Cooperation Failure: A Taxonomy Audit

Evidence: [F] sourced fact (source mandatory) · [B] claim (originator named, or flagged as our own thesis) · [I] inference (names the [F]/[B] it follows from).

Unless noted otherwise, every [F] below was verified against the primary text, not a secondary summary. Where only a secondary source was obtainable, the statement carries [B] and names who is being relied on.

Key findings

• [F] Fearon’s canonical statement of the bargaining puzzle names two mechanisms, and neither is “coordination”: “essentially two mechanisms … (1) the combination of private information about resolve or capability and incentives to misrepresent these, and (2) states’ inability, in specific circumstances, to commit to uphold a deal” (Fearon 1995, p. 409).
• [F] The AI-cooperation literature adopts the same axis and says so explicitly: “Prominent scholarship even argues that cooperation failure between rational agents requires either informational problems or commitment problems [Fea95, Pow06]” (Dafoe et al. 2020, §4.3).
• [F] A primary source separates three families in one sentence, which our page collapses into two: coordination failure “is not due to conflicting objectives as in ‘prisoner’s dilemma’ games or to asymmetric information as in ‘moral hazard’ games. Rather, coordination failure results from strategic uncertainty” (Van Huyck, Battalio & Beil 1990, p. 244).
• [F] The most-cited construction of secrecy-preserving mechanism implementation explicitly refuses to change any economic property: the authors “regard this loss of privacy … as inherent, and do not wish to rectify or modify this — or any other — property of a normal-form mechanism”, and require that the implementation “always enjoy economic efficiency to exactly the same degree (whatever it may be) as the abstract mechanism” (Izmalkov, Lepinski & Micali, Perfect Implementation, §1.1 p. 2).
• [F] Where MPC has been shown to circumvent an impossibility, the budget-type bound — the family Myerson–Satterthwaite belongs to — survived: “the MPC-assisted model does not help us circumvent the zero miner revenue lower bound” (Shi, Chung & Wu, ITCS 2023, §1.1.2).
• [F] The gap between “the computation was correct” and “the inputs were true” now has a canonical citation: secure multiparty computation “typically do[es] not specify how these inputs are chosen” (IACR ePrint 2022/1648).
• [B] Our own thesis, and the single most useful output of this round: CCTI does not take any of Myerson–Satterthwaite’s four published escape routes. It attempts a fifth one that lies outside the theorem’s setting — making types certifiable — and there is an existing formal literature on exactly that (mechanism design with evidence) which we have not read and which will decide whether the claim survives contact with an economist.
• [F] Two objections currently on the wiki page do not have the status we gave them: the standard-setting-as-distributive-conflict objection is a real, modelled literature (Farrell & Simcoe 2012 model standards wars as a war of attrition), while the Ostrom-monitoring objection is not a named controversy in the literature at all.

Analysis

1. The axis the fields actually converge on

Four literatures that do not cite each other arrive at the same two-way cut, and it is not ours.

[F] In international relations, Fearon’s Rationalist Explanations for War reduces the entire puzzle of costly conflict between rational actors to two mechanisms: private information combined with incentives to misrepresent it, and the inability to commit to uphold a deal (Fearon 1995, p. 409). [F] Powell’s follow-up hardens the second half against the leading alternative: “Bargaining indivisibilities do not solve the inefficiency puzzle by rendering it moot … all three are fundamentally commitment problems” (Powell 2006, p. 176). [F] Dafoe et al.’s research agenda for cooperative AI imports the same distinction verbatim, with the Fearon and Powell citations attached (Dafoe et al. 2020, §4.3). [F] Contract theory keeps the two apart with unusual precision: Grossman & Hart derive the hold-up problem while stating “Since the parties are assumed to have symmetric information …” (Grossman & Hart 1986, p. 696), and Hart & Moore note that “the x_i are observable even though they are not verifiable” (Hart & Moore 1990, fn. 6).

[I] Follows from the four statements above: the hold-up problem is a pure commitment failure with no information asymmetry between the parties. That is a stronger claim than our page makes anywhere, and it is the cleanest available proof that “information asymmetry” cannot be the common mechanism underneath everything — our page’s layering block asserts exactly that (“Common mechanism: private information that cannot be made credible”), and contract theory falsifies it with a worked counterexample.

