One’s Modus Ponens: Modality, Coherence and Logic∗†

Forthcoming in Philosophy and Phenomenological Research

Una Stojnić
Department of Philosophy and Center for Cognitive Science,
Rutgers University
October 3, 2016

Abstract
Recently, there has been a shift away from traditional truth-conditional accounts of mean-
ing towards non-truth-conditional ones, e.g., expressivism, relativism and certain forms of
dynamic semantics. Fueling this trend is some puzzling behavior of modal discourse. One
particularly surprising manifestation of such behavior is the alleged failure of some of the
most entrenched classical rules of inference; viz., modus ponens and modus tollens. These
revisionary, non-truth-conditional accounts tout these failures, and the alleged tension be-
tween the behavior of modal vocabulary and classical logic, as data in support of their
departure from tradition, since the revisionary semantics invalidate some of these patterns.
I, instead, offer a semantics for modality with the resources to accommodate the puzzling
data while preserving classical logic, thus affirming the tradition that modals express or-
dinary truth-conditional content. My account shows that the real lesson of the apparent
counterexamples is not the one the critics draw, but rather one they missed: namely, that
there are linguistic mechanisms, reflected in the logical form, that affect the interpretation
of modal language in a context in a systematic and precise way, which have to be captured
by any adequate semantic account of the interaction between discourse context and modal
vocabulary. The semantic theory I develop specifies these mechanisms and captures pre-
cisely how they affect the interpretation of modals in a context, and do so in a way that
both explains the appearance of the putative counterexamples and preserves classical logic.

∗ Acknowledgments: Earlier versions of this paper were presented at various conferences and departmental

colloquia. I would like to thank all of these audiences. I am indebted to Chris Barker, Justin Bledin, Elisabeth
Camp, Simon Charlow, Andy Egan, Graeme Forbes, Thony Gillies, Gil Harman, John Hawthorne, Ben Holguin,
Nathan Howard, Angelika Kratzer, Barbara Partee, Jim Pryor, Craige Roberts, Timothy Williamson, and two
anonymous referees for Philosophy and Phenomenological Research for comments on earlier drafts. I am especially
indebted to Jeff King, Ernie Lepore and Matthew Stone for numerous discussions from the very early stages of
this project and extensive comments on previous drafts.
† The paper is structured in two parts, followed by a formal appendix. In Part I (§ 1), I develop my account

of modality. Part II (§ 2) describes the formal tools that allow us to build the semantics described in Part I.
A formal appendix, (§ A) provides the complete formalization, along with a proof that the semantics preserves
classical logic. Part II and the appendix should thus be seen as complementing Part I, by providing a formal
implementation of the account developed in Part I.

1
0 Introduction
Recently, there has been a shift away from traditional truth conditional accounts of meaning
towards non-truth-conditional ones, e.g., expressivism, relativism and certain forms of dynamic
semantics. Fueling this trend are some puzzling behavior of modal discourse. One particularly
surprising manifestation of such behavior is the alleged failure of some of the most entrenched
classical rules of inference; viz., modus ponens (MP) and modus tollens (MT). Thus, several
authors have independently touted counterexamples to MP and MT.1 Because each challenge
arises in the presence of modal language (typically involving embedded modals or conditionals),
these critics believe they have uncovered a tension between the behavior of modal vocabulary and
classical logic, with the moral being to revise the semantics for modality in order to invalidate
certain classical patterns. These revisionary, non-truth-conditional accounts tout the alleged
tension between the behavior of modal vocabulary in natural language and classical logic, as data
in support of their departure from tradition, since the revisionary semantics invalidate some of
these patterns.2 I, instead, offer a semantics for modality with the resources to accommodate the
puzzling data while preserving classical logic, thus affirming the tradition that modals express
ordinary truth-conditional content. My account shows that the real lesson of the apparent
counterexamples is not the one the critics draw, but rather one they missed: namely, that there
are linguistic mechanisms, reflected in the logical form of a discourse, that affect the interpretation
of modal language in a context in a systematic and precise way, which have to be captured by any
adequate semantic account of the interaction between discourse context and modal vocabulary.
I develop and defend a theory that captures this interaction: according to my account, context
affects the interpretation of modals through mechanisms that are independently motivated, and
are required to explain the effects of context on the interpretation of context-sensitive expressions
quite generally, even in the most basic case—the case of pronouns. In particular, I shall argue,
the relevant mechanisms are the ones that govern how individual utterances are organized to
form a coherent discourse. While the impacts of these mechanisms explain (and predict) the
appearance of counterexamples, the underlying logic, as I shall prove, is classical.3
In short: I defend a theory that captures how context affects the interpretation of modals,
and does so in a way that reconciles classical logic with the semantics of modal language. My
1 See in particular McGee (1985) and Willer (2010) for counterexamples to MP, and Yalcin (2012) and Veltman

(1985) for counterexamples to MT. See also Kolodny and MacFarlane (2010) who reject MP as a reaction to certain
puzzling behavior of deontic modals, and Cantwell (2008) who argues that MP, MT and reasoning by cases all fail
when they involve modals in the consequents of conditionals. The revisionary accounts are typically non-truth-
conditional, insofar as they deny that an utterance expresses a proposition that is true or false depending on the
way the world is. In other words, these accounts deny utterances express propositional content. An exception to
this is McGee (1985) , who rejects MP, and thus endorses a non-classical account, but one that is nevertheless
not non-truth-conditional. In this respect, it differs significantly from these other accounts, which invalidate the
relevant patterns of inference in part by rejecting truth-conditionality of modal vocabulary.
2 This is the course explicitly taken by Yalcin (2012), who appeals to his counterexample as a basis for a

rejection of MT. Some relativists also reject MP and/or MT based on the puzzling behavior of modal language
(Kolodny and MacFarlane, 2010). And some dynamic semanticists, too, invalidate MT, where the diagnosis for
rejection lies once again in the behavior of modal language (Gillies, 2010, 2004). I should note that, though the
example I will focus on involves epistemic modality, the problem arises for other flavors of modality, as well; in
particular, deontic modals, too, exhibit the behavior that prima facie violates classical patterns of inference. (See
e.g. Kolodny and MacFarlane (2010).) The theory I will develop is not limited to epistemic modality, and my
treatment of the apparent failure of classical patterns of inference naturally extends to other examples involving
other types of modality, as well.
3 To be precise, the underlying logic is a classical modal logic—a simple extension of classical logic with a

modal necessity operator (in particular, system S4 .) This logic preserves a classical inference system, and is
sound and complete. Importantly, unlike another famous attempt at a defense of classical logic, pioneered by H.
P. Grice (1989), I shall not be arguing that the truth-conditions of the English indicative conditional are those of
the material conditional, but rather those of the strict conditional.

2
strategy will be to focus on one particular case, due to Yalcin, but the account I develop naturally
accommodates other known cases as well.4 The paper divides into two parts. In part § 1, I
argue that a systematic impact of discourse context on the interpretation of modals explains
the apparent counterexample, through general mechanisms, while preserving the validity of MT,
and I sketch an account of modality and context-change that captures this interplay between
discourse context, and modal expressions. In part § 2, I demonstrate that we can formalize this
systematic impact, in a way that explains the counterexample, and preserves classical logic.

1 Part I
1.1 The Challenge to Modus Tollens
The pattern [If p, q; ¬q ∴ ¬p] is known as modus tollens. Yalcin argues for its invalidity as
follows:

Take an urn with a 100 marbles. 10 of them are big and blue, 30 big and red, 50 small
and blue, and 10 are small and red. One marble is randomly selected and hidden
(you do not know which). Given this setup, (1) and (2) are licensed:

(1) If the marble is big, then it is likely red.

(2) The marble is not likely red.
But, surely, (3) does not follow:
(3) So, the marble is not big.
Since the inference from (1) and (2) to (3) looks to be an application of MT, we would seem to
have a counterexample.5
Possible reactions to the apparent counterexample are:
i embrace the invalidity of MT, or
ii deny that (1)–(3) is a genuine instance of MT.

(i) requires modifying our semantics for modals in such a way so as to invalidate MT. This is
Yalcin’s option, one that embraces a “revisionary” semantics for modal language that denies that
modal and conditional sentences express ordinary truth-conditional content.6 It is important to
note that these revisionary frameworks do not deny that, intuitively, the big premise in (1)–(3) is
(in some sense) about conditional probability, i.e., that it concerns the probability of the marble
being red, conditional on it being big, while the small premise is, intuitively, ‘unrestricted’ in
this sense. Everyone in the debate concedes this. What the revisionists deny, however, is that
4 E.g., the apparent counterexamples due to McGee (1985) and Kolodny and MacFarlane (2010).
5 Itis easy to construct other counterexamples along similar lines. See Veltman (1985) for an earlier counterex-
ample to MT with right-nested conditionals. I shall focus on Yalcin, but my considerations extend to Veltman.
6 See e.g. Yalcin (2012), Yalcin (2007) and Moss (2015) for an expressivist version of revisionary semantics, see

e.g. Kolodny and MacFarlane (2010) for a relativist version, and see Gillies (2010, 2004) for dynamic semantics
version. These semantics are ‘revisionary’ precisely insofar as they invalidate classical patterns of inference, and
deny that modal discourse expresses standard truth-conditions. It is worth noting that, although Yalcin revises
the standard compositional semantics for modals and conditionals, he defends expressivism as a pragmatic thesis.
In particular, he does not hold that the semantic content of a modal utterance is the informational content the
utterance expresses; in fact, he denies that modal utterances express any informational content. The semantics
I will develop and defend will require a substantial departure from the traditional semantics for modal discourse
(as in Kratzer (1977), and Kratzer (1983)), but one that allows us to vindicate the idea that modal utterances
express truth-conditional content. Thus, my semantics is conservative, insofar as it preserves classical patterns of
inference, and the classical truth-conditions for modal vocabulary.

