Paper deep dive
Epistemic Disturbance in the Graph Model for Conflict Resolution: State-Preserving Actions, Four-Valued Assessments, and the Distinction between Capability and Intention
Yukiko Kato
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 10/10/2026, 2:13:55 AM
Summary
This paper extends the Graph Model for Conflict Resolution (GMCR) by introducing 'state-preserving actions' that alter decision makers' (DMs) epistemic states without changing the physical conflict state. Using a four-valued logic (Belnap's lattice) to represent evidence for and against moves, the authors distinguish between capability and intention assessments. The study demonstrates that evidence for moves enables perceived moves while evidence against disables them, and that stability judgments change monotonically based on the type of evidence and the observer's reduction operator. The framework is applied to the 1995 DVD format negotiation to show how sequential stability can distinguish negotiation phases where general metarationality fails.
Entities (10)
Relation Signals (7)
State-Preserving Actions → alters → Epistemic States
confidence 95% · A state-preserving action changes only the epistemic state; inaction is the absence of a transition.
Evidence Against → disables → Perceived Moves
confidence 92% · evidence against can only disable them
Evidence For → enables → Perceived Moves
confidence 92% · evidence for the availability of a move can only enable perceived moves
Capability Assessments → affects → Sanction-Based Stability Concepts
confidence 90% · Capability assessments affect all sanction-based stability concepts
Intention Assessments → affects → Sequential Stability
confidence 90% · intention assessments affect only sequential stability.
Sequential Stability → distinguishes → 1995 DVD Format Negotiation Phases
confidence 88% · sequential stability, which asks whether it would, can [distinguish the phases].
General Metarationality → failstodistinguish → 1995 DVD Format Negotiation Phases
confidence 88% · general metarationality cannot distinguish its phases
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In the graph model for conflict resolution (GMCR), a decision maker (DM) either moves the conflict to another state or does nothing. The basic definitions leave inaction implicit, so every action that leaves the state unchanged is treated as doing nothing. Yet announcements, exercises, leaks and selective disclosures are neither moves nor inaction: they leave the state unchanged but change what other DMs believe about which moves are available and which moves others would want to make. We introduce such state-preserving actions by augmenting states with the DMs' epistemic states: a physical move changes the physical state, a state-preserving action changes only the epistemic state, and inaction is the absence of a transition. Actions generate evidence through observer-specific interpretation maps. Building on a four-valued extension of GMCR from the author's earlier work, which separates evidence for and against, we show that evidence for a move can only enable perceived moves and evidence against can only disable them, that two of the four reduction operators ignore one kind of evidence, and that contradictory assessments are absorbing under monotone accumulation. With the monotonicity of stability in move sets, this fixes the direction in which any action moves a DM's stability judgements and characterizes when actions can enable provocation or deterrence. Capability assessments affect all sanction-based stability concepts, and on the DM's own side also Nash stability, whereas intention assessments affect only sequential stability. Hedging between two candidate types weakly expands or shrinks an observer's sequentially stable set according to how it reads contradiction. In the 1995 DVD format negotiation, general metarationality cannot distinguish its phases, since the computer industry group could always sanction; sequential stability, which asks whether it would, can.
Tags
Links
- Source: https://arxiv.org/abs/2610.11690v1
- Canonical: https://arxiv.org/abs/2610.11690v1
Trouble viewing inline? Open PDF directly →
Full Text
78,207 characters extracted from source content.
Expand or collapse full text
Epistemic Disturbance in the Graph Model for Conflict Resolution: State-Preserving Actions, Four-Valued Assessments, and the Distinction between Capability and Intention Yukiko Kato Affiliation: Lynx Technologies Inc., Tokyo, Japan Affiliation: Institute of Science Tokyo, Tokyo, Japan Email: ykato@lynxjp.com Draft Abstract In the graph model for conflict resolution (GMCR), a decision maker (DM) either moves the conflict to another state or does nothing. The basic definitions leave inaction implicit: remaining at a state is not a move and receives no analysis of its own, so every action that leaves the state unchanged is treated as doing nothing. Many actions in real conflicts are neither moves nor inaction: announcements, exercises, leaks and selective disclosures leave the state of the conflict unchanged but change what other DMs believe about the moves available to whom and about the moves each DM would want to make. We introduce such state-preserving actions into GMCR by augmenting states with the DMs’ epistemic states. A physical move changes the physical state; a state-preserving action changes only the epistemic state; inaction is the absence of a transition. A state-preserving action generates evidence in each observer through an observer-specific interpretation map, so that the same action may be read differently by different DMs. To represent assessments that may lack evidence or contain conflicting evidence, and in which evidence accumulates, we build on a four-valued extension of GMCR proposed in the author’s earlier work. We show that evidence for the availability of a move can only enable perceived moves and evidence against can only disable them, that two of the four reduction operators ignore one kind of evidence entirely, and that contradictory assessments are absorbing under monotone accumulation of evidence. Combined with the monotonicity of stability in move sets, these facts determine the direction in which any state-preserving action moves a DM’s stability judgements, and characterize the channels through which an action can enable provocation or deterrence. Capability assessments affect all sanction-based stability concepts and, on the focal DM’s own side, Nash stability as well, whereas intention assessments affect only sequential stability. Hedging between two candidate types weakly expands the set of states an observer judges sequentially stable if the observer reads contradictory evidence as credibility, and weakly shrinks it otherwise. The 1995 negotiation that unified the two competing DVD formats illustrates the distinction: general metarationality cannot distinguish the phases of the negotiation, because the technical working group of computer companies could always sanction, whereas sequential stability, which asks whether it would, can. 1 Introduction A graph model records the state of a conflict as a combination of options taken by the decision makers (DMs), and a move as a change of that combination [19, 6]. In the basic definitions, a move is an arc between two distinct states, and the reachable lists from which stability is analysed exclude the current state; excluding it is a standard convention in the literature [23]. Remaining at a state is therefore not a move. It is not examined in its own right either: it enters the definitions only implicitly, as what a DM does when it does not move, and stability analysis asks whether a DM has reason to leave a state, never what a DM does while it stays. Everything that happens without a change of state is thereby treated as one and the same: doing nothing. This implicit treatment of inaction is the starting point of this paper. Many actions in real conflicts leave the combination of options unchanged and still matter: a declaration that a sanction is ready, a military exercise, a leak suggesting that a threat is empty, a private signal to one party but not another, a deceptive signal about future capability or intention. What such an action changes is not the state of the conflict but what other DMs believe about which moves are available and which moves the others would want to make. This paper introduces such actions into the graph model. The central distinction is between three things that the physical state alone cannot tell apart. Let s be a physical state and κ a profile of the DMs’ epistemic states. Then (s,κ)⟶(t,κ)s≠tphysical move,(s,κ)⟶(s,κ′)κ≠κ′state-preserving action,(s,κ)no transitioninaction. array[]l(s,κ) (t,κ)&s≠ t&physical move,\\[2.0pt] (s,κ) (s,κ )&κ≠κ &state-preserving action,\\[2.0pt] (s,κ)&no transition&inaction. array Under the projection onto the physical state, a state-preserving action looks like a self-loop s→s→ s, and in the basic definitions of the graph model it is indistinguishable from inaction. It is not a null action; it is null only under that projection. Making inaction explicit, as the absence of a transition, is what