[F] Fearon does use the word “coordination” — once, at p. 404, for the stag-hunt reading of preemptive war — and dismisses it: it “would seem easily resolved by communication”, and “it seems farfetched to think that small numbers of states would have trouble reaching the efficient solution here, if coordination were really the only problem” (Fearon 1995, p. 404).

[I] From the Fearon p. 404 passage: an economist or political scientist reading our page will not merely find the coordination/bargaining split unfamiliar — they have a canonical citation for treating coordination failure as the easy case, resolvable by cheap talk. Any dossier of ours that presents coordination and bargaining as symmetric siblings must answer that citation directly. This is the single most likely point of academic pushback, and it is answerable (see §3), but not by ignoring it.

2. Why the two-column table is a cross-cut, not a taxonomy

[I] Following from §1: our table is not wrong so much as differently indexed. It cuts on what is scarce — a focal point versus a credible commitment — while the fields cut on what kind of thing is unknown or unenforceable. Both cuts are legitimate; only one has a literature behind it. The cost of the mismatch is that our page’s rows silently mix the two indices: “Missing information is about others’ intentions / others’ types” is an informational row, while “Scarce resource: common knowledge / credible commitment” is a commitment row, and the two do not line up. Intentions can be a commitment problem (I know what you would like to do; I cannot believe you will still want to do it tomorrow), and types can be a coordination problem (we would all adopt the same standard if we knew who else could actually implement it).

[B] Our own thesis, offered for challenge: the defensible reconstruction is a two-by-two, not a two-column table. One axis is what is missing — information about states of the world versus enforceability of a promise. The other is what the payoff structure does — whether the parties’ interests are aligned on the outcome and only their expectations fail (coordination), whether there is a rent to divide (bargaining), or whether defection is individually dominant regardless (social dilemma, §3). Coordination-with-information-problem and coordination-with-commitment-problem are then both real and distinguishable cells, which the current table cannot express.

[I] From that thesis and from the diagnostic test on the wiki page: the test — “Would the parties act if they simply knew what the others would do?” — is a test on the information axis only. It cannot distinguish a coordination failure from a commitment failure, because in a commitment failure the answer is also “no”, and for a reason the test does not capture: the parties know perfectly well what the others would like to do. The test should be split in two, and the second question is “would the parties act if a promise could be made binding?“.

3. The genus we were missing: the social dilemma

[F] Dawes gives a criterion that is formal enough to adjudicate membership: “a dilemma game is one in which all players have dominating strategies that result in a deficient equilibrium” (Dawes 1980, p. 179), where the deficiency is that “there is at least one other outcome preferred by everyone” (p. 170).

[I] From Dawes’s criterion together with the Van Huyck quotation in the key findings: the social dilemma is structurally distinct from both of our categories. It is not coordination — a coordination game has no dominant strategy and multiple equilibria; it is not bargaining — no private information is required, and the failure survives full transparency. Our page has no cell for it, which means it has no cell for the prisoner’s dilemma, the tragedy of the commons, or free-riding. [F] That is a conspicuous absence given that the same page cites Ostrom, whose entire subject matter sits in this genus, and that Olson’s foundational statement of the problem — “rational, self-interested individuals will not act to achieve their common or group interests” (Olson 1965, p. 2) — is a social-dilemma claim, not a coordination claim.

[F] The boundary is genuinely contested in one place: the Stag Hunt is classified as a coordination game in the game-theoretic tradition and as a social dilemma (an “assurance game”) in the social-psychology and commons traditions, depending on whether the analyst emphasises the multiplicity of equilibria or the risk-dominance of defection.

[I] From that classification split: the ambiguity is the finding, not a defect to be resolved. Any use case where the parties would all prefer mutual cooperation but the safe unilateral move is to withhold — which is a fair description of most industrial data-sharing, PACT included — sits precisely on this boundary, and its diagnosis will therefore depend on which tradition the reader belongs to. [B] Our own thesis: for CCTI this matters at M1, because a use case sitting on the Stag Hunt boundary is one where an assurance mechanism (proof that others have committed) does the work, and assurance is much cheaper to build than the full EIC stack.

[F] No unified taxonomy spanning coordination, bargaining, and social dilemmas as three siblings was found in any of the seven fields surveyed. Each field treats its own genus as primary and the others as neighbours.

[I] From that absence: writing one down is a genuine, if modest, contribution — and it is also a genuine exposure. There is no citation to hide behind, so the construction has to be defensible on its own terms.

4. The clause we never taxonomised: search and matching

The first verb in CCTI’s own thesis is “find counterparties”. We had no failure category for it.