3
this intuition can be adequately captured by saying that the truth-conditional content of the big
premise describes the conditional probability, while the truth-conditional content of the small
premise describes the unrestricted one. They all argue, in one way or another, that there is no
plausible and systematic way of deriving the right truth-conditions given a context of utterance.7
Note that, though it might seem that the revisionary tendencies are localized to particular
counterexamples to a particular inference pattern—MT, this reaction opens the floodgates: it’s
easy to devise similar counterexamples to numerous other inference patterns that are classically
valid. The question then becomes which of the deductive rules of inference we should reject. To
illustrate, consider the same scenario as in Yalcin’s original counterexample: there’s an urn with
100 marbles, etc. But then:
(4) Suppose that the marble is big.
(5) Then it is likely red.
(6) But the marble is not likely red.
(7) So, the marble is not big.
(4)–(7) is also horrible. If (1)–(3) provides grounds for rejecting MT, then by parity (4)–(7)
provides grounds for rejecting MP or reductio. We might then reject both modus ponens and
modus tollens (as in fact Kolodny and MacFarlane (2010) do). Or we might reject reductio.
Should we reject all of these rules? Of course, we can derive MT through MP and reductio, on
the assumption of the monotonicity of logical consequence. More generally, we know that rules
of inference are holistic in this sense. But that is part of the problem—how do we isolate the
culprit(s)? How do we choose?
None of this is, strictly speaking, an argument against option (i), since, the proponents of
the revisionary views already accept that our modal vocabulary is at odds with classical logic.8
But it does show that there is a lot more at stake than the loss of MT. Moreover, as we shall
see shortly, contrary to what the proponents of the revisionary views assume, the problem is not
tied specifically to modal vocabulary, but can be replicated with the most basic case of context-
sensitive expressions—demonstrative pronouns. This suggests that the real problem may lie
elsewhere. A theory of context-sensitivity resolution that preserves the aforementioned inference
rules, while explaining away the appearance of counterexamples has a lot going for it. Since I
am precisely interested in developing such a theory, I shall reject option (i).
One way to pursue option (ii) that has received some attention in the literature is to claim
that epistemic modals—in particular, the modal operator ‘likely’—takes obligatory wide-scope
over the conditional, so that (1)–(3) is not a genuine instance of MT.9
If this were the only way to implement option (ii), we would face a hard choice indeed, for
obligatory wide-scoping faces well-known problems. For one, it generates intuitively incorrect
predictions. For example, on the (supposedly obligatory) wide-scope reading of the modal ‘likely’,
we get the intuitively wrong interpretation for (8).10
7 Again, assuming the standard notion of truth-conditions.
8A number of authors have suggested that the failure of MT is best explained by adopting the informational
approach to logic and consequence relation. We can then provide a proof system for such logic, and study its
relations to classical logic. See e.g. Yalcin (2007, 2012); Bledin (2014, forthcoming). See also Veltman (1985);
Gillies (2004); Willer (2010) for a related approach to logical consequence. These authors point out that various
classically valid inference patters—MT and reductio for instance—are valid within a restricted fragment that
does not contain modal expressions. However, here, we are precisely interested in whether the presence of modal
expressions gives rise to failures of classically valid patterns, and whether the best semantics for modality violates
classical logic. In other words, the interesting question is whether MT or reductio are unrestrictedly valid. I shall
argue, contrary to what these authors maintain, that the answer to this question is positive.
9 This proposal has been discussed by e.g. Yalcin (2012), Dorr and Hawthorne (2014), and Kratzer (1983),

though all of these authors reject it as unsatisfactory.

10 Yalcin (2012) provides further discussion of this problem.

4
 (8) If Bill comes to the party, then John will come and it is likely that Margaret will
come, too.
(8) does not have a reading according to which it is likely that if Bill comes to the party, John
and Margaret will come, too.
We can, alternatively, defend option (ii) by appeal to the context-sensitivity of modal oper-
ators. It is not particularly controversial that the interpretation of modal operators depends on
context (although, exactly how is a matter of great controversy). A familiar view is that modals
are quantifiers over possible worlds, but just which worlds depends on the context (Kratzer, 1977,
1981). We can exploit this to argue that the problematic counterexample can be explained away
by maintaining that the modal ‘likely’ in (1) contributes a different semantic content than the
one in (2), due to contextual effects on the interpretation of the two occurrences of the modal;
and so, (2) and the consequent of (1) fail to contradict each other. Accordingly, (1)–(3) is not
really an instance of MT.
This strategy captures the intuition that the consequent of (1) talks about a restricted (con-
ditional) probability, while (2) talks about an unrestricted one.11 The challenge is to explain
exactly why and how the context secures different (and intuitively correct) interpretations for
the two occurrences of the modal. To do so in a non-ad hoc way is notoriously difficult.12
To give more bite to the ad hoc-ness charge, note that it becomes even more pressing once we
acknowledge that contextual effects are not freely available with many other uncontroversially
context-sensitive expressions. To illustrate, consider the following example:
(9) If John ate the food from the fridge, then the fridge is empty.
(10) But the fridge is not empty.
(11) So, John didn’t eat the food from the fridge.
One cannot freely shift the context so that ‘empty’ in (10) means in the state of a vacuum, and
thus avoid the MT reading. So, why assume that context can freely shift between (1) and (2),
but not between (9) and (10)?13
This apparent asymmetry raises a worry about the relation between language and logic on
the contextualist accounts. These accounts typically leave the resolution of context-sensitivity
to broadly open-ended pragmatic mechanisms (e.g. speaker’s intentions, or non-linguistic con-
textual cues).14 But, if logical forms are partly fixed by broadly unsystematic, open-ended
mechanisms, then validity becomes partly a matter of such mechanisms, as well. As a result, on
such accounts, there is no systematic, rule-governed way of determining whether (1)–(3), or (9)–
(11), is expressing a valid argument, or not. For sure, given a set of fully-specified logical forms,
one can determine a subset of valid ones; but given a surface form, like (1)–(3), one derives the
fully-specified logical form only through open-ended, defeasible pragmatic processes. The rules
of language alone do not determine whether a pattern like (1)–(3) is valid or not. Thus, logical
11 However, as Yalcin (2012) notes, one should not be too quick to conclude, given this intuition, that the

consequent of (1) and (2) express different contents, since there are non-truth-conditional, non-contextualist
accounts available, Yalcin’s included, that capture this intuition without positing a difference in content.
12 See e.g King (2014), and Stanley (2000).
13 Of course, one could try to modify (9)–(11) to achieve the context-shifting effect by, e.g. putting a focal

stress on ‘empty’. But focal stress is one of the linguistic mechanisms that systematically affect the interpretation
of context-sensitive expressions, in a way that is predicted by the theory I shall defend. (More on this below.)
The point is just that one has to additionally signal when there is a contextual shift via a mechanism such as,
e.g., focal stress. The need for some such mechanism precisely shows that context cannot shift freely. So, the
challenge of spelling out the mechanism(s) that would yield the interpretation of (1)–(3) that the proponent of
option (ii) defends becomes only more pressing.
14 This kind of an approach was initially advocated by Kaplan (1989a), as an account of the resolution of

demonstrative pronouns. The idea is roughly that, with the possible exception of the so-called pure indexicals—
expressions like ‘I’, whose linguistic meaning fully determines the interpretation given a context—context-sensitive
expressions require pragmatic supplementation in order to be interpreted in a context.

5
inference becomes dependent on non-linguistic and psychological factors, like epistemic cues and
communicative intentions. The link between grammar and logic is thus only indirect, mediated
by pragmatic principles. This is worrisome if we cannot provide a systematic, yet non-ad hoc,
account of when and how such mechanisms affect the resolution of context-sensitive expressions.
And the problem is that the pragmatic mechanisms are too flexible to provide such systematic
constraints: recall, if we want to claim that (9)–(11) is an instance of MT, while (1)–(3) is not,
we need a principled story of why context affects the resolution in a particular way in one case,
but not in the other. And we also need a story about why, by contrast with (9)–(11), we do not
find contexts in which (1)–(3) expresses a valid logical form. It’s hard to see what such a story
would be, if it is to rely on speaker’s intentions and general, non-linguistic contextual cues.
By contrast, I offer a systematic formal account of the effects of context on the interpretation
of modals that is not ad hoc. I shall argue in what follows that we have independent evidence
for context-change in (1)–(3). Moreover, I shall argue that we have evidence for the kind of
context-change that is governed by mechanisms that affect the interpretation of context-sensitive
vocabulary quite generally, and is resolved in a systematic, rule-governed way.15 Once we see
the import of these mechanisms, we will see that the puzzling behavior of modals is not in
tension with classical logic. In particular, we will be able to devise a precise theory of context
that predicts the appearance of examples in (1)–(3), while at the same time, provably preserves
classical logic.