allows state-preserving actions to be defined at all. To represent the phenomena considered here, we distinguish at least three epistemically different situations that a binary representation (available or not) collapses: the absence of any evidence, evidence in one direction, and conflicting evidence. We also let evidence be added without erasing what was there before, since an announcement or a leak adds to what others know rather than replacing it. A four-valued extension of GMCR proposed by the author in earlier work [13, 17] is a particularly economical way to make these distinctions, because it separates evidence for from evidence against explicitly. It uses Belnap’s four values [2, 5]: evidence only for ( Y), only against ( N), neither ( U) or both ( YN). This extension is recent and not yet widely adopted, and we do not presuppose it as the standard representation of beliefs in the graph model. We adopt it because it makes the situation studied here straightforward to model and its effects explicit. Probabilities could also express uncertainty, but a single precise probability does not in general distinguish the absence of evidence from balanced conflicting evidence without additional structure, whereas =(0,0) U=(0,0) and =(1,1) YN=(1,1) are distinct values here; precise probabilities also require quantitative belief information that the standard graph model does not require. Richer formalisms such as imprecise probabilities or belief functions can make the distinction, at the cost of further quantitative structure. Hypergame models represent misperception through subjective games and generally do not separate the lack of evidence from conflicting evidence. When every assessment is Y or N, the framework reduces to the standard graph model with perceived move sets. A state-preserving action generates evidence in an observer, and the same action may generate different evidence in different observers. We therefore separate the action from its effect: each observer has an interpretation map that turns an action into a change of its own assessment. A DM acts on its four-valued assessments through a reduction operator that sends U and YN to Y or N; the four operators of this kind react to evidence in structurally different ways. Our results are as follows. ∙ We make explicit the treatment of inaction, which the basic definitions of the graph model leave implicit, and distinguish it from state-preserving actions by augmenting states with epistemic states (Section 4). Inaction is the absence of a transition; a state-preserving action is a transition that changes only the epistemic state; the convention that remaining at a state is not a move is preserved. ∙ Evidence for the availability of a move can only enable perceived moves, and evidence against can only disable them, for all four operators (Proposition 5.1); two operators ignore one of the two kinds of evidence entirely (Proposition 5.2); and the assessments that no further evidence can change, as far as action is concerned, are the contradictory ones and the definite ones that the operator resolves in the same way as contradiction (Proposition 5.3). ∙ Stability is monotone in the opponents’ move sets and antitone in the focal DM’s own (Lemma 6.2). Together these facts determine the direction in which a state-preserving action moves a DM’s judgement of its own stability (Theorems 7.1 and 7.3), and give a sufficient condition for a judgement to be immune to every action (Proposition 7.4). ∙ We define actions that enable provocation or deterrence, and show which operators are open to which, through which channel (Corollary 8.2). ∙ Capability assessments affect all sanction-based stability concepts and, on the focal DM’s own side, Nash stability as well, whereas intention assessments affect only sequential stability (Theorems 7.1 and 7.3, Table 1). ∙ When intention assessments derive from evidence about candidate types of a DM, hedging between two types makes exactly the disputed moves contested, and weakly expands or weakly shrinks an observer’s set of sequentially stable states according to how the observer resolves contradiction (Theorem 9.3). ∙ In the 1995 DVD format negotiation, general metarationality cannot distinguish the phases, because the working group of computer companies could always sanction, whereas sequential stability can, because it asks whether the group would (Section 10). The central result, Theorem 7.1, can be summarized as follows, where “stabilizing” means that the set of states a DM judges stable can only grow: evidence forevidence againstabout opponents’ movesstabilizingdestabilizingabout own movesdestabilizingstabilizing array[]l|c&evidence for&evidence against\\ opponents' moves&stabilizing&destabilizing\\ about own moves&destabilizing&stabilizing array How DMs anticipate the evolution of the conflict under their assessments, and how an actor can plan a sequence of state-preserving actions, is the subject of a companion paper [14]; the present paper concerns the effect of state-preserving actions on stability judgements. 2 Related work and positioning 2.1 Moves and inaction in the graph model The graph model represents a conflict by states, the moves that each DM controls between them, and the DMs’ preferences [19, 6]. Moves are arcs between distinct states, and the reachable lists used in stability analysis exclude the current state, a convention stated explicitly in later work [23]. Stability concepts ask whether a DM has an improvement that it would take, and status quo analysis traces the evolution of a conflict through sequences of moves [26]. In both, what a DM does without changing the state is not represented: it coincides with doing nothing. To our knowledge, the distinction between inaction and actions that leave the state unchanged has not been drawn in the graph model. The categorical formulation of [15] makes identities explicit as morphisms, which suggests the distinction, but does not model what such actions change. 2.2 Misperception and unawareness Misperception of options and preferences in the graph model has been studied through hypergames [24, 1], generalized to several levels of misperception [35], through perceptual graph models [22], and through interactive unawareness, including higher-order awareness [23]. These frameworks take a structure of misperception or unawareness as given and ask what stability and equilibria follow from it; in the model of [23], awareness may also increase along moves. None of them treats actions whose purpose and effect is to change the others’ perceptions while the state stays the same. 2.3 Manipulation and deception Within the graph model, strategic manipulation has been studied as the costly adjustment of other parties’ preferences through a social trust network, formulated as an optimization problem [27], and Inohara et al. showed, in a framework of interdependent preferences, conditions under which deceit about preferences cannot succeed [12]. Outside the graph model, deception in 2×22× 2 games [3], deception in hypergames, including stealthy deception in which a player manipulates another’s perception without being detected [36, 37], and strategic information transmission [4] concern related phenomena, typically with cardinal payoffs or probabilistic beliefs. 2.4 Representations of evidence Belnap’s four-valued logic was designed for reasoning under possibly inconsistent incoming information [2, 5], and its bilattice structure separates the order of truth from the order of information [9]. The four-valued extension of the graph model proposed in [13, 17] uses it to represent the status of options, with reduction operators that turn four-valued assessments into decisions. It has so far been used for given assessments, not for their change through actions. 2.5 Positioning Against this background, the contributions of this paper are the following. First, it makes explicit the treatment of inaction that the basic definitions leave implicit, and distinguishes it from state-preserving actions within the graph model itself. Second, where the literature on misperception analyses given structures of perception, this paper analyses how state-preserving actions transform them, through observer-specific interpretation maps that cover both the common signal read differently and selective disclosure. Third, where manipulation in the graph model has concerned preferences, this paper keeps preferences fixed and acts on assessments of capability and intention, and it shows that these two kinds of assessment act on different stability concepts. Fourth, the direction of each effect is determined ordinally, by the kind of evidence and the observer’s reduction operator, without cardinal payoffs or probabilities; the results depend on the four-valued extension only through the evidence-pair structure of Section 5. 