[F] Roth’s market-design programme names three failure modes: markets fail when they lack thickness, when they suffer congestion, and when they are not safe — the last defined as markets failing “when participants can’t safely reveal or act on their true preferences” (Roth, NBER WP 13530, pp. 1, 3; Roth, HBR, October 2007). [F] In macroeconomics, the multiple steady states of search equilibria were “dubbed a coordination failure” (Diamond, Nobel lecture 2010).

[I] From Roth’s three modes and the wiki page’s table: “safety” is functionally the bargaining-failure problem under a different name — it is the condition that a participant can reveal a preference without that revelation being used against them. Roth never cross-references the Myerson–Satterthwaite literature, and the mechanism-design literature does not cite Roth’s safety condition, so the two remain unjoined. [B] Our own thesis: WP-C’s “credible matching” is best positioned in Roth’s vocabulary rather than ours — it is a safety intervention on a matching market, and framing it that way gives it a mature literature, an empirical track record, and a set of established evaluation criteria that a purely cryptographic framing does not.

[F] Cryptographic matching already exists as a small literature — private stable matching (Golle, Financial Cryptography 2006) and secure stable matching at scale (Doerner, Evans & Shelat, CCS 2016) — but it is published in security venues and does not cite Roth’s market-design work.

[I] From that publication split: WP-C sits in an unoccupied position between two literatures, which is favourable for novelty and unfavourable for reviewer sympathy. Either community will ask why the other’s work was not cited.

5. Does cryptography escape Myerson–Satterthwaite?

This is the question the wiki page implicitly answers “yes” — it calls the EIC “a direct assault on the Myerson–Satterthwaite obstacle”. The literature’s answer is no.

[F] The theorem itself: with two-sided private information and overlapping valuation supports, no incentive-compatible, individually-rational, budget-balanced mechanism achieves ex-post efficient trade (Myerson & Satterthwaite 1983, Corollary 1, p. 273). [F] The paper names its own escape routes: the positive-density assumption is load-bearing — “our proofs have used the assumption that the valuations have positive density … Without this assumption the corollary is untrue” (p. 273) — and it computes “the smallest lump-sum subsidy required from an outside party” (p. 272), i.e. relaxing budget balance.

[I] From the theorem’s own statement: the four published escapes are non-overlapping or non-full-support valuations, an outside subsidy, relaxing incentive compatibility or individual rationality, and the large-market limit. None of them is cryptography, and none of them is secrecy.

[F] The construction most often invoked as if it were a cryptographic escape says the opposite in its own introduction. Izmalkov, Lepinski & Micali’s Perfect Implementation states that they “regard this loss of privacy … as inherent, and do not wish to rectify or modify this — or any other — property of a normal-form mechanism”, and that a perfect implementation “should always enjoy economic efficiency to exactly the same degree (whatever it may be) as the abstract mechanism” (§1.1, p. 2). [F] The best-known industrial deployment claims confidentiality and the removal of a conflicted auctioneer — not efficiency (Bogetoft et al., Danish sugar-beet auction, Financial Cryptography 2009).

[F] The one paper that does show MPC circumventing an impossibility is also the one that shows the limits of doing so. Shi, Chung & Wu prove that MPC assistance evades an impossibility only by restricting the strategy space — an information-and-timing move — and record explicitly that “the MPC-assisted model does not help us circumvent the zero miner revenue lower bound” (ITCS 2023, §1.1.2).

[I] From the Shi–Chung–Wu result read against the Myerson–Satterthwaite corollary: the bound that survived MPC is a budget bound, which is the same family as Myerson–Satterthwaite’s budget-balance constraint. That is the closest thing in the literature to a direct answer to our question, and the answer is unfavourable. Cryptography changes how a mechanism is implemented and what leaks during implementation; it does not relax the constraint set the mechanism must satisfy.

[B] Our own thesis, and the strongest claim in this dossier: CCTI’s actual move is none of the four escapes. Myerson–Satterthwaite assumes types are soft — a valuation is a private number that cannot be proven, so the only lever is incentives. Attestation and zero-knowledge proof make some type-relevant facts hard: provable without being disclosed. A mechanism in which agents can certify claims about themselves is not a mechanism in the Myerson–Satterthwaite setting at all; it is a mechanism with evidence.