1.2 Modals and Pronouns

Consider the following example:
(12) If Jane is out, then she is having fun.
(13) She (pointing at Mary) is not having fun.
(14) So, Jane is not out.
(12)–(14) is obviously bad. But no one would invoke this argument to present a counterexample
to MT. Obviously, (13) does not really contradict the consequent of (12). The occurrences
of ‘she’ in (12) and in (13) resolve to different referents. This explanation works because it
identifies non-ad hoc context-shifting in the inference. We can point to clear reasons why the
reference is resolved in a particular way in the relevant cases. In (12), the pronoun ‘she’ occurs
in an elaboration of a hypothetical scenario about Jane, and thus, is resolved to Jane. In (13),
the pronoun occurs in tandem with a demonstrative gesture—a pointing at Mary—and thus is
resolved to Mary. The elaboration in (12) and the pointing gesture in (13) render a certain referent
the most prominent for the subsequent anaphora; the subsequent pronoun then, as a matter
of its meaning—the character, in the sense of Kaplan (1989b)—picks out the most prominent
referent.16 As I have argued elsewhere (Stojnić, Stone, and Lepore, 2013, 2014), and will amplify
on in the next section, we can capture these observations by rendering explicit the mechanisms
15 As we shall see, moreover, these mechanisms do not merely affect the resolution of the context-sensitive

items, but are also reflected in the logical from of an argument like (1)–(3). See also Stojnić (2016a) for further
discussion.
16 Note that the example in (12)–(14) is not merely trading on the difference between deictic and anaphoric (uses

of) pronouns. Similar examples can easily be constructed where all the pronouns are interpreted anaphorically:
(i) Mary is upset because Jane is much luckier than she is.
a. If Jane buys a ticket, she always wins.
b. But, she does not always win.
c. #So, Jane didn’t buy a ticket.
No one would treat (i) as a serious threat to MT. We can point to reasons why a certain interpretation is naturally
retrieved in the context: while ‘she’ in (i-a) is uttered in the course of an elaboration of a hypothetical scenario
about Jane, and is thus resolved to Jane, the contrastive focus on the occurrence of ‘she’ in (i-b) requires that it
refers to the contextually most prominent referent other than Jane—and this is Mary.

6
that systematically govern the resolution of the pronoun. We can pinpoint the mechanisms that,
for our purposes, play the key role in governing the resolution of context-sensitivity by building
on the resources of so called Discourse Coherence Theories. As we shall see in what follows,
these mechanisms are reflected in the logical form of the discourse like (12)—(14), which will be
particularly important insofar as we want to treat validity as a matter of logical form.17

1.3 Coherence
The key insight behind Coherence Theory is the simple but often neglected observation that a
discourse is more than a random sequence of sentences. To flesh this out, we begin with an
illustrative example from Hobbs (1979):
(15) John took the train from Paris to Istanbul. He has family there.
(16) John took the train from Paris to Istanbul. He likes spinach.
There is a stark contrast between (15) and (16). While (15) is a perfectly felicitous piece of
discourse, (16) (out of the blue) is odd. What explains this contrast? Note that (15) does not
merely report two random, unrelated facts about John. It signals that John took the train from
Paris to Istanbul because he has family there; we understand the second sentence as providing an
explanation of the events described in the first. Recognizing this explanatory connection between
the two bits of discourse in (15) is part of understanding the contribution in (15)—unless we
recognize it, we have simply failed to fully understand the discourse. Unless we understand
how the two bits of discourse are related, we cannot fully understand the speaker’s contribution.
By contrast, due to difficulties in establishing such a connection, (16) seems off. We are left
wondering: is Istanbul famous for its spinach? Does spinach cause a fear of flying? That such
an interpretive effort is in play in an attempt to understand (16) suggests the requirement of a
readily recoverable implicit organization of the discourse that renders it coherent.
Drawing on these observations, Coherence Theorists postulate an implicit organization of dis-
course that establishes inferential connections—coherence relations—among utterances (Hobbs,
1979; Kehler, 2002; Asher and Lascarides, 2003). This implicit organization arises from the com-
municative strategies that interlocutors exploit to convey and organize their ideas through an
ongoing discourse. As demonstrated by the contrast between (15) and (16), successive contri-
butions to a discourse must be linked by a recognizable network of interpretive relationships.
The speaker must signal how she structures her contributions according to shared standards and
conventions.18 So, for instance, (15) is understood as connected by the coherence relation of
Explanation.19 Failure to establish such a connection in (16) makes it seem off.20
Crucial for us is that the task of establishing discourse coherence and resolving semantic
ambiguities turn out to be mutually correlated processes. In particular, as has been confirmed
17 See Hobbs (1979), Kehler (2002), and Asher and Lascarides (2003). I do not mean to suggest that mechanisms

of discourse coherence are the only mechanisms affecting the resolution of either pronouns or modals. Other
linguistic mechanisms, such as prosody, for example, likewise can affect the prominence in a discourse. See
Stojnić, Stone, and Lepore (2013, 2014) for an account that incorporates the effects of various different linguistic
mechanisms on the resolution of demonstrative pronouns.
18 This is not to say that a discourse cannot be ambiguous with respect to coherence relations it harbors. We

see one such example of ambiguity in (17) below. Part of the interpretive effort in understanding a discourse is
in resolving such ambiguities.
19 Coherence Theorists typically capture these observations by representing inferential relations—coherence

relations—in the logical form of a discourse. Cf. Asher and Lascarides (2003). There is both good syntactic and
good semantic evidence for representing coherence relations in the logical form of a discourse. See, in particular,
Asher and Lascarides (2003) and Webber et al. (2003) for further discussion.
20 Of course, we could make (16) felicitous, were we to provide a sufficiently rich context, which would allow us

to establish the relevant relation—for example, if it were a part of the common ground that the best spinach is
grown in Istanbul. This is just as expected.

7
by a number of empirical studies, pronoun resolution co-varies with the choice of coherence
relation.21 Here is an illustration:
(17) Phil tickled Stanley, and Liz poked him. (Smyth, 1994)
There are (at least) two ways we can understand (17). The second clause could be taken to
describe the result of the event described by the first: Phil tickled Stanley, and so, Liz poked
him (i.e. “Liz is avenging Stanley”); or, one could understand the two clauses as comparing
and contrasting two parallel events: Phil tickled Stanley, and Liz poked him as well (i.e. “What
happened to poor Stanley?”). Crucially, if the discourse is understood as connected by the Result
relation, the pronoun refers to Phil; if it is organized around the Parallel relation, then ‘he’ is
Stanley. The choice of a coherence relation guides the choice of pronoun resolution.
Moreover, the data suggest that the mutual constraints between these two tasks are both
systematic and robust. That is, given the choice of a coherence relation, the interlocutors are
radically constrained in the possible interpretation of a pronoun. To illustrate, consider the
following example from Kehler (2002):
(18) Margaret Thatcher admires Ronald Reagan, and George W. Bush absolutely wor-
ships her.
Kehler reports that (18) is generally judged infelicitous by his subjects. The subjects expect the
pronoun in the second sentence to resolve to Reagan, and intuitively feel the speaker has erred
in uttering ‘her’ instead of ‘him’. This is explained, again, by the interaction between the task
of establishing coherence and that of resolving a pronoun. The discourse follows ‘admires’ by
‘absolutely worships’—a stronger term in a scalar relationship—thus signaling that the discourse
is organized by the coherence relation of Parallel. Parallel requires that the occurrence of a
pronoun in the object position be resolved to Reagan (the object of the first clause). Given
the gender mismatch, the utterance is judged infelicitous. This is surprising given the available
referent for ‘her’ in the first conjunct, one that is a well-known subject of Bush’s admiration. If the
correlation were really a matter of mere general pragmatic defeasible reasoning, the perceived
infelicity of (18) would be mysterious. The infelicity is, by contrast, expected if the effect of
Parallel is a matter of an underlying convention—the convention determines that the referent
has to be Reagan, and that is why we are stuck with infelicity, even in the presence of a nearby
plausible interpretation.22
I take these (and other23 ) observations to show that coherence relations render certain refer-
ents prominent for subsequent anaphora resolution.24 Here’s one way to capture the constraints
21 See e.g. Wolf, Gibson, and Desmet (2004), Kehler et al. (2008), and Kaiser (2009).
22 Though Coherence Theorists agree that, due to the semantic effects of coherence relations on the truth-
conditions of a discourse, as well as due to syntactic constraints on discourse structure, coherence relations need
to be represented in the logical form of a discourse (cf. Asher and Lascarides (2003); Webber et al. (2003)), they
typically take the process of pronoun resolution and establishing coherence to be merely mutually constraining.
They understand the preference for a particular coherence relation to be an input to general holistic process of
reasoning that attempts to find the overall most plausible interpretation of the discourse. By contrast, I have
argued elsewhere that we should understand these dependencies not merely as mutual constraints between two
related tasks, but rather as linguistic effects of coherence relations on pronoun resolution; coherence relations
as a matter of linguistic convention make particular entities preferred candidates for subsequent anaphora. (For
detailed defense of this view, see Stojnić, Stone, and Lepore (2013, 2014).) I advance considerations below that
the effects of discourse coherence on modals is likewise conventionalized. For a more detailed defense of this
position see Stojnić (2016b, chs. 1 and 2, and 5).
23 It’s worth noting, moreover, that languages differ with respect to the effects of coherence relations on the

interpretation of pronouns, which suggests that, indeed, the effect is a matter of linguistic convention. See Stojnić,
Stone, and Lepore (2014).
24 The notion of prominence within a discourse that I’m relying on here should not be confused with the intuitive

notion of real-world salience. That a referent is salient is not sufficient (or necessary) to make it prominent in my
sense. Consider an utterance of “Jim came in. He sat down,” while Bill is jumping up and down, making himself
the most salient entity in the given situation. Unless the speaker points or somehow demonstrates that Bill is the

8
that the choice of coherence relation places on pronoun resolution. Suppose we rank candidate
referents for anaphora in a discourse according to their relative prominence.25 The import of dis-
course coherence is to affect this ranking, by making certain referents prominent for subsequent
anaphora. The pronoun, then, according to its meaning, refers to the most prominent referent
that satisfies the grammatical constraints encoded by the linguistic meaning of the pronoun (e.g.
‘he’ refers, roughly, to the top-ranked, third-person, singular, male candidate referent).26 All of
this supports our observations about (12)–(14). The consequent in (12) stands in an Elaboration
relation with the antecedent; it is because ‘she’ in (12) occurs in an elaboration of the hypothet-
ical scenario described by the antecedent that the pronoun refers to Jane. And it is because in
(13) the pronoun occurs in a tandem with a pointing gesture, that it refers to the individual
pointed at, namely, Mary. As a result, (12)–(14) is not an instance of MT.27
So far, I have demonstrated that even in the case of pronouns, if we fail to appreciate how
they are resolved within a discourse, we would be misled to interpret examples such as (12)–
(14) as “counterexamples”. Moreover, I have argued that pronoun resolution is responsive to
discourse structuring mechanisms, in particular, mechanisms of discourse coherence. Next, I
shall argue that modals are analogous to pronouns in two crucial respects. First, like pronouns,
their interpretation is an anaphoric process, by which I mean that they require a contextually
available antecedent which is either linguistically introduced or available from a non-linguistic
context.28 And second, like the resolution of pronouns, the interpretation of modals is guided
by the mechanisms that structure the information in discourse, in particular, mechanisms of
discourse coherence. These mechanisms govern the interpretation of both types of expressions
in a systematic, and rule-governed way. Once we incorporate the effects of these mechanisms on
the resolution of context-sensitivity, we will see that we can devise a semantics for modals, the
underlying logic of which remains classical.

