstein bargaining | Alternating offers, two players | Discount factor δ < 1 imported from outside |
| Iterated PD cooperation | Iterated game | “Unknown end” — finite-end games backward-induct to defection |
| Folk Theorem | Infinitely repeated game | “Any individually-rational payoff is sustainable” — produces a continuum, doesn’t select |
| Schelling focal points | Multiple equilibria | Explicitly psychological — round numbers, salience, convention |
The imports are sometimes presented as parts of the model (especially the discount factor), but they originate outside the strategic situation:
Strip them, and the bare game has either no unique solution (continuum of equilibria) or a different solution than the famous result.
The cleanest demonstration. Two players, $100, propose splits.
The symmetry argument ([[bilateral-trade-valuation]] §”asymmetric-externality reading”): if Player 1 and Player 2 are internally symmetric (same time-preference, same alternatives, same information, same utility-over-money), the game itself contains no information distinguishing them. Therefore the bare game cannot produce an outcome distinguishing them. Any asymmetric outcome requires importing asymmetric external factors.
The axis-points (100, 0) and (0, 100) are correctly excluded — those aren’t bargaining outcomes, they’re walk-away states equivalent to (0, 0). The interior is the bargaining space, and the bare game says nothing about which interior point gets played.
A perfect illustration that the “outside” of a game is itself manipulable:
Mr. Beast invents games like “whoever keeps their hand on X the longest keeps X.” Player 2’s BATNA is “leave when tired” — a utility threshold at exhaustion. Standard bargaining theory treats this as a fixed parameter of the situation.
Now Mr. Beast adds: “last one to leave gets $10k.”
He hasn’t changed the rules of the hand-on-X game. He’s rewritten Player 2’s BATNA from outside. The strategic situation flipped because the outside changed.
This is the canonical demonstration that the boundary of “the game” is arbitrary. If the outside is itself contestable — and in real strategic situations it almost always is, via information manipulation, manufactured time pressure, framing, third-party offers — then the bargaining-theory practice of importing BATNA as a fixed parameter is foundationally wrong. Real negotiators manipulate the counterparty’s BATNA; that manipulation IS the strategic move. Standard bargaining theory cuts the boundary of “the game” too narrow and assumes a stable outside; the stable outside is itself manipulable.
There’s a small literature trying to fix this — Aumann’s correlated equilibria, mechanism design with strategic rule-setters — but mainstream bargaining theory still treats BATNA as exogenous data.
Backward induction in a known-finite-end iterated Prisoner’s Dilemma:
So the theoretical prediction in any fixed-end iterated PD is universal defection.
Empirically, this doesn’t happen. Axelrod’s tournaments (1980, 1981) used a fixed-length 200-round iterated PD. Strategies were submitted by game theorists, economists, etc. Tit-for-tat won both tournaments by a clear margin. TFT cooperates on round 1, then copies opponent’s last move. It’s nice (never defects first), retaliatory (punishes immediately), forgiving (returns to cooperation after one round), clear (easy to recognize). Always-defect strategies did poorly because they got punished by TFT-like opponents and lost the cooperation surplus.
This is a clean empirical refutation of the backward-induction prediction. In a strict game-theoretic setting (known-finite PD), the predicted equilibrium (universal defection) lost to a strategy theory said shouldn’t win. The tournament was the test; theory failed it.
More recent: Press-Dyson 2012 introduced zero-determinant strategies — provably-extracting strategies that can force any linear combination of payoffs from an adapting opponent. ZD strategies are the “AlphaGo of iterated PD.” But against other ZD strategies, they break down. The race-to-the-top of strategic sophistication doesn’t converge on always-defect; it converges on cleverer cooperation/exploitation balances. Sophistication doesn’t validate the backward-induction prediction; it complicates it further.
A natural first guess at why theory mispredicts the tournament: the extensive form can’t represent memory / history-dependent play, so it can’t “see” the winning strategies. That guess is wrong, and locating the error precisely is the whole payoff.
The extensive form represents memory perfectly well. A strategy in the repeated game is by definition a function from history to action — Tit-for-Tat (“cooperate, then copy the opponent’s last move”), Grim, and Pavlov are all completely well-defined extensive-form strategies. They live in the strategy space the entire time. Backward induction doesn’t fail to see TFT; it sees it and rejects it. So the gap is not representational — it is in the solution concept (backward induction / subgame perfection), one level up from the model of the game.
The combinatorics, exactly (the counts are the argument). For a fixed 10-round game:
A memory strategy isn’t a path; it’s the rule that selects which path you land on, and which path that is depends on the opponent (hence field-dependence). The key consequence: in the fixed-horizon game the strategy space, though astronomically large, is finite, and every history-dependent heuristic is already one of its points — TFT/Grim/Pavlov are specific elements. The extensive form is descriptively complete here; it isn’t missing the winners.