3 Preliminaries 3.1 Graph models, sanctions and stability We follow the conventions of [16]. A graph model G=(N,S,(Ai)i∈N,(≿i)i∈N)G=(N,S,(A_i)_i∈ N,( _i)_i∈ N) has a finite set N of DMs, a finite set S of states, arc sets Ai⊆S×SA_i S× S and complete, not necessarily transitive, preference relations ≿i _i on S, with strict part ≻i _i. Remaining at a state is always possible and is not a move. For s∈Ss∈ S, Ri(s)=t:(s,t)∈Ai,t≠s,Ri+(s)=t∈Ri(s):t≻is,Φi≃(s)=t:s≿it.R_i(s)=\t:(s,t)∈ A_i,\ t≠ s\, R_i^+(s)=\t∈ R_i(s):t _is\, _i(s)=\t:s _it\. For i∈Ni∈ N, RN−i(s)R_N-i(s) and RN−i+(s)R_N-i^+(s) are the states reachable from s by legal sequences of moves, respectively improvements, of the DMs other than i, in which no DM moves twice in succession [6]. We use the four basic stability concepts of the graph model [19, 6] and state them through an explicit sanction predicate. Let i∈Ni∈ N, s∈Ss∈ S and t∈Ri+(s)t∈ R_i^+(s). The improvement from s to t has ∙ a GMR-sanction if some s′∈RN−i(t)s ∈ R_N-i(t) satisfies s≿is′s _is ; ∙ an SMR-sanction if some such s′s also satisfies Ri(s′)⊆Φi≃(s)R_i(s ) _i(s); ∙ an SEQ-sanction if some s′∈RN−i+(t)s ∈ R^+_N-i(t) satisfies s≿is′s _is . We write Sanciσ(s,t)Sanc^σ_i(s,t) when the improvement has a σ-sanction. State s is Nash stable for i [20, 21] if Ri+(s)=∅R_i^+(s)= , and general metarational (GMR) [10], symmetric metarational (SMR) [10] or sequentially stable (SEQ) [7, 8] for i if every t∈Ri+(s)t∈ R_i^+(s) satisfies Sanciσ(s,t)Sanc^σ_i(s,t) for σ=GMRσ=GMR, SMRSMR or SEQSEQ respectively. Let =Nash,GMR,SMR,SEQX=\Nash,GMR,SMR,SEQ\ and write Stabiσ(G)Stab^σ_i(G) for the set of states σ-stable for i in G. Coalitional versions of these concepts [18, 11] are not treated in this paper. For s∈Ss∈ S let D^G(s) D_G(s) be the forward-reachable set of s, the smallest set containing s and closed under arc successors. By the localization lemma [16, Lemma 3.4, Cor. 3.5], the σ-stability of s for any DM is determined by D^G(s) D_G(s), the arcs with tails in D^G(s) D_G(s) and the preferences restricted to D^G(s) D_G(s). 3.2 Four values as evidence pairs We write each element of Belnap’s lattice 4 as a pair (f,a)∈0,12(f,a)∈\0,1\^2 recording whether there is evidence for (f) and against (a): =(1,0),=(0,1),=(0,0),=(1,1). Y=(1,0), N=(0,1), U=(0,0), YN=(1,1). The knowledge order is componentwise, (f,a)≤k(f′,a′)(f,a) _k(f ,a ) iff f≤f′f≤ f and a≤a′a≤ a , with bottom U and top YN. The truth order is (f,a)≤t(f′,a′)(f,a) _t(f ,a ) iff f≤f′f≤ f and a≥a′a≥ a , with bottom N and top Y. Conjunction is the truth-order meet (f,a)∧(f′,a′)=(min(f,f′),max(a,a′))(f,a) (f ,a )=( (f,f ), (a,a )), and the empty conjunction is Y [2, 9]. Lemma 3.1. Conjunction is monotone in both the knowledge order and the truth order. The f-component of a conjunction depends only on the f-components of its arguments, and the a-component only on their a-components. Proof. Immediate from the componentwise formula. ∎ 3.3 Reduction operators A reduction operator is a map δ:→,δ 4→\ Y, N\ with δ()=δ( Y)= Y and δ()=δ( N)= N. There are exactly four, determined by (δ(),δ())(δ( U),δ( YN)); we write δ _ N Y for the one with δ()=δ( U)= N and δ()=δ( YN)= Y, and similarly δ _ Y N, δ _ N N and δ _ Y Y. In terms of evidence, δ(f,a)= _ N Y(f,a)= Y ⇔f=1 f=1 (any evidence for suffices), (any evidence for suffices), δ(f,a)= _ Y N(f,a)= Y ⇔a=0 a=0 (any evidence against defeats), (any evidence against defeats), δ(f,a)= _ N N(f,a)= Y ⇔f=1anda=0 f=1\ and\ a=0 (only uncontested evidence for), (only uncontested evidence for), δ(f,a)= _ Y Y(f,a)= Y ⇔f=1ora=0 f=1\ or\ a=0 (anything but uncontested evidence against). (anything but uncontested evidence against). The operator δ _ N Y is the operator δprec _prec of [17], the only one that is monotone in the knowledge order. All four are monotone in the truth order [17]. The notation differs from that of [17]: here each operator is indexed directly by its values on U and YN. 4 State-preserving actions 4.1 Epistemic states and perceived models Let O=⨆i∈NOiO= _i∈ NO_i be a finite set of options, OiO_i being those controlled by DM i. Each arc a∈Aia∈ A_i carries a requirement req(a)⊆Oireq(a) O_i, the options of i whose availability the move depends on. For i∈Ni∈ N let A−i=⋃j≠iAjA_-i= _j≠ iA_j. Definition 4.1 (Epistemic state). The epistemic state of DM i is a pair κi=(ψi,χi)κ^i=(ψ^i,χ^i) of ∙ a capability assessment ψi:O→ψ^i O 4, recording the evidence i holds on the availability of each option; and ∙ an intention assessment χi:A−i→χ^i A_-i 4, recording, for each move of another DM, the evidence i holds that the move would be an improvement for the DM who controls it. The set of epistemic states of DM i is Ki=O×A−iK_i=4^O×4^A_-i, and an epistemic profile is an element κ=(κi)i∈Nκ=(κ^i)_i∈ N of the finite set K=∏i∈NKiK= _i∈ NK_i. We write κi≤kκ′iκ^i _kκ i if the knowledge order holds in every component of both assessments. Each DM i uses a reduction operator δi _i. The warrant of an arc a under ψ is wψ(a)=⋀o∈req(a)ψ(o)w_ψ(a)= _o (a)ψ(o). Definition 4.2 (Perceived model and judgements). The perceived arcs of DM i are Aj[κi]=a∈Aj:δi(wψi(a))=A_j[κ^i]=\a∈ A_j: _i(w_ψ^i(a))= Y\ for j∈Nj∈ N, and its perceived credible arcs are Bj[κi]=a∈Aj[κi]:δi(χi(a))=B_j[κ^i]=\a∈ A_j[κ^i]: _i(χ^i(a))= Y\ for j≠ij≠ i, while Bi[κi]B_i[κ^i] consists of the arcs of Ai[κi]A_i[κ^i] that are improvements for i. The perceived model G[κi]G[κ^i] is the graph model with arcs Aj[κi]A_j[κ^i]. DM i judges s to be σ-stable for itself if s is σ-stable for i in G[κi]G[κ^i], where for σ=SEQσ=SEQ the sets RN−i+R^+_N-i are replaced by the sets RN−iBR^B_N-i of states reachable by legal sequences along perceived credible arcs. We write Stabiσ[κi]Stab^σ_i[κ^i] for the set of states i judges σ-stable. Preferences are taken to be those of the model; what a DM may misjudge is which moves are available and which moves the others would want to make. When χi(a)=χ^i(a)= Y exactly for the arcs that are improvements for their controllers and χi(a)=χ^i(a)= N otherwise, Bj[κi]B_j[κ^i] is the set of perceived improving arcs, and the judgements are standard stability in the perceived model. When, moreover, every capability assessment is Y or N and agrees with the model, the perceived model is the model itself. 4.2 Three kinds of transition Definition 4.3 (Augmented state space and transitions). The augmented state space is S~=S×K S=S× K, with projection πS(s,κ)=s _S(s,κ)=s. A transition of S~ S is one of the following. (a) A physical move of DM i: (s,κ)→(t,κ)(s,κ)→(t,κ) with (s,t)∈Ai(s,t)∈ A_i and t≠st≠ s. (b) A state-preserving action of DM k: (s,κ)→(s,κ′)(s,κ)→(s,κ ) with κ′≠κ ≠κ, as specified in Definition 4.4. Inaction at (s,κ)(s,κ) is the absence of a transition. In the present model physical moves leave epistemic states unchanged: we abstract from the epistemic updating that observed physical moves cause in practice, which would require an additional observation rule. Under πS _S, a state-preserving action is sent to a self-loop at s, while inaction involves no transition at all. The convention that remaining at a state is not a move, and hence that Ri(s)R_i(s) excludes s, concerns inaction; it does not exclude state-preserving actions, which are transitions of the augmented space and do not enter the legal sequences of the physical model. Definition 4.4 (Actions and interpretation). For each DM k and state s, let Actk(s)Act_k(s) be a finite set of state-preserving actions available to k at s. For each observer i≠ki≠ k, an interpretation map Γi:Actk(s)×Ki→Ki _i _k(s)× K_i→ K_i assigns to each d∈Actk(s)d _k(s) and epistemic state κiκ^i an epistemic state Γi(d,κi) _i(d,κ^i) with κi≤kΓi(d,κi).κ^i _k _i(d,κ^i). Performing d at (s,κ)(s,κ) leads to (s,κ′)(s,κ ), where κ′i=Γi(d,κi)κ i= _i(d,κ^i) for i≠ki≠ k and κ′k=κkκ k=κ^k. The change κi↦Γi(d,κi)κ^i _i(d,κ^i) is the disturbance of i’s epistemic state induced by d, and k→i(s,κi)=Γi(d,κi):d∈Actk(s)D_k→ i(s,κ^i)=\ _i(d,κ^i):d _k(s)\ is the set of disturbances of i that k can induce at s. The dependence of Γi _i on the actor k and the physical state s is suppressed in the notation. We assume that a state-preserving action does not directly alter the actor’s own epistemic state. In practice an actor may learn from the reactions to its own action, as with a probe or a trial announcement; allowing such self-updating is a straightforward extension that we do not pursue here. An action is not itself evidence; it generates evidence in each observer, and different observers may read the same action differently. If i does not observe d, then Γi(d,κi)=κi _i(d,κ^i)=κ^i; selective disclosure is an action observed by some DMs and not others. Announcements, exercises, leaks and bluffs are distinguished by their sets Actk(s)Act_k(s) and their interpretation maps, not by their effect on the physical state, which is none. The requirement κi≤kΓi(d,κi)κ^i _k _i(d,κ^i) is the following assumption. Assumption 4.5 (Monotone accumulation). Evidence can be added but not removed: interpretation never lowers an assessment in the knowledge order. The assumption excludes retraction, correction and forgetting of evidence, and the discrediting of sources; its consequences below, in particular the absorbing character of contradiction, depend on it. 4.3 Kinds of disturbance A disturbance κi↦κ′iκ^i κ i changes the capability assessment ψiψ^i, the intention assessment χiχ^i, or both. A change of ψiψ^i is opponent-side if it concerns only options outside OiO_i and own-side if it concerns only options in OiO_i. A change of either assessment is a for-disturbance if it raises only f-components and an against-disturbance if it raises only a-components. Since the components are raised independently, every disturbance is a composite, in any order, of a for-part and an against-part, each of which splits further into opponent-side and own-side parts and into capability and intention parts. These componentwise updates commute at the level of epistemic states and lead to the same final assessment; the intermediate perceived models may differ with the order (for instance, → U→ Y→ YN and → U→ N→ YN). 