1.4 Modals as Pronouns

That modals exhibit anaphoric-like behavior has been observed in the literature on modal sub-
ordination, as in (19), where a modal is interpreted relative to some other modal expression
introduced earlier in a discourse (i.e. relative to a linguistically introduced antecedent):29
(19) A wolf might walk in. It would eat you first. (Roberts, 1989)
Perhaps the clearest argument for the full analogy between modals and pronouns with respect to
the range of interpretive effects they permit has been offered by Stone (1997, 1999).30 First, he
referent, the occurrence of the pronoun ‘he’ will be interpreted as Jim. This is not to say that prominence has
nothing to do with the entity being salient in the pre-theoretic sense; rather, it is just to say that interlocutors
tend to rely on a relatively narrow set of intra-discourse cues that language provides for guiding their attention
towards particular referents in a discourse. See Stojnić, Stone, and Lepore (2014), and Stojnić (2016b).
25 The idea of a ranking of referential candidates in a discourse comes from Centering Theory, which hypothesizes

that the referential candidates in a discourse are ranked according to relative prominence, those ranked higher
being preferred over those ranked lower as interpretations for subsequent anaphora. See Sidner (1983), Grosz,
Joshi, and Weinstein (1995), and Bittner (2014).
26 For a detailed development of this view, see Stojnić, Stone, and Lepore (2014).
27 Note that according to this explanation, there are changes in context—making certain referents prominent

as antecedents for subsequent pronouns at particular points in discourse—which explain why the pronouns in
(12)–(14) are resolved the way they are. But these contextual changes are induced by linguistic mechanisms,
reflected in the logical from of a discourse; in particular, they are governed by mechanisms of discourse coherence.
28 Some theorists like to reserve the terms “anaphora” and “anaphoric” for cases where an item is bound by a

linguistically introduced antecedent. The reader should bear in mind that I use these terms throughout in this
broader sense specified in the text.
29 See Roberts (1989). How exactly to account for modal subordination has been a matter of much debate, one

that has generated a vast literature. The account defended in this paper shows how to model modal subordination.
30 Stone’s arguments for the parallel between modals and pronouns are analogous to Partee’s famous arguments

for the parallel between pronouns and tense. See Partee (1973, 1984).

9
observes that both can depend for their interpretation on an antecedent introduced linguistically
(either by means of indefinite or definite reference) earlier in the discourse, as illustrated by (19).
Just as ‘he’ in (20) is naturally understood to refer to the man introduced in the first sentence,31
so ‘would’ in (19) is naturally understood restrictedly, as describing the hypothetical scenario
introduced by the modal ‘might’ in the first clause.
(20) John owns a donkey. He beats it. (Based on Partee (1984).)
Second, like pronouns, modals allow for reference to specific entities from a non-linguistic context.
In particular, just as (21), can be uttered out of the blue to refer to some significant woman
available in the discourse context, so too (22), uttered out of the blue, can be understood to
describe the hypothetical scenario that is salient in the discourse context (the scenario in which
the speaker buys the stereo):
(21) (Referring to a certain significant female) She left me. (Partee, 1973)
(22) (Looking at a high-end stereo in an electronics store) My neighbors would kill me.
(Stone, 1997)
Third, both types of expression allow for bound readings, where intuitively their semantic inter-
pretation co-varies with the domain of some higher binding operator, as in (23) and (24):
(23) Every woman believes that she is happy.32 (Partee, 1984)
(24) If a concert goer arrives late, he or she will not be permitted into the auditorium.33
(Stone, 1997)
Finally, both types of expression allow for so called “donkey anaphora” readings, as witnessed by
(25) and (26); crucially, just like ‘it’ in (25) co-varies with the indefinite NP ‘a donkey’ (without
being within its syntactic scope), so in (26) the modal in the consequent ‘will’ co-varies with the
sub-constituent of the antecedent clause ‘if the enemy captures it’.
(25) If a man owns a donkey, he beats it. (Partee, 1984)
(26) If a submarine cannot self-destruct if an enemy captures it, the enemy will learn its
secrets. (Stone, 1997)
These data strongly suggest an analogy between modals and pronouns with respect to the kind of
interpretive dependencies they allow; both search for an antecedent either previously linguistically
introduced, as in (19)–(20), and (23)–(26) or provided by the context, as in (21)–(22).34
Observe that conditionals, too, display this type of anaphoric behavior. For example, in
(27), the second conditional depends on a scenario introduced by the first one; it is not simply
evaluated against all epistemically accessible worlds, available in the context discourse initially.
And the modal in the consequent of the second conditional is thus interpreted only relative to
the hypothetical scenario introduced by the antecedent of the second conditional relative to the
hypothetical scenario described by the first conditional.35
(27) If a wolf walks in, we will use the tranquilizer gun. If we manage to shoot it, we
will be safe.
31 I assume that there are no accompanying pointing gestures in (20).
32 Similar examples are found in e.g. Partee (1973) and May (1977).
33 I cite the original example that Stone provides, but the effect is easily replicated without ‘will’: “If a concert

goer arrives late, he or she might not be permitted into the auditorium.” The same goes for Stone’s other examples
involving ‘will’ (in particular, example (26)).
34 Further support for the analogy between modals and pronouns with respect to their interpretive range is

provided by examples from languages such as Warlpiri or American Sign Language (ASL) that allow for a single
anaphoric expression to be ambiguous between pronominal and modal interpretation. See Bittner (2001) for data
on Warlpiri, and Schlenker (2013) for data on ASL.
35 Just as, as we shall see, the anaphoricity of modals plays a key role in explaining the counterexamples like the

one in (1)–(3), so this anaphoricity of conditionals plays a key role in explaining the counterexamples to MP and
MT involving right-nested conditionals, as McGee’s (1985) original counterexample to MP, or Veltman’s (1985)
counterexample to MT.

10
One way we can think about the observed anaphoricity (in the sense of “anaphoricity” described
above) of modals and conditionals is as follows. It is customary to treat modals as quantifier
expressions, quantifying over possible worlds.36 We know that quantifiers require a contextually
supplied domain restriction—“Everyone had fun today” does not mean that everyone in the
universe had fun today.37 The same goes for modals—like other quantifiers, they also require
that their domain of quantification be further contextually restricted.38 “A wolf might walk
in” does not convey the meaning that in at least one world out of all possible worlds a wolf
walks in. Rather it conveys a more restricted meaning—at least one world out of the relevant,
epistemically accessible worlds is such that in it a wolf walks in. Since in the case of the modals
the domain of quantification is just a set of worlds, the restrictor on the domain will likewise be
a set of worlds (a possibility, for short). I suggest that modal anaphora resolution is a matter
of retrieving the possibility that serves as the restrictor on the domain of quantification of a
modal operator. That is, modals (and conditionals) are anaphoric insofar as they require an
anaphorically retrieved restrictor: just as with the antecedent of a pronoun, the restrictor can
either be explicitly linguistically introduced in the discourse (e.g. by some modal utterance prior
in the discourse), as in (19), (24), (26) and (27), or otherwise made prominent in the context, as
in (22).39
How this anaphoric dependency is resolved—i.e. how the restrictor is retrieved in a context—
brings us to our second analogy between pronouns and modals. Crucially, modals exhibit the
same kind of responsiveness to mechanisms of discourse coherence as pronouns. The idea is that
mechanisms of discourse coherence affect the prominence of possibilities that are candidates for
the restrictor of a subsequent modal, just as they can affect the prominence of referents that are
candidates for subsequent pronominal anaphora; they render certain possibilities prominent to
serve as a restrictor of a subsequent modal. We see this already in (19), where the second sentence
elaborates on the hypothetical possibility described by the first. It is because this Elaboration
relation holds that the modal ‘would’ in the second sentence is understood as restricted by
the possibility described by ‘might’ in the first— all of the hypothetical scenarios out of those
epistemically accessible ones in which the wolf walks in are such that the addressee is eaten
first. (Similarly, it is because there is an Elaboration relation between the antecedent and the
consequent in (24) that the modal ‘will not’ in the consequent is understood as restricted by the
possibility described by the antecedent.) The Elaboration relation is what makes this scenario
prominent for the subsequent modal to pick up on.
The import of coherence is crucial in making a certain possibility prominent as a restrictor
of a subsequent modal. Note that mere sequencing of modals is not sufficient for the correct
interpretation. That is, it is not always the case that the modal will be restricted by a possibility
introduced by the immediately preceding modal (when there is an immediately preceding modal).
The fact that one modal follows another does not suffice to establish that the hypothetical
scenario described by the second modal further elaborates the one introduced by the first. We
easily see this with examples like (28).40
36 Cf. Kratzer (1977, 1981, 2012).
37 For more on quantifier domain restriction, see von Fintel (1994), and Stanley and Szabó (2000).
38 Typically, in an ordinary discourse, the restrictor will eliminate at least some worlds from the domain of

quantification, on pain of redundancy. In principle, however, the restrictor need not eliminate anything. This can
happen, for instance, if the restrictor is a necessary proposition.
39 Does this mean that quantifiers are also anaphoric in my sense? The answer is yes. Like modals they require

a restrictor, provided either by the non-linguistic context, or the prior discourse. Moreover, though the details
exceed the scope of the present paper, there are good reasons to hold that the way in which the restrictor of a
quantifier is retrieved in a context is analogous to the way in which the restrictor of a modal and an antecedent of
a pronoun is, i.e. that quantifier domain restriction is sensitive to discourse structuring mechanisms. (See Stojnić
(2016b).)
40 For similar examples, see Asher and McCready (2007).