Unboundedness enters only when the horizon does. Drop the known end and a strategy becomes a function on arbitrarily-long histories — genuinely infinite (uncountable for arbitrary functions; countable for computable ones). That is the same unknown-horizon regime where cooperation becomes sustainable (Folk Theorem), so the unboundedness and the interesting cooperation are the same far side of the finite-horizon assumption. The extensive form can represent this case but cannot select within it — the failure is selection, not representation.
The hardness is real and P=NP-flavored (not literally P=NP): verification is cheap (simulate a match, O(rounds)), search is hard. Computing a Nash equilibrium is PPAD-complete (Daskalakis–Goldberg–Papadimitriou 2009); equilibrium refinements are NP-hard (Gilboa–Zemel 1989). If the equilibrium is intractable to compute, real agents aren’t computing it either — the computational hardness and the empirical failure are one fact seen twice, and both point at heuristics.
The deeper resolution: backward induction and the tournament are answering different optimization problems.
All-defect wins A and loses B. It is a best response to itself, but not to a field containing forgiving cooperators — against TFT it eats the sucker’s payoff once and then forgoes the entire cooperation surplus forever. So “rational” (equilibrium sense) and “winning” (ecological sense) are genuinely different objects. Classical game theory answers A and then gets blamed for failing B — but it was never the same question. Minmax is the same story: it optimizes against a worst-case adversary (so it picks defect), but the tournament field is fixed and non-adversarial, not worst-case.
The signature that separates A from B: change the field and the B-winner changes; the A-answer never moves. TFT wins Axelrod only because the zoo contains enough cooperators to reward it — drop it alone into an all-defect population and it can’t invade (it needs to arrive in a cluster). Equilibrium is field-independent by construction; an evolutionary answer is field-dependent by construction. That field-dependence is precisely what an equilibrium concept cannot express — and the cleanest proof that the tournament is a Question-B object.
This relocates — but does not yet dissolve — the two original cracks: equilibrium as a pointer, not a result (a flashlight over the stable set, with selection imported from outside), and the dynamic player (identity/strategy formed in the play, the “memory” intuition). The field that bridges A and B is evolutionary game theory — replicator dynamics, best-response-to-distribution — where TFT’s victory is a theorem, not an anomaly. That EGT is where this should go (cf. the evolutionary-capitalist critique, which gets the population-dynamics point but inflates it into objective-value teleology — right tool, wrong metaphysics).
Why EGT dodges the enumeration wall: it never optimizes over the (unbounded) strategy space. It asks the local question “can a mutant invade this population?” — a stability check, not a global search — so it sidesteps the intractability above. ESS is a filter on a given population, and it genuinely prunes (not every Nash equilibrium is an ESS).
The most promising selection lever — bounded complexity. Model strategies as finite automata with ≤ k states (Rubinstein 1986; Abreu–Rubinstein 1988): now the space is finite and parameterized by k (TFT = 2 states, always-defect = 1), and you can enumerate. Crucially, adding even a tiny lexicographic cost for complexity collapses the equilibrium set toward cooperation — a real selection principle, and a partial answer to crack (3). It selects within a population rather than depending on which population you started from, which is exactly EGT’s weak spot: replicator outcomes depend on initial conditions, and ESS prunes but does not uniquely select. So the standing guess: EGT refines (3) (and complexity-cost is the sharpest known lever); whether it closes (3) or just relocates the indeterminacy into “which initial population” is the open test. Parked here 2026-06-27 as deferred future work.
Moral application (2026-07-05). This same finite-horizon → default-defect result is the formal core of the Ring of Gyges / “might makes reality” problem in ethics: at the 1-person level always-defect is correct and default, so cooperative morality is the constructed structure that escapes it, and its grounding must be population/durability (EGT), never single-agent rationality. Worked through in Morality — Open Problems #4 and the Sitch/Mullally debate review (incl. God-as-infinite-iterator as the shadow-extension device, and the self-reprogramming defector as the residual edge).
Even in one-shot Prisoner’s Dilemma — no iteration, no future — humans cooperate at 30-50% rates in lab experiments. Theory predicts 0%.
Why? Behavioral game theory (Camerer, Fehr, Falk, Bowles) proposes:
These are patches, not foundational fixes. They treat the theory’s failure as a parameter problem (“the utility function was wrong; add more terms”) rather than a structural problem (“the model isn’t modeling the right thing”). The patches make the model fit observed cooperation rates by tuning extra parameters — but the resulting model is no longer parsimoniously game-theoretic; it’s a behavioral-economics model that uses game-theoretic notation.
The sharper read: what actually drives cooperation is internalized norms, reputation across unrelated games, convention, focal points — and these are not properties of individual utility functions. They’re structural features of human social organization that game theory keeps trying to import as inputs rather than describing as the actual mechanism.
The cleanest framing of the empirical pattern: in real strategic interaction, norms guide behavior more than self-interest does.