[I] From that thesis: the correct framing of the CCTI research question is therefore not “does cryptography beat Myerson–Satterthwaite” — which invites, and deserves, a flat no — but “how much of the efficiency loss is recovered as the certifiable fraction of the type space grows?”. That is a well-posed question with a plausible answer between zero and everything, and it is the question WP-C should be measuring.

[B] Named as our own claim and explicitly not yet verified: there is an existing formal literature on mechanism design with verifiable or certifiable evidence — the tradition running from partially verifiable information through costly verification — and it, not the MPC literature, will decide whether the fifth-escape claim survives review. This round did not read it. [I] From that gap: it is the highest-value target for the next round, because it is the one place where the project’s central economic claim could be either grounded or refuted outright, and because an economist reviewer will know it exists.

6. Where cryptography actually does work — and its one unguarded flank

[F] Verifiable disclosure has a real theoretical result behind it: when information is certifiable and disclosure is costless, full revelation unravels — the seller of a good who does not disclose is inferred to have the worst type (Grossman 1981; Milgrom 1981). This statement is carried at secondary-source grade: neither primary text could be obtained in machine-readable form in either round, and it should be verified before use in an external document.

[F] The unguarded flank is input authenticity. Secure multiparty computation “typically do[es] not specify how these inputs are chosen” (IACR ePrint 2022/1648); DECO proves the provenance of data from a TLS session, not its veracity, and hedges in its own §3.1 with multi-oracle majority agreement.

[F] The trusted-hardware assumption is falsifiable and has been falsified in practice: the SGAxe attack extracted actual SGX private attestation keys, allowing forged attestation of arbitrary enclaves (SGAxe, 2020).

[I] From SGAxe read against CCTI’s M5 gate: the project’s decision to treat TEEs as “a starting crutch, not the target architecture” is not conservatism, it is the minimum defensible position — the assumption it removes at M5 is one that has already been broken. [I] From the same, plus the ePrint 2022/1648 statement: the honest description of the attested-data layer is that it relocates trust — from a counterparty to a hardware vendor and its key-management practice — rather than removing it. That formulation has no single canonical citation; it is triangulated from the attack literature and from vendor-risk commentary, and should be marked as inference wherever we use it.

7. Program equilibrium — and a retraction

[F] The literature closest to CCTI’s own thesis is program equilibrium: agents that can read each other’s source code can achieve cooperation unavailable to agents that cannot, because the code itself is a commitment device (Tennenholtz 2004; carried at secondary grade, the primary text is behind a publisher paywall).

[F] Retraction, round 1. Two critiques of program equilibrium that the round-1 reconnaissance note stated as findings could not be sourced in round 2 and are hereby withdrawn: that mutual transparency can produce equilibria worse for all parties than ordinary Nash play, and that a trusted mediator reintroduces the commitment problem it was meant to solve. Targeted search across four program-equilibrium sources returned zero hits for the mediator claim, and the closest real statement to the first is Critch, Dennis & Russell’s remark that “certain formal institution designs that one might expect to defect against each other will instead turn out to cooperate, or conversely” — which is not the same claim.

[I] From the retraction: round 1’s failure mode is worth naming, because it will recur. Both withdrawn statements were plausible, both were the kind of thing the literature could say, and neither was invented from nothing — they were over-confident compressions of adjacent material. The defence against this is exactly the marker discipline used here: a claim that cannot carry an inline locator is not a finding.

[F] The Byzantine-Altruistic-Rational model formalises the same insight from the systems side: “the classical dichotomy between correct and faulty nodes becomes inadequate” once participants may be rational rather than merely correct or broken (BAR, TR-05-10, p. 1). [F] And there is a matching impossibility on the cryptographic side: rational secret sharing admits no bounded-round mechanism under standard assumptions (Halpern & Teague 2004, Theorem 3.1).

8. WP-E: the three candidate institutions are not three of a kind

CCTI’s WP-E commits to characterising one institutional equivalent among insurance, arbitration, and reputation. The choice is less free than it looks.

[F] Insurance bifurcates into two literatures with different failure modes. Formal, market insurance is risk-bearing that degrades through adverse selection — a hidden-type problem in the Rothschild–Stiglitz tradition. Informal, mutual insurance is a self-enforcement problem: Coate & Ravallion model it as “a simple repeated game of two self-interested households” (1993), and Fafchamps characterises such arrangements by “the limitations imposed by their self-enforcing nature”.