11
 (28) If a wolf walks in, it would eat you. But one probably won’t walk in.
As before, the consequent of the first sentence in (28) elaborates on the information provided
by the antecedent. The Elaboration relation between the antecedent and the consequent makes
the hypothetical scenario introduced by the antecedent the most prominent one, and as a result,
‘would’ in the consequent further describes this scenario. Crucially, however, the modal ‘won’t’
in the second sentence does not further elaborate upon scenario described by the two modals in
the first sentence. The two sentences stand in a relation of Contrast, signaled by the discourse
marker ‘but’, and are understood as contrasting two hypothetical scenarios—one in which a wolf
walks in, and one in which one does not.
Intuitively, the Contrast relation requires that the first and the second sentence provide
contrasting information about some body of information. A bit more precisely, the two bits
of discourse provide contrasting information about some body of information regarding some
common topic—in our example the topic is what is possible regarding a wolf’s entrance.41 The
first sentence already sets the stage in determining the body of information the contrast has to be
about—(assuming that the conditional is uttered discourse initially) it is interpreted relative to
the set of epistemically accessible worlds determined by the context discourse initially, describing
what might be the case if a wolf walked in, given this overall body of knowledge. The second
sentence, then, has to provide a contrasting bit of information, regarding the possibility of
a wolf’s entrance, about this body of information available discourse initially. The Contrast
relation makes this body of information prominent, and consequently, the modal in the second
sentence selects it as its restrictor—given this overall body of knowledge, a wolf probably won’t
come in. This is the intuitively correct interpretation: given the overall body of knowledge, if a
wolf walks in, it would eat the addressee, but given the same body of knowledge, one probably
won’t walk in.
What this sort of example establishes is that, just as with pronouns, the impact of discourse
coherence on making a certain possibility prominent as a restrictor for a subsequent modal is
crucial. As the contrast between Elaboration and Contrast shows, precise discourse mechanisms
govern the prominence of possibilities in a discourse. We can capture this idea in a way similar to
the way in which we captured the effects of coherence on the prominence of candidate referents
for pronoun resolution. Here is a first pass proposal: let the context represent a ranking of sets of
worlds (possibilities, for short) that are candidates for domain restrictors of modals in a discourse
according to their relative prominence—the most prominent being the top-ranked one. A modal
simply retrieves the most prominent epistemically live possibility as the restrictor for its domain
of quantification.42 The prominence ranking of candidate possibilities, in turn, is affected by a
range of linguistic mechanisms, most notably, those of discourse coherence; coherence relations
make certain possibilities prominent for subsequent modal anaphora.43
41 The relevant topic is typically signaled by the cues in the information structure—the way the information is

packaged together. One way of signaling this, in English, is by exploiting prosodic accents. For example, compare
the following two utterances:
(i) John likes MARY.
(ii) JOHN likes Mary.
(i) is fine in the context in which we are wondering whom John likes, say, Mary or Sue, but not in the context
in which we are wondering who likes Mary, say, Bill or John; the opposite is true of (ii). For more on sentential
focus, see Rooth (1992), and for more on information structure, see e.g. Roberts (1996).
42 I will use the term “epistemically live possibility” to denote the possibility that is not ruled out given the

relevant body of knowledge, i.e., which is accessible given the relevant epistemic accessibility relation. For a more
precise notion of epistemic accessibility, see § 2.1 and § A.2.
43 Though limited space prevents me from going into details here, I have argued elsewhere that there are

good reasons to treat these effects of coherence on the resolution of modal anaphora as linguistically encoded,
rather then as byproduct of general reasoning. (See Stojnić (2016a,b).) For instance, as Asher and McCready
point out, the direct translation of (19) in Japanese is infelicitous, unless there is an overt discourse marker

12
 I shall assume that at the beginning of a conversation the top-ranked possibility is just a
set of epistemically possible worlds—a set of worlds epistemically accessible from the actual
world. The intuitive idea behind this assumption is the familiar idea that the ultimate goal of
a conversation is to narrow down possible ways the actual world could be.44 However, as the
discourse progresses, the prominence ranking changes; this change in ranking is precisely what
we want to model, since this is precisely the change relevant for the interpretation of modals. The
changes arise as an effect of introducing novel possibilities (e.g. through utterances containing
modals or conditionals), and through discourse structuring mechanisms that change and affect
the prominence ranking, in the ways described earlier.
Connecting this with our previous observations, the anaphoricity of modals is captured by
requiring that the restriction on the domain of quantification be retrieved in a way similar to
how the antecedent of a pronoun is—either provided by the context, or explicitly by the prior
discourse. The way anaphora is resolved, in both cases, is determined by discourse structuring
mechanisms, in particular, mechanisms of discourse coherence. Just as the mechanisms of dis-
course coherence affect the prominence of candidate referents for the resolution of a pronoun, so
too they affect the prominence of a candidate possibility for a restrictor on a subsequent modal.
To sum up: modals are like pronouns in two crucial respects: (a) they are anaphoric ex-
pressions, and (b) the resolution of modal anaphora is responsive to the same mechanisms that
pronoun resolution is responsive to. We now have all the ingredients we need to tackle the
original counterexample.45
signaling an Elaboration relation (Asher and McCready, 2006).) This would be surprising if the elaboration
reading, where the second modal is understood as elaborating on the scenario described by the first one, were
merely a byproduct of general reasoning. Moreover, as I argue in Stojnić (2016b,a) the mechanisms of discourse
coherence sometimes force inconsistent interpretations of modal discourse, even when there are possible plausible
alternative interpretations that could have been retrieved instead. That the inconsistent readings are retrieved
even in the face of alternative common-sense interpretations is easily explained if the mechanisms of discourse
coherence conventionally mandate such interpretations; however, this would be mysterious if the effects of these
mechanisms were byproducts of pragmatic reasoning, since in such cases, we have overwhelming common-sense
reasons to disprefer the inconsistent interpretations, and so override the effects of discourse coherence. These
considerations, among others, suggest that the effects of the mechanisms of discourse coherence are a part of our
linguistic repertoire, and should hence be reflected in the logical form of the discourse.
44 Cf. Stalnaker (1978). More precisely, since typically there is uncertainty about which world is the actual one

(given that our knowledge is limited), the initial set of epistemically live possibilities will be a set of worlds W
that contains, for each world w that is a candidate for the actual world, the set of worlds accessible from w. Of
course, if we are concerned with the interpretation of discourse initial uses of modals, it will matter a great deal
whose body of knowledge is relevant for determining the set of epistemically accessible worlds. Since I shall not
deal with this issue in the present paper, I shall simply assume that the relevant body of epistemically accessible
worlds will be contextually provided discourse initially. However, pace Kolodny and MacFarlane (2010), Yalcin
(2007), and Moss (2015), the resources of my account can show how this body of information is selected in a
context, even discourse initially. (See Stojnić (2016b).) Relatedly, as I argue at length in Stojnić (2016b), the
resources of my account naturally explain the patterns of intra- and inter-contextual (dis)agreement involving
modal language, that have been argued to favor the revisionary theories precisely on grounds of the alleged failure
of context to determine an adequate body of information in such cases. (For the revisionary arguments based on
(dis)agreement patterns, see e.g. Egan, Hawthorne, and Weatherson (2005); MacFarlane (2014). For alternative
contextualist responses to these arguments see, e.g., Dowell (2011) and von Fintel and Gillies (2009). These
contextualist accounts (much like Kratzer’s account) assume general pragmatic mechanisms of context-sensitivity
resolution that are too unconstrained to systematically account for the apparent counterexamples to MT, as well
as the puzzling behavior of modals in various embedding environments, including, but not limited to antecedents
of conditionals. See Stojnić (2016b) for a detailed discussion.)
45 Note that, according to the theory developed here, modals express truth-conditional content, and the ex-

planation of the counterexample exploits the difference in the truth-conditional content between the consequent
of the big premise, and the small premise in (1)–(3). Several authors have challenged the view on which modal
vocabulary expresses truth-conditional content on grounds that are not directly related to the failure of classical
inferences. (See e.g. Egan, Hawthorne, and Weatherson (2005), Yalcin (2007), and Moss (2015).) Though the
theory developed here, given its systematic account of context-change, has means for accounting for those chal-
lenges as well, addressing them here is beyond the scope of the present paper. (I address these issues in detail in