5 How reduction operators respond to evidence Proposition 5.1 (Direction of evidence). For every reduction operator δ and every x∈x 4: raising the f-component of x can change δ(x)δ(x) only from N to Y, and raising the a-component can change it only from Y to N. Proof. Raising f is an increase in the truth order and raising a a decrease, and every reduction operator is monotone in the truth order. Directly: in each of the four evidence formulas, the condition for Y is preserved by raising f and by lowering a. ∎ Proposition 5.2 (Sensitivity). δ(x) _ N Y(x) depends only on the f-component of x, and δ(x) _ Y N(x) only on the a-component. Each of δ _ N N and δ _ Y Y depends on both components. Proof. The first two claims are read off the evidence formulas. For δ _ N N, raising f at U changes the value from N to Y, and raising a at Y changes it from Y to N. For δ _ Y Y, raising f at N changes it from N to Y, and raising a at U from Y to N. ∎ Since a disturbance raises both components in general, these facts recover the behaviour of the operators in the knowledge order: δ _ N Y is monotone, δ _ Y N is antitone, and δ _ N N and δ _ Y Y are neither. This is the pattern behind the failure of knowledge monotonicity for three of the four operators in [17]. Proposition 5.3 (Immune assessments). Call x∈x 4 immune for δ if δ(y)=δ(x)δ(y)=δ(x) for every y with x≤kyx _ky. Then U is immune for no operator, YN is immune for every operator, Y is immune iff δ()=δ( YN)= Y, and N is immune iff δ()=δ( YN)= N. Hence the immune values are x:f=1\x:f=1\ for δ _ N Y and δ _ Y Y, and x:a=1\x:a=1\ for δ _ Y N and δ _ N N. Proof. The upper sets are 4 for U, \ YN\ for YN, ,\ Y, YN\ for Y and ,\ N, YN\ for N, and δ is not constant on 4. ∎ Under Assumption 4.5, ignorance is always open to further evidence, and a contradictory assessment is absorbing: it cannot be changed by any further evidence. This is a consequence of monotone accumulation, not of four-valued logic alone. A definite assessment resists further evidence exactly when the DM resolves contradiction in the same direction: a DM that treats contested evidence as availability cannot be talked out of believing that a move is available, but can be talked into believing it; the dual DM the other way round. Lemma 5.4 (Immune warrants). Let a be an arc and ψ≤kψ′ψ _kψ . If ψ(o)ψ(o) is immune for δ for every o∈req(a)o (a), then δ(wψ′(a))=δ(wψ(a))δ(w_ψ (a))=δ(w_ψ(a)). Proof. If req(a)=∅req(a)= both warrants are Y. Otherwise, if δ()=δ( YN)= Y, every ψ(o)ψ(o) and every ψ′(o)ψ (o), o∈req(a)o (a), has f=1f=1 by Proposition 5.3; by Lemma 3.1 both warrants have f=1f=1, and δ _ N Y and δ _ Y Y send every such value to Y. If δ()=δ( YN)= N, the same argument applies with a=1a=1. ∎ 6 Stability and move sets Lemma 6.1 (Monotonicity of reachability). Let G and G′G have the same states and preferences, let i∈Ni∈ N, and suppose that Aj⊆Aj′A_j A _j for every j≠ij≠ i. Then RN−i(s)⊆RN−i′(s)R_N-i(s) R _N-i(s) and RN−i+(s)⊆RN−i′+(s)R^+_N-i(s) R +_N-i(s) for every s. The same holds for reachability along any family of designated arc sets with Bj⊆Bj′B_j B _j for every j≠ij≠ i. Proof. A legal sequence in G is a legal sequence in G′G : its steps are arcs of G′G , and its labels and preference comparisons are unchanged. ∎ Lemma 6.2 (Stability in move sets). Let G and G′G have the same states and preferences, and let i∈Ni∈ N and σ∈σ . (a) If Ai=Ai′A_i=A _i and Aj⊆Aj′A_j A _j for all j≠ij≠ i, then Stabiσ(G)⊆Stabiσ(G′)Stab^σ_i(G) ^σ_i(G ), with equality for Nash. (b) If Aj=Aj′A_j=A _j for all j≠ij≠ i and Ai⊆Ai′A_i A _i, then Stabiσ(G′)⊆Stabiσ(G)Stab^σ_i(G ) ^σ_i(G). The same holds for SEQ judgements with the credible arc sets BjB_j in place of the improving arcs of j≠ij≠ i. Proof. (a) The improvements Ri+(s)R_i^+(s) and the escape sets Ri(s′)R_i(s ) depend only on AiA_i; by Lemma 6.1 the sets in which sanctions are sought can only grow, so every sanction persists. Nash stability involves AiA_i only. (b) Now Ri+(s)R_i^+(s) and every Ri(s′)R_i(s ) can only grow, while the sets in which sanctions are sought are unchanged; every condition requires a sanction for every improvement, and SMR also requires Ri(s′)⊆Φi≃(s)R_i(s ) _i(s), so both become harder to satisfy. ∎ 7 The direction of disturbance Theorem 7.1 (Direction for capability). Let DM i use δi _i, let κi↦κ′iκ^i κ i be a disturbance that changes only the capability assessment, let σ∈σ , and write Stab=Stabiσ[κi]Stab=Stab^σ_i[κ^i] and Stab′=Stabiσ[κ′i]Stab =Stab^σ_i[κ i]. (a) Opponent-side for-disturbance: Stab⊆Stab′Stab . (b) Opponent-side against-disturbance: Stab′⊆StabStab . (c) Own-side for-disturbance: Stab′⊆StabStab . (d) Own-side against-disturbance: Stab⊆Stab′Stab . For Nash stability, (a) and (b) hold with equality. Proof. Under a for-disturbance each changed assessment has its f-component raised and its a-component unchanged; by Lemma 3.1 the same holds for every warrant, and by Proposition 5.1 every perceived arc remains perceived. Hence Aj[κi]⊆Aj[κ′i]A_j[κ^i] A_j[κ i] for every j, with equality for the DMs whose options are unaffected. Since the intention assessment is unchanged, it follows that Bj[κi]⊆Bj[κ′i]B_j[κ^i] B_j[κ i] for every j as well. For an against-disturbance the inclusions are reversed. The four cases follow from Lemma 6.2. ∎ Corollary 7.2 (Operator-specific direction). Let κi↦κ′iκ^i κ i be an arbitrary opponent-side disturbance of the capability assessment. If i uses δ _ N Y, then Stab⊆Stab′Stab ; if i uses δ _ Y N, then Stab′⊆StabStab . For own-side disturbances the inclusions are reversed. For δ _ N N and δ _ Y Y no such conclusion holds in general. Proof. Decompose the disturbance into a for-part and an against-part. Under δ _ N Y perceived arcs depend only on f-components (Lemmas 3.1 and 5.2), so the against-part has no effect; the case δ _ Y N is symmetric. Example 8.3 shows both directions for the other two operators. ∎ Theorem 7.3 (Direction for intention). Let κi↦κ′iκ^i κ i be a disturbance that changes only the intention assessment. (a) i’s Nash, GMR and SMR judgements do not change. (b) Under a for-disturbance, StabiSEQ[κi]⊆StabiSEQ[κ′i]Stab^SEQ_i[κ^i] ^SEQ_i[κ i]; under an against-disturbance the reverse inclusion holds. (c) Under δ _ N Y every such disturbance can only enlarge the set of states judged SEQ; under δ _ Y N it can only shrink it. Proof. (a) These judgements do not involve χiχ^i. (b) By Proposition 5.1, a for-disturbance can only enlarge each Bj[κi]B_j[κ^i], j≠ij≠ i, and an against-disturbance only shrink it, while Ai[κi]A_i[κ^i] and the improvements of i are unchanged; conclude by Lemma 6.2. (c) Follows from (b) and Proposition 5.2. ∎ Capability and intention thus act on different parts of a stability certificate. Assessments of capability determine which moves exist in the perceived model. Through the focal DM’s own options they affect its improvements and hence every concept, Nash stability included; through the opponents’ options they affect the sanctioning paths and hence every sanction-based concept, but not Nash stability. Assessments of intention determine only which of the opponents’ moves count as credible, and affect only SEQ. The same two dimensions separate structural admissibility from preferential credibility in the analysis of state exclusion [16, Sec. 6]. Assessment disturbed Concepts affected Counterpart in [16] Own capability (ψ on OiO_i) Nash, GMR, SMR, SEQ (available improvements and escapes) — Opponents’ capability (ψ on O∖OiO O_i) GMR, SMR, SEQ; Nash unchanged (existence of sanctioning paths) structural admissibility Opponents’ intention (χ) SEQ only (credibility of sanctioning paths) preferential credibility Table 1: Effects of capability and intention assessments on stability concepts. For a state-preserving action d of k, the direction of its effect on an observer i is obtained by decomposing the induced disturbance κi↦Γi(d,κi)κ^i _i(d,κ^i) into its parts and applying Theorems 7.1 and 7.3 to each. Proposition 7.4 (Judgements immune to every action). Let s∈Ss∈ S, and let O^(s) O(s) be the set of options in the requirements of arcs of G whose tail lies in D^G(s) D_G(s). If ψi(o)ψ^i(o) is immune for δi _i for every o∈O^(s)o∈ O(s), and χi(a)χ^i(a) is immune for δi _i for every arc a of another DM with tail in D^G(s) D_G(s), then no state-preserving action of any DM changes any of i’s stability judgements at s. Proof. By Lemma 5.4 and Proposition 5.3, every arc with tail in D^G(s) D_G(s) keeps its status as perceived and as credible under every disturbance. Perceived arcs are arcs of G, so the forward-reachable sets of