[I] From that bifurcation: choosing “insurance” for WP-E is not one choice but two, and the two land on opposite sides of the information/commitment axis established in §1. Market insurance is an information problem, and attestation attacks it directly — there is a real applied literature on attestation technology reducing adverse selection, in telematics-based motor insurance. Mutual insurance is a commitment problem sustained by repeated interaction, which makes it structurally a reputation mechanism wearing a different name. [B] Our own thesis: WP-E’s menu is therefore effectively two options, not three — an information institution (market insurance, arbitration’s fact-finding function) and a commitment institution (reputation, mutual insurance) — and the project should choose on that axis rather than on the institution’s name.

[F] The best historical model for what CCTI is actually building is the medieval Law Merchant as analysed by Milgrom, North & Weingast: the private judges could not enforce their judgments, and their function was the transmission of information that fed a reputation mechanism (Milgrom, North & Weingast 1990, pp. 2, 3, 10, 19).

[I] From the Law Merchant analysis read against the CCTI architecture: this is the closest institutional analogue in the literature to an attestation layer — an institution that supplies verified information to a decentralised reputation system while possessing no enforcement power of its own. It is also a warning, since the Law Merchant’s effectiveness depended on merchants’ willingness to consult the record and to act on it, which is an adoption condition, not a technical one.

9. The empirical headwind at the M4 AI-Advantage gate

[F] A controlled study with 216 human participants finds that “Humans and LLMs achieve comparable aggregate surplus within their groups” and warns that “performance parity … can mask substantive procedural differences” (arXiv:2509.09071). [F] A second study finds that LLM negotiators “falter on consistency, legitimacy, and commitment when stakes rise” (Bhattacharya et al., HFES 2025). [F] In the opposite direction, autonomous pricing algorithms have been shown to learn supracompetitive pricing without communication (Calvano et al., AER 2020), though a 2026 replication finds that collusion fragile under realistic perturbation.

[I] From the parity finding read against CCTI’s headline thesis: the current empirical baseline does not support “AI agents negotiate better than humans”, and a reviewer will produce this citation. But the parity finding is not a refutation of CCTI’s claim, because CCTI’s claim is conditional — agents reach agreements given verifiable trust infrastructure that humans lacking prior trust history would not. [F] No study located in either round has deployed verifiable commitment or code transparency between negotiating agents and measured the effect on agreement rates or surplus.

[I] From that absence together with the parity finding: WP-D’s evaluation is first-of-kind, which is exactly the value proposition, and the correct experimental design is therefore not “LLM versus human” but “agents with attested claims versus agents without”, holding the agent constant. Any design that compares against a human baseline invites the parity literature and answers a question the project is not asking. [B] Our own thesis: this should be settled before M1, because the use case selected at M1 determines whether such a comparison is even constructible.

10. Consequences for the wiki page

[I] Three amendments follow from §§1–4, all of them corrections rather than additions:

1. The social dilemma must be added as a third genus with Dawes’s dominant-strategy criterion as its membership test, and the page’s claim that private information is the “common mechanism” underneath both existing categories must be withdrawn — Grossman & Hart’s symmetric-information hold-up is a counterexample.
2. The “Scarce resource” row conflates credible revelation (an information problem, addressed by proof) with credible binding (a commitment problem, addressed by escrow, enforcement, or a repeated game). These are the two mechanisms Fearon separates, and cryptography’s contribution is overwhelmingly on the first. Splitting the row is the single highest-value edit.
3. The third objection on record — that Ostrom’s monitoring works partly socially and that SINE builds monitoring that shows nothing — is a good objection but it is ours. [F] It is not a named controversy in the literature: adjacent real debates exist on informational versus normative social influence (Deutsch & Gerard 1955) and on the generalisability of the design principles (Cox, Arnold & Villamayor-Tomas 2010), but no source frames a debate about the nature of monitoring. It should be marked as inference, not presented as received scholarly dispute. By contrast, [F] the second objection — that standards bodies are distributive arenas — is a modelled literature, with Farrell & Simcoe treating standard-setting as a war of attrition, and should be upgraded with that citation.

Process note. Two reconnaissance rounds, 14 subagents in round 2 (all on Sonnet per standing policy), roughly 1.1 M subagent tokens against 287 k in round 1. The increase is entirely primary-text acquisition: fetching and reading PDFs is fetch-heavy and several agents spent most of their budget failing to obtain a paywalled or image-only text. One agent burned 161 k tokens across 17 tool uses and returned almost nothing.