13
1.5 Yalcin’s Counterexample Explained Away
We can now explain what is going on with (1)–(3) as follows: in (1), the consequent of the
conditional is understood as elaborating on the possibility introduced by its antecedent (which
is just the set of epistemically accessible worlds in which the antecedent holds), and thus, the
Elaboration relation renders this possibility the most prominent one for subsequent anaphora.
Consequently, the modal ‘likely’ in the consequent, which is searching for the most prominent
epistemically accessible possibility, selects this possibility as the restrictor for its domain of
quantification. The consequent is thus understood as further describing the possibility introduced
in the antecedent, providing the intuitively correct restricted reading—the marble is likely red,
given that it is big. In turn, we naturally understand (2) as being linked to (1) by the relation
of Contrast. This is seen even more clearly if we insert an explicit discourse marker ‘but’ in (2):
‘But the marble is not likely red.’ Note that some such way of signaling contrast is required, for
the discourse consisting of (1) followed by (2) to be felicitous.
As always, the question of which relation holds (and between which relata) is a matter of
disambiguation in a discourse, much as in the case of (17). There are often discourse-internal,
linguistic cues that signal a particular relation (e.g. a discourse marker ‘but’), but context can
play a role in disambiguation as well. For example, here the initial context sets up a topic—
the color of the marble, depending on a certain assumption about its size. (1) and (2) are
then understood as providing contrasting bits of information about some body of information
regarding this topic. As before, the first sentence already sets the stage in determining the
body of information the contrast has to be about—since (1) is uttered discourse initially, it is
interpreted relative to the overall body of knowledge available discourse initially. Thus, one
understands (1) and (2) as providing two contrasting bits of information about the overall body
of knowledge or information available discourse initially, regarding the likelihood of the marble
being red, depending on a certain assumption about its size. Namely, given the overall available
information discourse initially, if the marble is big, then it’s likely red, but additionally, given the
overall available information, the marble is not likely red (given no particular assumption about
its size).
A bit more precisely, the effect of Contrast in (1) and (2) is the same as in (28). (1) and
(2) are understood as contrasting two different bits of information about some initial overall
Stojnić (2016a) and Stojnić (2016b).) I shall just note that we can, for example, easily explain the problematic
behavior of epistemic modals embedded in antecedents of conditionals—one of the key data points used by Yalcin
to argue against the truth-conditional accounts (Yalcin, 2007, 2011). Yalcin notes the discrepancy between the
following two conditionals:
(i) If it is raining and it might not be raining, ...
(ii) If it is raining and I/we don’t know it, ...
The first one is odd, while the second perfectly fine, pointing to an apparent problem for the standard truth-
conditional accounts of conditionals, that interpret modals as quantifying over some salient body of knowledge.
Yet, the natural way to understand “it might not be raining” in the conditional above is as elaborating on the
hypothetical raining scenario. This affects the resolution of modal anaphora—‘might’ is understood as quantifying
over all the relevant epistemically accessible worlds in which it is raining, and so it is no surprise that the
conditional winds up being bad. Note that my account does not predict that the conditional with the reversed
order of conjuncts in the antecedent, i.e. of the form: “If it might p, and not p, then...,” will automatically be
bad. This is a desired result since, as Dorr and Hawthorne (2014) note, switching the order of the conjuncts in
some cases (in particular, in Yalcin’s original example) makes the conditional felicitous. Yet, my account does not
predict that all such cases will automatically be felicitous either. This is because the badness of such a conditional
will depend on which coherence relations, and other interpretive dependences, can be established between the two
conjuncts in the antecedent, and between the conditional and the rest of the discourse in which it is embedded, for
these factors can all affect the resolution of modal anaphora in a particular case. This is, again, a desired result.
And the point holds more generally—embedding any sentence (including the original Yalcin’s conditional) in a
broader discourse might give raise to various interpretive dependences that might yield a different interpretation
than the one we get when interpreting the same sentence-type out of the blue.

14
body of information, regarding the likelihood of the redness of the marble, given some or no
assumption about its size. The first sentence already sets the stage in determining the body of
information that the contrast has to be about—the conditional is interpreted relative to the set
of epistemically accessible worlds determined by the context discourse initially,46 describing what
the likelihood of the redness of the marble is, given this body of knowledge, provided that it’s big.
The second sentence thus has to provide a contrasting bit of information about this same body
of information (i.e. the set of epistemically accessible worlds, discourse initially) regarding the
likelihood of the marble being red. Thus, this body of information is made prominent by Contrast.
Consequently, the modal in (2), selects this body of information as its restrictor, conveying that
the marble is not likely red given this overall body of knowledge (given no particular assumption
about its size). Thus, we see that the two occurrences of ‘likely’ in (1) and (2) are interpreted
differently, much like the two occurrences of ‘she’ are in (12) and (13). And thus, (1)–(3) is no
more an instance of MT than (12)–(14) is.47

1.6 Conclusion to Part I

In § 1, I argued that the lesson of the alleged counterexamples to classically valid patterns of
inference is not that we need a revisionary, non-truth-conditional semantics for modal vocabulary
that invalidates these patterns, but rather that we need a different, more constrained picture of
how context affects interpretation. In particular, I argued that modals are like pronouns not
merely in that they are anaphoric expressions, but in that they are sensitive to some of the same
interpretive mechanisms of anaphora resolution as pronouns are—in particular, the mechanisms
of discourse structure. The “counterexamples” arise due to a failure to appreciate the import
of these mechanisms on the interpretation of modals. But, a failure to appreciate how context
changes with the evolution of a discourse would lead us to mistakenly treat even the examples
like (12)–(14) as counterexamples. That the effect of context-change on modals is more subtle
and less well explored only makes the failure to appreciate it that much more dangerous.48
The mechanisms of context-sensitivity that I have argued are relevant for the interpretation of
modals, though pervasive and nuanced, are not unruly. We can devise a formal theory of context-
change that models the import of these mechanisms on the interpretation of modals in a rule
governed and systematic fashion, and indeed, this is what I do in § 2, and in the Appendix (§ A).
Thus, the formalism developed there should be understood as an existence proof in support of the
main argument of § 1. The theory accommodates the anaphoricity of modals, while preserving
their standard truth-conditions, and maintaining the validity of the classical patterns of inference.
Though the resulting semantics departs from the standard semantics for modal discourse, it does
so in a way that allows us to capture the interplay between discourse structuring mechanisms
and the interpretation of modal anaphora. It is precisely this departure that preserves classical
logic, and permits modal utterances to express standard truth-conditional content.
It is important to note that the account developed here opens a promising line for further
development. In particular, philosophers often identify context-sensitivity of various philosophi-
46 Provided that the conditional is uttered out of the blue, which, by assumption, it is.
47 Note that, while my account maintains that there is a change in context that transpires in (1)–(3), affecting
the interpretation of modal expressions, the account does not maintain that such a change is somehow ‘illicit’. In
an important sense, there is no way of holding the context fixed throughout the course of (1)–(3). This is because
(1)–(3) harbors linguistic elements (in particular, modals, antecedents of conditionals, and coherence relations)
part of the meaning of which is to change the context in a way that affects the truth-conditions (by introducing
certain possibilities, making them prominent, and demoting others).
48 As I argue in more detail in Stojnić (2016b), the revisionary accounts in fact are not well suited to explain

the full range of data concerning modal anaphora. For instance, they do not capture the modal anaphora data
described in § 1.4. My account, in turn is precisely designed to account for these data. Thus, apart from preserving
truth-conditions and classical inference patterns, the account is well motivated on independent grounds.

15
cally interesting terms—e.g., of implicit restriction on quantifier domains, knowledge ascriptions,
vague predicates, normative terms—and use this context-sensitivity to motivate broad philosoph-
ical conclusions. But in doing so, they typically assume a model of context-sensitivity that is
resolved by freely selecting one candidate resolution out of an open-ended list of potential ones,
through general pragmatic mechanisms (e.g. speaker’s intentions, and non-linguistic contextual
cues). This predicts a level of flexibility that often fails to be born out in practice, and this flexi-
bility in turn shapes the philosophical arguments that appeal to such context-sensitivity. Though
arguing for this in full generality obviously exceeds the scope of the present paper, going beyond
modals and pronouns, the kinds of tools developed here open a path for exploring the potential
systematic constraints on other context-sensitive expressions. If the approach advocated here
can be extended to capture the contextual-sensitivity of these other kinds of expressions, then
that would show that contexts are much less powerful, and the resolution of context-sensitivity
a much more constrained process than what philosophers typically assume in their arguments.49

2 Part II: A New Semantics for Modality

In what follows, we have two tasks: first, to modify the standard truth-conditional account,
developed in Kratzer (1977, 1981, 1983, 2012), in order to accommodate the anaphoricity of
modals and conditionals, and second, to develop an account of prominence and context-change
that explains how modal anaphora is resolved in a discourse. I will undertake the first task in
§ 2.1, and the second task in § 2.2. As I shall demonstrate, the formal semantics that emerges
assigns standard truth-conditions to modals and conditionals, and as I prove, in the Appendix
(§ A), preserves classical logic.

2.1 Anaphoricity and Truth-Conditions

As mentioned earlier, the classic account treats modals as quantifying over a contextually supplied
modal base, i.e. a set of contextually relevant worlds that comprise the domain of quantification.
A modal then requires that a particular relation holds between the proposition expressed by
the clause embedded under a modal (which is typically called the prejacent 50 ) and the modal
base.51 For example “It might rain” requires that there is some possible world in the contextually
supplied modal base in which it rains. Let ‘q’ denote the proposition expressed by the prejacent
φ of ‘might φ’ in a given context and ‘R’ an accessibility relation, that specifies the epistemically
accessible worlds, i.e. the modal base.52 Then the truth-conditions expressed by ‘might φ’,
relative to a context of utterance, are:
49 Indeed, in Stojnić (2016b) I argue that the kind of account developed here extends to numerous other cases

of context-sensitivity, and that the kind of mechanisms I described here are suited to explain context-sensitivity
quite generally.
50 So, in ‘might φ’, φ is the prejacent.
51 This captures the main insights from Kratzer’s account, though I suppress formal details for simplicity.

In particular, I suppress an ordering source parameter, which provides an ordering of worlds in a modal base
according to some contextually provided standard, and the formal machinery that serves to derive a modal base
parameter. (Cf. Kratzer (1977, 1981, 1983, 2012).) We could easily factor these elements back in. As Stone
points out, modal base and ordering source parameters, as specified in Kratzer’s account, cannot accommodate
the anaphoricity of modals, since both are determined in complex ways, and neither provides a semantic parameter
that can be contributed by prior discourse, so a modification of Kratzer’s account is needed regardless (Stone,
1997).
52 The relation R plays the role of a Kratzerian modal base. On Kratzer’s account the modal base is contextually

determined in complex ways. For our purposes, we can simplify even further, and let R be supplied by the model,
since we are not dealing with discourse-initial uses of modals. See Stojnić (2016b) for details on how R is
determined in context, discourse initially.