s in the perceived models before and after the disturbance are both contained in D^G(s) D_G(s), and they coincide with each other by induction on path length. The localization lemma gives the claim. ∎ 8 Actions that enable provocation or deterrence A state-preserving action does not change the physical state and therefore is not an improvement in the underlying preference relation on S. Its strategic value lies instead in the change it induces in the stability judgements of others. We define this change without assuming how the others then choose among their moves; the choice of responses is treated in [14]. Definition 8.1 (Provocation- and deterrence-enabling actions). Let k≠ik≠ i, s∈Ss∈ S, d∈Actk(s)d _k(s), σ∈σ , and κ′i=Γi(d,κi)κ i= _i(d,κ^i). Write Ri+[κi](s)R_i^+[κ^i](s) for i’s improvements from s in G[κi]G[κ^i]. (a) d is provocation-enabling for k against i at s if s∈Stabiσ[κi]∖Stabiσ[κ′i]s ^σ_i[κ^i] ^σ_i[κ i] and some t∈Ri+[κ′i](s)t∈ R_i^+[κ i](s) satisfies t≻kst _ks. (b) d is deterrence-enabling for k against i at s if s∈Stabiσ[κ′i]∖Stabiσ[κi]s ^σ_i[κ i] ^σ_i[κ^i] and s≻kts _kt for every t∈Ri+[κi](s)t∈ R_i^+[κ^i](s). A provocation-enabling action turns a state that i judges stable into one it judges unstable, where at least one of i’s improvements is welcome to k; whether i takes that improvement is not asserted. A deterrence-enabling action turns a state that i judges unstable into one it judges stable, where k prefers the state to every improvement i was contemplating. The true availability of the options and the true preferences of the DMs concerned play no role: an action works through i’s interpretation, so bluffs are included. A part of a disturbance of i is effective if applying it changes at least one of i’s perceived arc sets Aj[κi]A_j[κ^i] or perceived credible arc sets Bj[κi]B_j[κ^i]. Corollary 8.2 (Channels). Decompose the disturbance induced by d into its parts. Then: (a) A provocation-enabling action induces at least one effective part that is an opponent-side against-part (of capability or intention) or an own-side for-part of capability; a deterrence-enabling action induces at least one effective part that is an opponent-side for-part or an own-side against-part of capability. (b) Under δ _ N Y against-parts are never effective, so provocation can be enabled only through evidence for i’s own options, and deterrence only through evidence for its opponents’ options or intentions. (c) Under δ _ Y N for-parts are never effective, so provocation can be enabled only through evidence against its opponents’ options or intentions, and deterrence only through evidence against its own options. (d) Under δ _ N N and δ _ Y Y all channels are possible. In particular, if k’s actions can generate in i only evidence about options and intentions of DMs other than i, then an observer using δ _ N Y can be subjected to deterrence-enabling but never to provocation-enabling actions, and an observer using δ _ Y N the reverse. Proof. The updates of the parts commute on evidence pairs and lead to the same final epistemic state, whatever the order; fix any order of application. By Theorems 7.1 and 7.3, opponent-side for-parts and own-side against-parts can only enlarge the set of states i judges stable, and the other parts can only shrink it. Parts that are not effective change nothing. Applying the parts one after another in the fixed order, a state can leave the set of states judged stable only at a step that applies an effective shrinking part, and enter it only at a step that applies an effective enlarging part. Parts (b) and (c) follow from Proposition 5.2, and (d) from Example 8.3. ∎ Example 8.3 (A threat and its assessment). Let N=1,2,3N=\1,2,3\ and S=s,t,uS=\s,t,u\. DM 1 can move from s to t unconditionally; DM 2 can move from t to u, a threat whose requirement is an option o∈O2o∈ O_2; DM 3 has no moves. Preferences satisfy t≻1s≻1ut _1s _1u, s≻2ts _2t and t≻3st _3s; the other comparisons play no role. The graph is shown below; both arcs are irreversible, and the dotted arc is the one whose availability DM 1 assesses. ssttuuDM 1DM 2 (o) DM 1 judges s GMR-stable exactly when it perceives the threat, that is, when δ1(ψ1(o))= _1(ψ^1(o))= Y; the same holds for SMR, since DM 1 has no move from u. DM 2 has an action d2d_2 (an exercise) that DM 1 reads as evidence for o, and DM 3 an action d3d_3 (a leak) that DM 1 reads as evidence against. ψ1(o)ψ^1(o) Action and induced change δ1()= _1( YN)= Y (δ _ N Y, δ _ Y Y) δ1()= _1( YN)= N (δ _ Y N, δ _ N N) N (threat not believed) d2d_2: → N→ YN threat perceived; d2d_2 is deterrence-enabling for DM 2 no effect Y (threat believed) d3d_3: → Y→ YN no effect threat no longer perceived; d3d_3 is provocation-enabling for DM 3 Table 2: Effect of state-preserving actions on DM 1’s judgement at s. Starting from U, evidence for (→ U→ Y) makes the threat perceived under δ _ N Y and δ _ N N, and evidence against (→ U→ N) removes it under δ _ Y Y and δ _ Y N. Under δ _ N N both deterrence (→ U→ Y) and provocation (→ Y→ YN) can be enabled, and under δ _ Y Y both deterrence (→ N→ YN) and provocation (→ U→ N). As an implementation check of the definitions, rather than as additional evidence for the proofs, the operator-level facts of Section 5 were recomputed exhaustively over 4, and Theorem 7.1 was tested on 3000 randomly generated three-DM graph models with option requirements and complete, possibly intransitive preferences, for all four operators, all four stability concepts and random for- and against-disturbances on both sides, with no violation in 2.4×1052.4× 10^5 checks; Theorems 7.3 and 9.3 were tested on 2000 further models with no violation in 2×1052× 10^5 checks. 9 Types and hedging Observers often assess a party not move by move but by asking which side it is on. We model this by candidate types and derive intention assessments from evidence about types. Definition 9.1 (Types, live types and derived intention). Let k≠ik≠ i. A set of candidate types of k is a finite set Θk _k of complete relations on S, each a preference that k might have. A type assessment of i is a map τ:Θk→τ _k 4, and the live types are L=θ∈Θk:δi(τ(θ))=L=\θ∈ _k: _i(τ(θ))= Y\. The live types induce the intention assessment of k’s moves χL(u,v)=(f,a),f=1⇔v≻θu for some θ∈L,a=1⇔v≻θu fails for some θ∈L. _L(u,v)=(f,a), aligned f=1& v _θu for some θ∈ L,\\ a=1& v _θu fails for some θ∈ L. aligned The observer applies its operator twice: to decide which types are live, and to decide which of k’s moves are credible. If no candidate type is live, the induced intention assessment is U for every move of k, representing the absence of any surviving basis for a judgement either way; the observer’s operator then decides how such moves are treated. Evidence against a type matters as far as the observer’s operator lets it: an assessment Y that becomes YN keeps the type live under δi()= _i( YN)= Y and removes it under δi()= _i( YN)= N. If exactly one type θ is live, we say that k has revealed θ to i; if exactly two types θ1,θ2 _1, _2 are live, k is hedging between them in the eyes of i. Proposition 9.2 (What hedging contests). If k hedges between θ1 _1 and θ2 _2, then χL(u,v) _L(u,v) is Y if v is an improvement over u under both types, N if under neither, and YN if under exactly one; the contested assessments are absorbing under Assumption 4.5. Proof. Immediate from Definition 9.1 with L=θ1,θ2L=\ _1, _2\ and from Proposition 5.3. ∎ Theorem 9.3 (Hedging). Let i use δi _i and let k≠ik≠ i hedge between θ1 _1 and θ2 _2. Let hedgeJ_hedge, θ1J_ _1 and θ2J_ _2 be the sets of states i judges SEQ for itself when k hedges, reveals θ1 _1 or reveals θ2 _2, that is, when the live set is θ1,θ2\ _1, _2\, θ1\ _1\ or θ2\ _2\, all other assessments being fixed. (a) If δi()= _i( YN)= Y, k’s perceived credible arcs under hedging are the union of those under the two revelations, and hedging weakly expands i’s set of SEQ-stable states: θ1∪θ2⊆hedgeJ_ _1 _ _2 _hedge. (b) If δi()= _i( YN)= N, they are the intersection, and hedging weakly shrinks it: hedge⊆θ1∩θ2J_hedge _ _1 _ _2. Proof. By Proposition 9.2, the intention assessments under hedging and under the revelations differ only on disputed moves, which are YN under hedging and Y under one revelation and N under the other. The statements on arcs follow, and the inclusions follow from the monotonicity of SEQ judgements in credible arcs (Lemma 6.2). ∎ Theorem 9.3 compares the observer’s judgements, not the hedger’s payoffs, and makes no assumption about k’s true preferences. Example 9.4 (Two powers and a hedging party). There are three DMs, N=P1,P2,HN=\P_1,P_2,H\, and five states, S=s,t1,t2,u1,u2S=\s,t_1,t_2,u_1,u_2\, with status quo s. Power P1P_1 can move from s to t1t_1 and power P2P_2 from s to t2t_2, each an encroachment it prefers. The third party H can respond to t1t_1 by moving to u1u_1, siding with P2P_2, and to t2t_2 by moving to u2u_2, siding with P1P_1. Preferences are t1≻P1s≻P1t2≻P1u2≻P1u1,t2≻P2s≻P2t1≻P2u1≻P2u2,t_1 _P_1s _P_1t_2 _P_1u_2 _P_1u_1, t_2 _P_2s _P_2t_1 _P_2u_1 _P_2u_2, and s≻Ht1≻Ht2≻Hu1≻Hu2s _Ht_1 _Ht_2 _Hu_1 _Hu_2: H ranks s first and prefers staying at t1t_1 and t2t_2 to siding with either power. The graph is shown below; all arcs are irreversible. sst1t_1t2t_2u1u_1u2u_2P1P_1P2P_2H The candidate types of H are θA _A, under which siding with P1P_1 is an improvement and siding with P2P_2 is not, and θB _B, the reverse. observers with δ()=δ( YN)= Y observers with δ()=δ( YN)= N H reveals θA _A P2P_2 judges s SEQ, P1P_1 not P2P_2 judges s SEQ, P1P_1 not H reveals θB _B P1P_1 judges s SEQ, P2P_2 not P1P_1 judges s SEQ, P2P_2 not H hedges both judge s SEQ neither Table 3: SEQ judgements of the status quo in Example 9.4. Against observers who read contradiction as credibility, hedging makes both powers judge the status quo SEQ-stable, although in the true model it is SEQ-stable for neither, since H’s responses are not improvements for H. It is GMR-stable for both in the true model, because GMR asks only whether H can respond, not whether it would. 