Game theory keeps importing these structural facts as parameters. The deeper reading is that what game theory imports as ‘preferences’ or ‘discount factors’ are actually norms and conventions — structural features of the social context that exist prior to and independent of the individual utility calculations the theory tries to perform on top of them.
This is the praxeological point applied to strategic interaction. Mises argued that human action follows from value frameworks, not from utility maximization. Game theory’s empirical failures are exactly where it tries to derive action from optimization rather than describing how value frameworks (norms, conventions, institutions) actually structure behavior.
The vault’s transfer-problem checklist has six rows. Rows 4-6 (counterparty modeling, 3+-player asymmetric bargaining, multi-step trade chains) are the open ones. They’re open in the vault because they’re open in the field, and they’re open in the field because game theory doesn’t actually solve continuous-bargaining-with-private-values.
What real Monopoly play does — and what AI play needs to do to be strong — is operate at the convention layer, not the utility-optimization layer. Strong human players use:
These aren’t game-theoretic moves; they’re convention-mediated coordination on which equilibrium gets played. A Monopoly AI that tries to solve the bargaining problem via pure utility maximization will be empirically beaten by an AI that learns to operate at the convention layer — exactly because actual gameplay terminates at conventions, not at game-theoretic equilibria.
Engineering implication: the Monopoly AI’s trade evaluator currently uses bilateral NPV-trajectory math (game-theoretic). The improvement isn’t a smarter optimizer; it’s adding a convention layer that proposes/accepts focal-point trades over optimizing trades, and that includes reputation across simulated games. The behavioral parameters Camerer would add to the utility function should instead be modeled as a separate convention component — that’s the structural fix, not the parametric patch.
This is the planner-LM composite thesis applied to bargaining: pure planning (game-theoretic optimization) doesn’t solve the problem; you need a separate layer for the convention/communication channel where actual coordination happens. Cicero needed an LM because Diplomacy is a bargaining game and bargaining is a convention game, not a utility-maximization game. The vault has been saying this in pieces; this is the unified statement.
The point isn’t that game theory is useless — it’s that it’s misframed when used as descriptive theory. Game theory is excellent as:
The vault’s bilateral trade valuation, frontier trade theory, and Pareto-filter constructions are normative game theory used well. They’re tools for “given a stated objective and stated rules, here’s the optimal path.” They don’t pretend to predict what actual humans will do; they prescribe what a rational player should do.
The mistake worth avoiding: don’t use these tools to predict behavior. Use them to plan moves, to filter the decision space, and to identify where the convention layer kicks in. The convention layer is where actual gameplay happens, and it doesn’t belong inside the game-theory toolkit — it belongs in the planner-LM composite as a separate component.
The clearest framing of where bargaining theory sits in economics, recorded 2026-05-18 from the trade-theory discussion:
Free markets are good because they incentivize profit creation. But they don’t explain how the profit gets divided. That’s all bargaining.
Most market-economics scholarship — Smith, Ricardo, the Austrians (Mises, Hayek, Rothbard), neoclassical equilibrium theory — addresses why trade creates value: comparative advantage, marginal utility, gains-from-trade, division of labor. That side of economics is well-developed. The vault’s Value and Profit, Risk and Entrepreneurship, and the broader economics thread all live on the value-creation side.
The distribution side — how the gains-from-trade actually split between the trading parties — is much less developed. Standard treatments either:
The honest position: mainstream economic theory has a strong value-creation account and a weak value-distribution account. The distribution side IS the bargaining problem, and the bargaining problem is genuinely open. This isn’t a peripheral concern; it’s the open layer of microeconomics.
Disputes about value distribution are ubiquitous and politically charged:
These all reduce to “the value exists, who gets how much of it?” — i.e., bargaining. The political debates appear to be about whether markets work (they do, at value-creation); the actual disagreement is about distribution outcomes that bargaining theory doesn’t pin down. Treating “the market figured out the right division” as a value-creation argument is a category error — markets do creation, bargaining does distribution.
The vault has been doing the value-creation half cleanly (Austrian economics, comparative advantage, the trade-creates-mutual-benefit framework). The trade-theory work in bilateral-trade-valuation, frontier trade theory, and the Catan empirical analysis is the vault’s actual engagement with the distribution-side problem — at game-board scale, but the structure is the same as wage negotiation, M&A pricing, or any commercial trade.
The Monopoly/Catan trade-theory work isn’t just game analysis. It’s the bargaining-problem half of economics applied to a constrained domain where the rules are exact and the outcomes are observable. The insights — counterparty modeling, asymmetric externalities, focal-point conventions doing equilibrium selection, planner-LM composites where the LM is the convention layer — all transfer to real-economy bargaining problems where the same structure operates with messier data.
The “bargaining problem is the best unsolved research in game theory” framing is the right one. Breakthroughs here improve more than game-AI; they improve the theoretical engagement with how markets actually distribute the surplus they create.