16
Definition 2.1.
{w | ∃w0 : wRw0 & w0 ∈ q}
That is, ‘might φ’ is true at a world w (relative to a context c) just in case there is some world
among the worlds epistemically accessible from w in which the proposition, q, expressed by the
prejacent in c, holds. Anaphoric dependency is easily factored into the standard truth-conditions
of modals explicitly as follows. (I shall use ‘M (p, q)’ for the truth-condition expressed by an
utterance of ‘might φ’, where q is the proposition expressed by the prejacent φ of the utterance
of ‘might φ’, and p the proposition corresponding to an anaphorically retrieved restrictor. I omit
the details about how context determines truth-conditions (and, in particular, the restrictor p)
here. This will be the topic of § 2.2.)
Definition 2.2.
M(p, q) = {w | ∃w0 : wRw0 & w0 ∈ p & w0 ∈ q}
This gives us the resources to define other modal expressions: as is standard, ‘must’ is the
universal dual of ‘might’.53
For probability modals, such as ‘likely’, we need a probability measure over the accessible
worlds. Let P be a probability measure over the set of worlds in the universe W , that maps each
subset of W to [0, 1], satisfying the following constraints:
• P(W ) = 1
• P(p ∪ q) = P(p) + P(q), when p and q are disjoint subsets of W .54
Then, where q is the proposition expressed by the prejacent φ of an utterance of ‘likely φ’, and
p the proposition corresponding to the anaphorically retrieved restrictor, the truth-conditions,
P (p, q), expressed by the utterance of ‘likely φ’ are as follows:
Definition 2.3.
P(p, q):= {w | P({w0 | wRw0 & w0 ∈ p & w0 ∈ q}) / P({w0 | wRw0 & w0 ∈ p}) > .5}55
As expected, the truth-conditions expressed by an utterance of ‘likely φ’ are the set of worlds
such that, for each w in the set, the ratio of the probability that an R-accessible world from w be
a p and q world to the probability that the R-accessible world from w be a p-world is greater than
.5; i.e. an utterance of ‘likely φ’ is true in w just in case given our modal base, the conditional
probability of the prejacent, q, given the p-restricted modal base is greater than .5.56
53 For the definition of truth-conditions of ‘must φ’, see § A.2.
54 I shall assume that the probability measure is supplied by the model. Alternatively, it could be provided
by context. The choice is inessential for the context-sensitivity I’m aiming to model. For simplicity, I assume
that W is finite. I also assume that P is a regular probability measure, i.e. assigning non-zero probability to all
non-empty sets of worlds. Insofar as P(p) represents the prior, if P(p) = 0, p wouldn’t really be a possibility.
55 It is typically assumed that the restrictor on the domain of quantification, p, is not empty. But if we wanted

to allow for the possibility in which p is empty, we could modify our definition in a following way: P(p, q):=
{w | P({w0 | wRw0 & w0 ∈ p & w0 ∈ q}) >1/2 P({w0 | wRw0 & w0 ∈ p})}. Also, one might wonder whether we
should always impose a threshold of .5, or perhaps the threshold might vary with the context. For simplicity I
choose the former option. The choice is inessential for our purposes.
56 I depart from Kratzer (1991) in suggesting this quantitative characterization of the truth-conditional contri-

bution of probability operators, rather than a qualitative one cashed out in terms of relative likelihood. There
are well known problems with a purely qualitative account. For a detailed discussion, see Yalcin (2010); for a
discussion of the prospects of basing the quantitative notion of probability on a qualitative one, see Kratzer (2012,
ch 2.) and Holliday and Icard (2013). This particular choice I make here, though independently motivated, is
inessential to the overall point of the paper. What matters for us is that the anaphoric potential of a probabil-
ity modal is correctly captured. We can build the anaphoric dependency into the truth-conditions in the way
suggested, regardless of what we think the correct account of the truth-conditions is. The point holds for the
truth-conditions of other modals as well.

17
 Finally, we need to specify the truth-conditional contribution of a conditional. We can easily
factor in the anaphoric potential into the truth-conditions of a conditional, just as we did with
modals, while otherwise preserving the standard truth-conditions. Where p, as before, is the
anaphorically retrieved restrictor with respect to which the conditional is uttered, q corresponds
to the proposition expressed by the antecedent, and r to the one expressed by the consequent,
an utterance of a conditional expresses truth-conditions corresponding to a set of worlds such
that for each w in the set, all the worlds w0 , R-accessible from w, that are p and q worlds, are r
worlds as well; i.e. an utterance of a conditional is true in w if and only if all the p and q worlds
in the modal base are r worlds as well:57
Definition 2.4.
Cond(p, q, r):= {w | ∀w0 : wRw0 , if w0 ∈ p & w0 ∈ q, then w0 ∈ r}
This in essence preserves the standard truth-conditions associated with a conditional, but factors
in the fact that a conditional itself is always evaluated against some prominent body of infor-
mation, that need not always correspond to the complete unrestricted set of epistemically live
worlds.
This concludes my characterization of the truth-conditional contribution of modals and con-
ditionals. According to the proposed account, the utterances containing modals and conditionals
express the same truth-conditions one would expect given the canonical account,58 except that
the anaphorically retrieved restrictor is factored in separately. This allows us to flexibly track
the way it is retrieved in a context. It is important to note that, provided that all anaphoric
restrictors are resolved to the same set of worlds (e.g. to the set of all possible worlds), we
get exactly the truth-conditions for modals we would expect in standard propositional modal
logic (our conditional is a standard strict-conditional).59 However, we have yet to specify the
effects of context and context-change on the determination of truth-conditions of a given utter-
ance. As I have argued, the most important impact of context for us will be in the resolution
of anaphoric dependencies of modals and conditionals, i.e. in the role of context in determining
the restrictor of the modal base of modals (and conditionals). In § 2.2, I lay out a formal theory
of context-change, which will allow us to tackle the alleged counterexample we began with.

2.2 Truth-conditions and the Dynamics of Context-change

In order to capture the idea that the restrictor of a modal expression is the most prominent
possibility, we need to have a way of modeling the prominence of possibilities in a discourse
context.60 As suggested in § 1.4, we can let a context represent a ranking of possibilities—sets
of worlds, or propositions—according to relative prominence, the top-ranked possibility being
the most prominent one. One way to think about such a context is as a sort of a conversational
57 I ignore the ordering source, but we could easily factor this parameter in, and then state the truth-conditions

by making a conditional true (in a world w and at a context c) just in case all the p and q worlds that are closest
to w in the modal base are r worlds as well. Apart from factoring in the anaphora (and modulo the abstraction of
the ordering source parameter), the truth-conditions I propose here match the ones developed in Kratzer (1983).
Kratzer’s account makes a conditional true in a world and at a context just in case all the (closest) antecedent
worlds within a modal base are consequent words as well.
58 Cf. Kratzer (1977), Kratzer (1981), Kratzer (1983) and Kratzer (2012)
59 More precisely, on the assumption that R is reflexive and transitive, we would get the system S4. For a proof,

see section § A.7. This is a common and natural assumption. Reflexivity, e.g., ensures that must p entails p, and
also that p entails might p, and transitivity ensures, e.g., that might(might p) entails might p, again, provided
that all anaphoric restrictors are resolved to the same set of worlds (e.g. to the set of all possible worlds).
60 I sketch the key bits of the formal account here, but for the fully precise formal definitions and details, consult

§ A.

18
record, in the sense of Lewis (1979), an abstract “scoreboard” that tracks the moves and con-
tributions interlocutors make in the flow of a discourse, and that comprises information relevant
for interpretation, such as who’s speaking, what the conversation is about, etc. Crucially, for us,
the conversational record tracks propositions put into play in the course of a conversation as well
as their relative prominence. Since this is the only aspect of the scoreboard that will matter for
modal anaphora, we can abstract away from all other aspects that might be otherwise needed. In
this spirit, we can see my context as modeling one aspect of a Lewisean scoreboard—the relative
prominence of possibilities within a discourse.
We will let a context represent the ranking of possibilities according to their relative promi-
nence. To model a context, we will exploit the idea of an assignment function, modeled as a
stack, in the sense of theoretical computer science; an assignment function, or a stack, specifies
an ordered sequence of worlds.61 Let the context be a set of such stacks.62 So, a context is just a
set of sequences of worlds. For a given context G, let ‘wi ’ be a variable that stores a world at the
ith position of every stack in G.63 Then, relative to G, ‘wi ’ stores a set of worlds that collects
the worlds at the ith position of every stack in G. This, in particular, will allow us to keep track
of the propositions that have been introduced and promoted (as well as demoted), during the
course of a discourse. I shall assume that every stack in G begins with the 0th position, as the
top-ranked position on the stack, and that for each position n, the position n + 1 is one position
lower in the ranking. Let ‘Gi ’ denote the set of worlds that collects the worlds at the ith position
of every stack in G.64 Then, G0 is the top-ranked proposition in G, and for each n, Gn+1 is a
proposition one step lower in the ranking. This allows us to keep track of the relative prominence
of possibilities for modal anaphora.
As noted in § 1.4, I shall assume that at the beginning of a conversation, the top-ranked
possibility is just a set of epistemically accessible worlds. In turn, as the discourse progresses
utterances affect the prominence ranking of candidate possibilities. As we have already seen in
an informal way in § 1, utterances can promote novel possibilities and re-rank the ones introduced
prior in the discourse. The simplest way to model this is to represent utterances as updates to
the context that change the relative prominence of candidate possibilities. Formally, an update
is represented as a relation between an input context and an output context where the update
is true (relative to an input context and a world of evaluation w) just in case it relates an input
context to some non-empty output context (relative to w).65
Thus, utterances have a two-fold contribution—they express truth-conditions (as per § 2.1),
but they also contribute updates that change the context, by updating the ranking of possibilities.
Both aspects of the interpretation are crucial. The updates associated with utterances capture the
way in which these utterances change the context; in turn, such a dynamically maintained context
61 Formally, a stack is an assignment function, mapping a finite convex subset of N to a set of worlds together

with an undefined value ⊥.