10 Illustration: the 1995 DVD format negotiation 10.1 Background In late 1994 two incompatible high-density optical disc formats were announced: the Multimedia Compact Disc (MMCD) of Philips and Sony, and the Super Density (SD) disc of Toshiba, Matsushita, Time Warner and others [32, 25]. On 3 May 1995, technical representatives of Apple, Compaq, Hewlett-Packard, IBM and Microsoft publicly urged the two camps to converge on a single format, reporting that they had met each camp separately in April and urged it to present one [28]. The group, in which IBM played a prominent role, was open to further computer companies, and a near-contemporaneous account names additional members [28, 29]; we refer to it as the Technical Working Group (TWG). Both camps regarded its support as critical for the adoption of their format in personal computers [25]. A general agreement on the basic specifications of a single format was reached on 15 September 1995 [31, 33, 34, 30], and the final details were announced on 8 December 1995 [33]. The accounts agree that the TWG, with IBM prominent, exerted pressure on both camps, but they differ on the signalling in the summer of 1995. A later retrospective reports that by late July IBM let the camps know that it was ready to endorse SD, and that the SD camp, which regarded its format as the likely winner, was reluctant until IBM informed both camps in September that it would endorse the compromise [25]. A near-contemporaneous account reports audience-specific communications: Sony and Philips were told that IBM was about to adopt the longer SD format and were given a few weeks to reach a compromise, while Toshiba’s designers were told that Sony’s modulation and storage techniques were preferable [29]. We use the case to illustrate the distinction between capability and intention, not to establish the causes of the outcome. As shown below, the framework accommodates either account of the signalling. 10.2 Model Following the usual presentation of a graph model, we specify the decision makers and their options, the feasible states, the transitions with their reversibility, and the preferences. Decision makers and options. The DMs are the MMCD camp M, the SD camp S and the TWG T, with the options listed in Table 4. The TWG’s public position was a call for a single format [28], and its formal stance was initially to endorse neither format and to issue non-binding recommendations; its influence rested on the de facto market power of its members [25]. The option B stylises this as an explicit refusal of support for a format launched by one camp alone; it is not a literal boycott, which the sources do not report. DM Option Meaning M LML_M launch the MMCD format unilaterally AMA_M accept a unified format S LSL_S launch the SD format unilaterally ASA_S accept a unified format T EME_M endorse MMCD ESE_S endorse SD B refuse support to any single format launched unilaterally ECE_C endorse a compromise accepted by a camp Table 4: Decision makers and options of the DVD model. Feasible states. A camp takes at most one of its options: accepting a unified format supersedes a launch. The TWG takes at most one of its options. In addition, ECE_C requires that some camp has accepted, and T does not endorse a format whose own camp has accepted a unified format. When both camps have accepted, the format is unified and the conflict is resolved; all such configurations are identified with a single state, in which no statement of T is recorded. These constraints leave 3333 feasible states, numbered as in Fig. 2 and listed in option form in Table 6 (Appendix A). The status quo is state 11, in which no option is taken, and the unified format is state 3333. Transitions and reversibility. Figure 1 shows the option-level transitions of each DM. A camp can launch, or accept directly, from the status quo, and can accept after launching. The TWG can make one statement, while the format is not unified. All transitions are irreversible: no launch, acceptance or statement can be withdrawn. Figure 2 shows the resulting integrated graph on the 33 feasible states, with 8888 arcs. MMnoneLML_MAMA_MSSnoneLSL_SASA_STTnoneEME_MESE_SBBECE_C Figure 1: Option-level transitions of the three DMs. Every arrow is one-way: all transitions are irreversible. EME_M (ESE_S) is unavailable once M (S) has accepted, ECE_C requires that a camp has accepted, and no DM moves once the format is unified. M: noneMM: launchedMM: acceptedSS: noneSS: launchedSS: accepted123456789101112131415161718192021222324252627282930313233move of MMmove of SSmove of T Figure 2: Integrated graph of the DVD model. States are grouped by the positions of the two camps (columns: M; rows: S); within a group, the centre position is the state in which T has made no statement, and the other positions correspond to EME_M (upper left), ESE_S (right), B (lower left) and ECE_C (lower right). All arcs are irreversible. The status quo is state 11 and the unified format state 3333; the option form of every state is given in Table 6. Preferences. Each camp most prefers to launch its format with the TWG’s endorsement and least prefers the rival’s endorsed launch; unification lies in between and above the status quo; a launch without endorsement is valued above the status quo by the camp that launches (first-mover hope) but below it if the TWG refuses support. The complete rankings, which are weak orders, are given in Table 6; the principles from which they are constructed, and the comparisons on which the results depend, are set out in Section A.2. The TWG’s capability to endorse or to refuse support is not in doubt; what the camps assess is its intention, through five candidate types, whose rankings are also given in Table 6: θB _B (blocker: refuses support to any single launch and otherwise does nothing), θU _U (enforcer: refuses support to any single launch and endorses a compromise once one camp has accepted it), θUS _US (an enforcer that endorses SD if the stalemate persists), and the partisans θS _S and θM _M. The analyst takes T’s true preferences to be those of θU _U. The arcs that matter. The analysis below uses few of the arcs. Write ℓM _M (state 1313) for the state in which M has launched and nobody else has moved, and bMb_M, eMSe^S_M and eMMe^M_M (states 1616, 1515 and 1414) for the states reached from ℓM _M when T refuses support, endorses SD or endorses MMCD; ℓS _S, bSb_S, eSMe^M_S and eSSe^S_S (states 55, 88, 66 and 77) are defined symmetrically. The relevant arcs are s0→ℓM,ℓM→bM,ℓM→eMS,ℓM→eMM,s_0 M _M, _M Tb_M, _M Te^S_M, _M Te^M_M, and symmetrically for S, with ℓM≻Ms0≻MbM,eSM _M _Ms_0 _Mb_M,\,e^S_M and eMM≻Ms0e^M_M _Ms_0. Thus M’s launch is an improvement; it has a GMR-sanction under every assessment, because T can move to bMb_M or eMSe^S_M, both of which satisfy s0≻M⋅s_0 _M·; and it has an SEQ-sanction exactly when one of these sanctioning moves is also an improvement for T under a live type: under θB _B, θU _U and θUS _US (refusal) and under θS _S (endorsement of SD), but not under θM _M, whose best response is eMMe^M_M. The same holds for S with the roles of the partisans exchanged. The rival camp’s responses do not change this: launching its own format in reply starts a format war, which is a GMR-sanction but not an SEQ-sanction, since it is not an improvement for the camp that would start it; accepting after the rival’s launch does not lower the launcher’s position below the status quo and is no sanction at all. 10.3 Capability and intention in the case Table 5 shows whether each camp judges the status quo stable against its own unilateral launch, under SEQ and under GMR, for various sets of live TWG types. Live TWG types M judges s0s_0 stable (SEQ / GMR) S judges s0s_0 stable (SEQ / GMR) θB\ _B\, θU\ _U\ or θUS\ _US\ yes / yes yes / yes θS\ _S\ yes / yes no / yes θM\ _M\ no / yes yes / yes θM,θS\ _M, _S\ yes / yes yes / yes Table 5: Stability judgements of the status quo s0s_0 against unilateral launch. Under GMR every camp judges the status quo stable under every assessment, because the TWG can always endorse the rival format or refuse support; GMR asks whether the TWG can sanction and cannot distinguish the phases of the negotiation. Under SEQ the judgement depends on whether the sanctioning move is one the TWG would make, that is, on the assessment of its intention. The row θM,θS\ _M, _S\ is an instance of Theorem 9.3: a TWG that appears as a possible partisan of either side deters both, for observers who read contradiction as credibility. 