Texts that remain unobtained after two rounds, and which therefore cap what this dossier can claim at primary grade: Grossman 1981 and Milgrom 1981 (Milgrom’s own hosted PDF is an image-only scan; Chicago Unbound states full text unavailable); Maskin & Tirole 1999 (JSTOR paywall, Harvard mirror blocked by the proxy); Ostrom, Governing the Commons, ch. 2 (one PDF exceeds the fetch size limit, another 403s, the publisher reader carries only ch. 1); Harsanyi & Selten 1988; Aumann 1976 (image-only scan); Tennenholtz 2004 (Elsevier paywall); Kollock 1998; Rothschild & Stiglitz 1976 (all mirrors image-encoded).

Consequence for our Ostrom page: we currently carry the eight design principles and the appropriation/provision distinction without ever having read chapter 2 in primary form. That is a live exposure, not a footnote, and it is worth solving by buying the book rather than by another round of subagents.

What would refute this analysis?

• The fifth-escape claim (§5) is the most exposed statement in this dossier. It would be refuted by a result in the mechanism-design-with-evidence literature showing that certifiability does not relax the Myerson–Satterthwaite constraint set — for instance a theorem that budget balance, individual rationality and efficiency remain jointly unattainable even when types are verifiable at zero cost. We have not read that literature. If such a theorem exists, CCTI’s economic framing needs to change, not merely its wording.
• The three-genus taxonomy (§3) would be refuted by a source that subsumes social dilemmas under coordination. The risk-dominance tradition comes close, and the Stag Hunt’s dual classification is evidence that the boundary is soft rather than sharp. If a standard reference treats assurance games as the general case, our third genus collapses back into the first.
• The information/commitment axis (§1) would be refuted by a field that cuts differently and is not derivative. Behavioural and experimental economics is the obvious candidate and was surveyed only thinly here; a literature organised around trust, reciprocity, or norm-following rather than around information and enforcement would show that the convergence in §1 is a convergence among rationalist traditions specifically, not among all of them. That would not make §1 wrong, but it would narrow its scope considerably.
• The verifiable-disclosure result (§6) is currently carried at secondary grade. If the unravelling result turns out to require assumptions we cannot meet — costless disclosure, common knowledge that the sender is informed, a fully ordered type space — then the argument that attestation drives full revelation weakens sharply, and it weakens exactly where CCTI needs it.
• The M4 argument (§9) depends on an absence. We claim no prior study has measured verifiable commitment between negotiating agents. Absence of evidence from two rounds of search is weak evidence of absence, and a single overlooked paper in a multi-agent-systems venue would remove the first-of-kind claim.
• Missing evidence, stated plainly. Eight primary texts listed in the internal section were not obtained. Everything this dossier says about Ostrom, about Aumann’s common-knowledge result, about Rothschild–Stiglitz, and about program equilibrium’s original construction rests on secondary reading and is marked accordingly. None of those gaps affects the central argument of §5, which rests entirely on texts we did read.

Sources

Primary texts read in full or in the cited sections:

• Fearon, J. D. (1995). Rationalist Explanations for War. International Organization 49(3): 379–414. https://slantchev.ucsd.edu/courses/pdf/fearon-io1995v49n3.pdf — pp. 404, 409.
• Myerson, R. B. & Satterthwaite, M. A. (1983). Efficient Mechanisms for Bilateral Trading. Journal of Economic Theory 29: 265–281. https://www.cs.princeton.edu/courses/archive/spr10/cos444/papers/myerson_satterthwaite83.pdf — Corollary 1 and pp. 272–273.
• Dafoe, A. et al. (2020). Open Problems in Cooperative AI. arXiv:2012.08630. https://ar5iv.labs.arxiv.org/html/2012.08630 — §§4.3, 4.4.
• Shi, E., Chung, H. & Wu, K. (2023). What Can Cryptography Do for Decentralized Mechanism Design? ITCS 2023. arXiv:2209.14462. https://arxiv.org/abs/2209.14462 — §1.1.2, p. 7.
• Izmalkov, S., Lepinski, M. & Micali, S. Perfect Implementation of Normal-Form Mechanisms. — §1.1, p. 2.
• Roth, A. E. (2007). What Have We Learned from Market Design? NBER Working Paper 13530. https://www.nber.org/papers/w13530 — pp. 1, 3. With Roth, A. E. (2007), The Art of Designing Markets, Harvard Business Review, October 2007.
• Compute, but Verify: Efficient Multiparty Computation over Authenticated Inputs. IACR ePrint 2022/1648. https://eprint.iacr.org/2022/1648 — abstract.
• Van Huyck, J., Battalio, R. & Beil, R. (1990). Tacit Coordination Games, Strategic Uncertainty, and Coordination Failure. American Economic Review 80(1): 234–248 — p. 244.
• Dawes, R. M. (1980). Social Dilemmas. Annual Review of Psychology 31: 169–193 — pp. 170, 179.
• Powell, R. (2006). War as a Commitment Problem. International Organization 60(1): 169–203 — p. 176.
• Grossman, S. & Hart, O. (1986). The Costs and Benefits of Ownership. Journal of Political Economy 94(4): 691–719 — p. 696.
• Hart, O. & Moore, J. (1990). Property Rights and the Nature of the Firm. Journal of Political Economy 98(6): 1119–1158 — fn. 6.
• Klein, B., Crawford, R. & Alchian, A. (1978). Vertical Integration, Appropriable Rents, and the Competitive Contracting Process. Journal of Law and Economics 21(2): 297–326 — p. 298.
• Milgrom, P., North, D. & Weingast, B. (1990). The Role of Institutions in the Revival of Trade. Economics and Politics 2(1): 1–23 — pp. 2, 3, 10, 19.
• Olson, M. (1965). The Logic of Collective Action — p. 2.
• Halpern, J. & Moses, Y. (1990). Knowledge and Common Knowledge in a Distributed Environment. JACM 37(3) — Corollary 6.
• Halpern, J. & Teague, V. (2004). Rational Secret Sharing and Multiparty Computation. STOC 2004 — Theorem 3.1.
• Aiyer, A. et al. (2005). BAR Fault Tolerance for Cooperative Services. UT Austin TR-05-10 — p. 1.
• SGAxe: How SGX Fails in Practice (2020). https://sgaxe.com/files/SGAxe.pdf.
• Doerner, J., Evans, D. & shelat, a. (2016). Secure Stable Matching at Scale. CCS 2016. https://eprint.iacr.org/2016/861.
• Calvano, E., Calzolari, G., Denicolò, V. & Pastorello, S. (2020). Artificial Intelligence, Algorithmic Pricing, and Collusion. American Economic Review 110(10): 3267–3297. https://www.aeaweb.org/articles?id=10.1257/aer.20190623.
• Negotiation experiment with N=216 human participants. arXiv:2509.09071. https://arxiv.org/abs/2509.09071.

Consulted at secondary grade, primary text not obtained (each so marked in the text):

• Grossman, S. (1981) and Milgrom, P. (1981) on verifiable disclosure and unravelling.
• Tennenholtz, M. (2004). Program Equilibrium. Games and Economic Behavior 49(2): 363–373.
• Ostrom, E. (1990). Governing the Commons, ch. 2.
• Rothschild, M. & Stiglitz, J. (1976); Maskin, E. & Tirole (1999); Aumann, R. (1976); Harsanyi, J. & Selten, R. (1988).
• Coate, S. & Ravallion, M. (1993). Reciprocity Without Commitment. Journal of Development Economics 40(1): 1–24; and Fafchamps, M. on self-enforcing risk-sharing.
• Farrell, J. & Simcoe, T. (2012). Choosing the Rules for Consensus Standardization. RAND Journal of Economics 43(2): 235–252.
• Deutsch, M. & Gerard, H. (1955); Cox, M., Arnold, G. & Villamayor-Tomas, S. (2010).
• Bhattacharya et al. (2025), HFES Proceedings, on LLM negotiator behaviour under raised stakes.
• Critch, A., Dennis, M. & Russell, S. (2022). Cooperative and Uncooperative Institution Designs. arXiv:2208.07006.
• Golle, P. (2006). A Private Stable Matching Algorithm. Financial Cryptography 2006.
• Bogetoft, P. et al. (2009). Secure Multiparty Computation Goes Live. Financial Cryptography 2009.
• Diamond, P. (2010). Unemployment, Vacancies, Wages. Nobel Prize lecture.

Internal sources:

• Coordination failure vs. bargaining failure — the synthesis page this dossier audits.
• CCTI — thesis, work packages, milestones; derived from sources/proposals/2026-06-17-aria-ccti-proposal-draft-2.md.
• Reconnaissance note, round 1 — superseded in part; see the retraction in §7.