10.4 State-preserving actions and their interpretations The TWG’s public statement of May 1995, its communications in the summer and its September communication change nothing in the physical state: no format is launched, endorsed or accepted by them. They are state-preserving actions of T, and the sources suggest the following interpretations. ∙ Public call for a single format (May 1995): read by both camps as evidence for θB _B. Both camps judge the status quo SEQ-stable; neither launches. ∙ Summer communications. Under the later account, a common action d observed by both camps is read differently: ΓM(d,κM) _M(d,κ^M) generates evidence for θUS _US, an enforcer with an SD fallback, and ΓS(d,κS) _S(d,κ^S) evidence for θS _S, partisan support. Under the near-contemporaneous account, T performs distinct actions dM≠dSd_M≠ d_S: dMd_M, observed by M only, generates evidence for θUS _US in M, while dSd_S conveys technical information to S, whose reading of T as a likely supporter of SD rests on its own expectation of victory. Either way, the readings are θUS\ _US\ for M and θS\ _S\ for S. With θS\ _S\ live, S no longer judges the status quo SEQ-stable against its own launch (Table 5): its launch has no credible sanction, consistent with its regarding SD as the likely winner. M continues to judge the status quo stable. ∙ September communication: an action observed by S that makes an enforcer type live for it; S again judges the status quo SEQ-stable against its own launch. The two accounts of the summer correspond to the two ways in which Definition 4.4 lets the effect of actions differ across observers: observer-specific interpretation of a common action, and selective disclosure through distinct actions. Why M then accepted a unified format, which is worse for it than the status quo in itself, requires an account of how DMs anticipate the evolution of the conflict; it is given in [14]. 10.5 Limitations The dates and the content of the public statements rest on primary sources [28, 31, 33] and contemporary reports [34, 30]. The private communications of the summer of 1995 are known only from a near-contemporaneous account, which reports them at second hand [29], and from a later retrospective [25]. The preferences (Section A.2) and the readings are reconstructions. The Hollywood studios, whose advisory committee had set requirements of its own [25], are left out, and each camp and the TWG are treated as single DMs. 11 Discussion Interpretation. The four reduction operators become profiles of vulnerability to manipulation. A DM that reads any evidence for an option as availability (δ _ N Y) can be made more cautious but cannot be emboldened by talk about its opponents; a DM that reads any evidence against as unavailability (δ _ Y N) can be emboldened but not made more cautious; the other two can be moved either way. Under monotone accumulation, contradiction is absorbing, so injecting contradictory evidence is effective only once and the ambiguity created by hedging persists. A party that makes its capabilities ambiguous acts on every sanction-based judgement of its opponents; a party that makes only its intentions ambiguous acts on credibility alone. Limitations and further work. Relating the conditions under which manipulation of preferences fails [12] to those under which manipulation of assessments succeeds is an open question. Evidence is assumed to accumulate monotonically; retraction, correction and the discrediting of sources would allow contradictory assessments to be resolved and require a different treatment. Physical moves are assumed to leave assessments unchanged. Higher-order assessments (what a DM believes about another’s assessment) and coalitional stability [18, 11] are not treated; for coalitions, disturbances raise the questions whether members share assessments and whether an action can target a coalition as a whole. A third epistemic layer, in which a state-preserving action is perceived by others as a change of physical state, would model feints in the narrow sense. The distinction between inaction and state-preserving actions also suggests a categorical treatment in which the latter are non-identity endomorphisms [15]. How DMs anticipate the evolution of the conflict under their assessments, and how an actor can plan a sequence of state-preserving actions, is developed in [14]. Finally, all results are finite and relational, which makes them natural targets for formal verification alongside the existing Lean development of [17]. Appendix A States and preferences of the DVD model A.1 States and rankings Table 6 lists the feasible states of the DVD model in option form and gives each DM’s ranking: 11 is the most preferred, and states with the same number are equally preferred. The columns for T give the rankings of the five candidate types; the true type is θU _U. Table 6: Feasible states (option form) and preference rankings of the DVD model. Y: the option is taken; –: it is not. M S T camps types of T State LML_M AMA_M LSL_S ASA_S EME_M ESE_S B ECE_C M S θB _B θU _U θUS _US θS _S θM _M 1 – – – – – – – – 6 6 3 4 4 4 4 2 – – – – Y – – – 6 6 4 5 5 6 2 3 – – – – – Y – – 6 6 4 5 3 2 6 4 – – – – – – Y – 6 6 3 4 4 5 5 5 – – Y – – – – – 8 5 5 6 6 5 5 6 – – Y – Y – – – 5 10 6 7 7 6 2 7 – – Y – – Y – – 12 1 6 7 7 1 6 8 – – Y – – – Y – 6 9 2 3 3 5 5 9 – – – Y – – – – 3 7 3 4 4 4 4 10 – – – Y Y – – – 2 10 4 5 5 6 2 11 – – – Y – – Y – 3 7 3 4 4 5 5 12 – – – Y – – – Y 10 3 4 2 2 5 5 13 Y – – – – – – – 5 8 5 6 6 5 5 14 Y – – – Y – – – 1 12 6 7 7 6 1 15 Y – – – – Y – – 10 5 6 7 7 2 6 16 Y – – – – – Y – 9 6 2 3 3 5 5 17 Y – Y – – – – – 10 10 5 6 6 5 5 18 Y – Y – Y – – – 5 12 6 7 7 6 1 19 Y – Y – – Y – – 12 5 6 7 7 1 6 20 Y – Y – – – Y – 11 11 2 3 3 5 5 21 Y – – Y – – – – 5 8 5 6 6 5 5 22 Y – – Y Y – – – 1 12 6 7 7 6 1 23 Y – – Y – – Y – 9 6 2 3 3 5 5 24 Y – – Y – – – Y 10 5 6 7 7 5 5 25 – Y – – – – – – 7 3 3 4 4 4 4 26 – Y – – – Y – – 10 2 4 5 5 2 6 27 – Y – – – – Y – 7 3 3 4 4 5 5 28 – Y – – – – – Y 3 10 4 2 2 5 5 29 – Y Y – – – – – 8 5 5 6 6 5 5 30 – Y Y – – Y – – 12 1 6 7 7 1 6 31 – Y Y – – – Y – 6 9 2 3 3 5 5 32 – Y Y – – – – Y 5 10 6 7 7 5 5 33 – Y – Y – – – – 4 4 1 1 1 3 3 A.2 Construction of the rankings The rankings in Table 6 were not elicited from the parties; they are the analyst’s reconstruction. They are built in two steps. Each state is classified by the positions of the two camps (not moved, launched or accepted the compromise) and by the statement of T (none, endorsement of MMCD, of SD or of the compromise, or refusal of support). A small number of ordinal principles, stated below, then order these classes. The two camps are treated symmetrically: the same principles apply to each, with the roles of own and rival exchanged, so that S’s rank of a state equals M’s rank of the mirror state. Table 7 gives the resulting ranks of the camps. Table 7: Ranks of the camps by situation, from the viewpoint of the own camp. Each cell gives the rank (1: most preferred) and, in brackets, the state number when the own camp is M / when it is S; –: infeasible. Statement of T Own camp Rival camp none endorses own endorses rival refusal endorses compromise not moved not moved 6 [1/1] 6 [2/3] 6 [3/2] 6 [4/4] – not moved launched 8 [5/13] 5 [6/15] 12 [7/14] 6 [8/16] – not moved accepted 3 [9/25] 2 [10/26] – 3 [11/27] 10 [12/28] launched not moved 5 [13/5] 1 [14/7] 10 [15/6] 9 [16/8] – launched launched 10 [17/17] 5 [18/19] 12 [19/18] 11 [20/20] – launched accepted 5 [21/29] 1 [22/30] – 9 [23/31] 10 [24/32] accepted not moved 7 [25/9] – 10 [26/10] 7 [27/11] 3 [28/12] accepted launched 8 [29/21] – 12 [30/22] 6 [31/23] 5 [32/24] unified (both accepted) 4 [33/33] Principles for the camps. Write c for the own camp and r for the rival. (C1) Support of the TWG. A format needed the TWG’s support for use in personal computers [28]. Hence, for fixed positions of the camps, an endorsement of c’s position is better for c than silence, and an endorsement of r’s position worse; a refusal of support for c’s launch is worse than silence and a refusal for r’s launch better. In a format war, a refusal harms both launchers and lies between silence and the endorsement of the rival. An endorsement of the compromise counts as support for the camp that has accepted it. (C2) Statements without products. When neither camp has moved, a statement of T concerns no product on the market and leaves the camps’ ranks unchanged (rank 66 throughout). (C3) Endorsed sole format. The best outcome for c is its own format on the market with the TWG’s endorsement and no rival format (rank 11); the worst is the symmetric outcome for r, together with a format war in which T endorses r (rank 1212). (C4) First-mover hope. An unendorsed launch of c’s format, without a rival format, is better for c than the status quo (rank 55 against 66), but one refused support by the TWG is worse (rank 99). This reflects the expectation of each camp that its format could win; the SD camp is reported to have regarded its format as the likely winner [25]. (C5) Unification. A unified format is better for each camp than the status quo and than an unendorsed launch, and worse than an endorsed sole format and than the rival’s unilateral abandonment of its format (rank 44). That both camps finally agreed [31, 33] is consistent with unification ranking above the status quo; its position relative to an unendorsed launch is a modelling assumption. (C6) Format war. A format war without an endorsement is worse for each camp than the status quo (rank 1010); with the endorsement of c’s format it is as good for c as an unendorsed sole launch (rank 55). (C7) Concession. Accepting the compromise while the rival holds out is worse for c than the status quo (rank 77) unless T endorses the compromise. If T does, c is ranked above unification (rank 33) and r near the bottom (rank 1010): the camp that accepted first has the TWG’s backing in the negotiation on the content of the unified format, while the hold-out is isolated. Correspondingly, the rival’s unilateral concession without such an endorsement is good for c (ranks 22–33). This principle is a modelling assumption; the sources report the pressure on the hold-out camp [25, 29] but not the camps’ valuations. Rankings of the TWG types. The rankings of the five candidate types follow from their definitions in Section 10; they describe what a camp may believe about T, not evidence about T’s actual preferences. ∙ θB _B (blocker): unification is best and the refusal of support for a launched format next (rank 22); then come the states without a launch, in which silence and refusal (rank 33) are preferred to endorsements (rank 44); a launch left unanswered (rank 55) and an endorsed launch (rank 66) are worst. ∙ θU _U (enforcer): as θB _B, except that endorsing the compromise after one camp has accepted it is ranked second only to unification (rank 22). ∙ θUS _US: as θU _U, except that endorsing SD while no format has been launched or accepted is ranked with the refusal of support for a launch (rank 33): the enforcer’s fallback when the stalemate persists. ∙ θS _S and θM _M (partisans): endorsing the favoured format is best when that format has been launched (rank 11) and next best otherwise (rank 22); unification is ranked third, silence while no format has been launched fourth, and an endorsement of the other format last. Which comparisons the results use. The stability judgements of Table 5 rest on a few comparisons, those stated in Section 10: (i) ℓM≻Ms0≻MbM,eSM _M _Ms_0 _Mb_M,\,e^S_M and eMM≻Ms0e^M_M _Ms_0 (C1, C4); (i) the best response of each type of T at ℓM _M: refusal of support for θB,θU,θUS _B, _U, _US, endorsement of SD for θS _S and of MMCD for θM _M; (i) the rival’s responses at ℓM _M are not improvements for it, since a launch in reply leads to a format war (rank 88 to 1010 for the rival, C6) and an acceptance leaves its rank unchanged (88), and an acceptance does not lower the launcher’s rank below that of the status quo (55 against 66); (iv) after eMMe^M_M the rival is at its lowest rank, so it has no improving move; and the symmetric comparisons for S. Any rankings that satisfy these comparisons give the same judgements; in particular, the remaining choices (the exact position of unification, the principle (C7) and the ties) do not affect them. The anticipatory analysis of [14] uses further comparisons, notably (C5) and (C7). References [1] Y. M. Aljefri, M. A. Bashar, L. Fang, and K. W. Hipel, “First-level hypergame for investigating misperception in conflicts,” IEEE Trans. Syst. Man Cybern.: Syst., vol. 48, no. 12, p. 2158–2175, 2018. [2] N. D. Belnap, “A useful four-valued logic,” in Modern Uses of Multiple-Valued Logic, J. M. Dunn and G. Epstein, Eds. Dordrecht: Reidel, 1977, p. 5–37. [3] S. J. Brams, “Deception in 2×22× 2 games,” Conflict Manag. Peace Sci., vol. 2, no. 2, p. 171–203, 1977. [4] V. P. Crawford and J. Sobel, “Strategic information transmission,” Econometrica, vol. 50, no. 6, p. 1431–1451, 1982. [5] J. M. Dunn, “Intuitive semantics for first-degree entailments and ‘coupled trees’,” Philos. Stud., vol. 29, p. 149–168, 1976. [6] L. Fang, K. W. Hipel, and D. M. Kilgour, Interactive Decision Making: The Graph Model for Conflict Resolution. New York: Wiley, 1993. [7] N. M. Fraser and K. W. Hipel, “Solving complex conflicts,” IEEE Trans. Syst. Man Cybern., vol. 9, no. 12, p. 805–816, 1979. [8] N. M. Fraser and K. W. Hipel, Conflict Analysis: Models and Resolutions. New York: North-Holland, 1984. [9] M. L. Ginsberg, “Multivalued logics: A uniform approach to reasoning in artificial intelligence,” Comput. Intell., vol. 4, no. 3, p. 265–316, 1988. [10] N. Howard, Paradoxes of Rationality: Theory of Metagames and Political Behavior. Cambridge, MA: MIT Press, 1971. [11] T. Inohara and K. W. Hipel, “Coalition analysis in the graph model for conflict resolution,” Syst. Eng., vol. 11, no. 4, p. 343–359, 2008. [12] T. Inohara, S. Takahashi, and B. Nakano, “Impossibility of deception in a conflict among subjects with interdependent preference,” Appl. Math. Comput., vol. 81, no. 2–3, p. 221–244, 1997. [13] Y. Kato, “State definition for conflict analysis with four-valued logic,” in Proc. IEEE Int. Conf. Syst. Man Cybern., 2022, p. 3186–3191, doi:10.1109/SMC53654.2022.9945371. [14] Y. Kato, “Anticipation and disturbance plans in the graph model for conflict resolution,” companion manuscript, 2026. [15] Y. Kato, “Categorical approach to conflict resolution: A corrected correspondence between category theory and the graph model for conflict resolution,” arXiv:2306.13961v2, 2026. [16] Y. Kato, “Genuine and spurious equilibria in the graph model for conflict resolution under effective state exclusion,” manuscript, version 3.3. [17] Y. Kato, “A four-valued graph model for conflict resolution: Core framework and a machine-checked formalization in Lean 4,” arXiv:2609.11174, 2026. [18] D. M. Kilgour, K. W. Hipel, X. Peng, and L. Fang, “Coalition analysis in group decision support,” Group Decis. Negot., vol. 10, no. 2, p. 159–175, 2001. [19] D. M. Kilgour, K. W. Hipel, and L. Fang, “The graph model for conflicts,” Automatica, vol. 23, no. 1, p. 41–55, 1987. [20] J. F. Nash, “Equilibrium points in n-person games,” Proc. Natl. Acad. Sci. USA, vol. 36, no. 1, p. 48–49, 1950. [21] J. Nash, “Non-cooperative games,” Ann. Math., vol. 54, no. 2, p. 286–295, 1951. [22] A. Obeidi, D. M. Kilgour, and K. W. Hipel, “Perceptual graph model systems,” Group Decis. Negot., vol. 18, p. 261–277, 2009. [23] L. C. Rêgo and G. I. A. Vieira, “Interactive unawareness in the graph model for multilateral conflicts,” Discrete Appl. Math., vol. 318, p. 31–46, 2022. [24] M. Wang, K. W. Hipel, and N. M. Fraser, “Solution concepts in hypergames,” Appl. Math. Comput., vol. 34, no. 3, p. 147–171, 1989. [25] S. Wolpin, “The history of DVD,” 2015 (reprint of an article in DVD Etc., 2007), https://stewartwolpin.com/the-history-of-dvd/. [26] H. Xu, D. M. Kilgour, K. W. Hipel, and G. Kemkes, “Using matrices to link conflict evolution and resolution in a graph model,” Eur. J. Oper. Res., vol. 207, no. 1, p. 318–329, 2010. [27] H. Zhang, R. Zhang, J. Liu, and Y. Zhu, “Strategic manipulation behavior in graph model for conflict resolution: The role of social trust network,” Inf. Sci., vol. 686, art. 121313, 2025, doi:10.1016/j.ins.2024.121313. [28] Apple Computer, “Requirements for future high-capacity compact-disc format announced by computer industry technical group,” press release, 3 May 1995. Archived at https://web.archive.org/web/19981202113012/http://product.info.apple.com/pr/press.releases/1995/q3/950503.pr.rel.cd.html. [29] B. Doyle, “A Digital Video Disc standard for the 21st century,” Desktop Video Group, undated [c. 1996], https://w.dtvgroup.com/DigVideo/DVD/DVDstory.html. [30] A. Harmon and D. Holley, “Agreement reached on a new format for video: Technology: Toshiba and Time Warner end battle with Sony and Philips on disc for use in devices to replace VCRs,” Los Angeles Times, 16 September 1995. [31] Nokia, “Nokia welcomes single standard for next generation high density optical disc format,” press release, 26 September 1995. Archived at https://web.archive.org/web/19961220013314/http://w.nokia.com/news/news_htmls/nokia_950926.html. [32] J. Taylor, DVD Demystified: The Guidebook for DVD-Video and DVD-ROM. New York: McGraw-Hill, 1998. [33] Toshiba Corporation, “DVD format unification,” press release, 8 December 1995, https://w.global.toshiba/w/news/corporate/1995/12/pr0802.html. [34] United Press International, “Sony, Toshiba agree on DVD format,” 16 September 1995. [35] Y. M. Aljefri, K. W. Hipel, and L. Fang, “General hypergame analysis within the graph model for conflict resolution,” Int. J. Syst. Sci.: Oper. Logist., vol. 7, no. 1, 2020, doi:10.1080/23302674.2018.1476604. [36] B. Gharesifard and J. Cortés, “Stealthy deception in hypergames under informational asymmetry,” IEEE Trans. Syst. Man Cybern.: Syst., vol. 44, p. 785–795, 2014. [37] N. S. Kovach, A. S. Gibson, and G. B. Lamont, “Hypergame theory: A model for conflict, misperception, and deception,” Game Theory, vol. 2015, art. 570639, 2015.