62 This formalization, unlike the one exploiting a single stack ranking sets of worlds, allows us to keep track

of not only sets of worlds, but also relations between individual members of different sets. (Cf. Bittner (2014),
Brasoveanu (2010), and van den Berg (1996).) Though this is not strictly speaking crucial for our present purposes,
it does become crucial once we want to develop an integrated semantics for anaphora both in the modal, and in
the pronominal domain.
63 In describing context change I am adopting the following strategy. I define a dynamic language that models

the dynamics of prominence, and then provide a translation of a fragment of English into this language. The
dynamic language has atomic expressions (propositional constants (p, q, r), and variables (wi for i ∈ N), conditions
(propositional expressions comprising set of atoms closed under ∧ and ¬), and update expressions, which we will
define and describe shortly. For details see § A (in particular, § A.1, § A.4 and § A.7).
64 Thus, G is the proposition stored at the ith position in G.
i
65 This is the standard notion of truth exploited in dynamic semantics. (See Dekker (2011).) Recall that on

my account utterances also express ordinary truth-conditional content. The dynamic notion of truth defined here
is exploited to capture the logic of context-change. We will see that once we have the logic of context-change in
place, the ordinary truth-conditional content (and the ordinary notion of truth) falls out of it straightforwardly.

19
determines the relevant truth-conditional content of a given utterance. Thus, we effectively model
the two-way interaction between a context and an utterance: on the one hand, an utterance
changes the context, and on the other, such a dynamically evolving context determines its truth-
conditional content.
Before laying down the updates associated with modals and conditionals, we need some
preliminaries. We said that utterances contribute updates to a context by promoting and re-
ranking propositions, i.e. possibilities, but they also express propositions. I am going to separate
the role that propositions play in anaphora, from the one they play in semantic composition.
To illustrate why we need this separation, take the example: “A wolf might eat Harvey”. First,
we need to compose the proposition expressed by the prejacent, that a wolf eats Harvey, with
the modal ‘might’. But if this proposition automatically became the top-ranked possibility in
the context, since the modal ‘might’ selects the top-ranked epistemically live proposition as its
restrictor, so long as there’s some epistemically accessible world in which a wolf eats Harvey, we
would predict that the restrictor for the modal in “A wolf might eat Harvey” is the proposition
comprising all the epistemically live worlds in which a wolf eats Harvey. But, obviously, this
proposition should not automatically be made a restrictor, as witnessed by examples like: “A
wolf might walk in. It might eat Harvey.”
There are several ways to get around this problem, but one elegant way is to store sepa-
rately the propositions that enter into semantic composition, and the ones that are candidate
restrictors for subsequent modals. That insures that the truth-conditional contribution does not
automatically interfere with prominence ranking.66
To this end, I reserve a designated position on each stack in a context G that does not affect
the ranking. Let ‘comp’ (for compositional) denote the proposition that comprises all the worlds
stored at this position in every stack in G.67 We exploit ‘comp’ to store, relative to G, each bit
of propositional content that enters into semantic composition.68 Here is how we do that. We
will treat expressions that contain no proper propositional subparts as atomic formulae in our
system. When φ is an atomic formula (and so, does not contain modals, or conditionals), its
interpretation will just be the simplest update, defined below in 2.5, which stores the proposition
expressed by that formula in the input context G as a new value of ‘comp’. This update relates
the input context G to an output context H, (relative to a world of evaluation w) just in case
H differs from G in at most the value of ‘comp’, and the value of ‘comp’ is the proposition that
‘p’ expresses in G and w. That is, where G is the input context, H the output context, w the
world of evaluation, and G ∼ n
H just in case G differs from H in (at most) the nth position, we
can define this update as follows:
Definition 2.5.
∼ H & Hcomp = JφKG,w .
Jhcomp := φiK(w, G, H) iff G comp
When φ is non-atomic, it will be interpreted as a more complex update. However, crucially,
its truth-conditional contribution determined by the update will still be stored as a value of
‘comp’ of the output context, in a way to be specified presently.69
Observations in § 1.3 and § 1.4 show that modals require an anaphorically retrieved restrictor,
but also introduce possibilities that can subsequently be picked up by other modals. I suggested
66 This is not to say that one and the same proposition cannot play a role both in semantic composition and also

be prominent for anaphora resolution. This is just to say that the mechanisms by which a proposition completes
these two roles are best kept separate.
67 So, now, formally, a stack is just a function from a finite convex subset of N plus comp to a set of worlds

together with an undefined value, ⊥.

68 Similarly as before, ‘G
comp ’ denotes the value of ‘comp’ in G.
69 In particular, 2.5–2.12 are clauses of a single recursive definition of update expressions in our dynamic

language. See the Appendix (§ A) for details.

20
that the anaphorically retrieved restrictor will be the top-ranked epistemically live possibility.
Let ‘@E’ denote the top-ranked epistemically live possibility (the top-ranked possibility that is
a subset of epistemically accessible worlds) in a given context.70 Intuitively, we want the update
associated with an utterance of ‘might φ’ to have the following effect: first, (as with all other
updates) its truth-conditions (as defined in 2.2) are stored as the value of ‘comp’ of the output
context. Second, it introduces a possibility comprising the top-ranked epistemically live worlds
in which the prejacent holds. Here is how we can informally describe the effect on context that
an update associated with ‘might φ’ carries out. Where G is an input context, and H an output
context, the update relates an input context G to an output context H just in case there are
intermediate contexts G0 and G00 and:
a. G0 is a result of updating G with the update associated with prejacent of the modal (recall
that since updates associated with both atomic and non-atomic formulae will store their
truth-conditional content as the value of ‘comp’, G0comp will just be the truth-conditional
contribution of the prejacent in G, i.e. JφKG,w ),
b. G00 is just like G0 except that it stores the top-ranked epistemically live possibility in which
the prejacent holds (which is just (G0comp ∩ J@EKG,w )) as a novel top-ranked possibility,
and pushes all other values one position down, and
c. the final output context H is just like G00 except that it stores the truth-conditional content
expressed by the utterance as a new value of ‘comp’, which by 2.2, is the set that comprises
all the worlds w such that there’s some world epistemically accessible from w in which both
the restrictor (the top-ranked epistemically live possibility in the input context G) and the
prejacent hold (i.e. M (J@EKG,w , G0comp )).
Putting all this together, we can now define the update associated with ‘might φ’. Let us
define a relation between contexts, ≈
n
, where for any two contexts G and H, this relation holds
just in case H is obtained from G by storing a novel value for n and pushing all other values one
position down in the ranking; more precisely, G ≈n
H just in case H is identical to G up until n,
it differs from G in (at most) the nth position, and for all m, such that m ≥ n, Gm = Hm+1 .71
Then, where K is an update associated with the prejacent φ, we define the update associated
with an utterance of ‘might φ’, that carries out the steps in (a.)–(c.), in the following way:
Definition 2.6.
Jmight(@E, K)K(w, G, H) iff there is a G0 and G00 such that JKK(w, G, G0 ) & G0 ≈
0
G00 & G000 =
G0comp ∩ J@EKG,w & G00 comp
∼ H & Hcomp = M (J@EKG,w , G0comp ).
Let us see how this works through an example.
(29) A wolf might walk in.
Given 2.5 and 2.6, we can represent (29) as follows. Let ‘p’ stand for the prejacent (‘a wolf walks
in’); since atomic, by 2.5, it’s just interpreted as hcomp := pi. Thus, we get:
• Jmight(@E, hcomp := pi)K(w, G, H) which by 2.6 holds just in case there is a G0 and G00
such that: Jhcomp := piK(w, G, G0 ) & G0 ≈
0
G00 & G000 = G0comp ∩ J@EKG,w & G00 comp
∼ H &
Hcomp = M (J@EKG,w , G0comp ).
70 Formally, we include in the basic vocabulary of our dynamic language unary predicates and a unary operator

‘@’. Where ‘E’ is a unary predicate and ‘@ a unary operator, ‘@E’ is an atom. ‘@E’ is interpreted as taking a
property denoted by ‘E’ and delivering the top-ranked proposition satisfying it, denoted by ‘@E’.
0 ,n + gn+1... |g ∈ H} = G and Gcomp = Hcomp . See section § A.3 of the
71 Even more precisely: G ≈ H iff {g
n
Appendix for details.

21
 • By 2.5, Jhcomp := piK(w, G, G0 ) holds just in case G comp
∼ G0 & G0comp = JpKG,w . That
is, it holds just in case G0comp is the proposition expressed by p in G and w, namely, the
proposition that a wolf walks in.
• Moreover, the possibility corresponding to a set of top-ranked epistemically live worlds in
which a wolf walks in is introduced as a novel top-ranked possibility, pushing all other
possibilities further down in the ranking (thus we get G00 ).
• Finally, the proposition expressed by the modal utterance is stored as the value of ‘comp’ in
the final output context H, which is otherwise just like G00 ; by 2.2, this proposition stored
as the value of ‘comp’ in H corresponds to a set of worlds R-related to some world in
which both the anaphorically retrieved restrictor and the prejacent of the modal hold (i.e.
M (J@EKG,w , G0comp )). The anaphorically retrieved restrictor is the top-ranked epistemi-
cally live possibility in the input context G, (i.e. J@EKG,w ) which, assuming that (29) is
uttered out of the blue, just is the set of epistemically accessible worlds discourse initially.
Thus, the proposition expressed by (29), the value of comp in the output context H, is
the proposition that some of the epistemically accessible worlds discourse initially are such
that in them a wolf might walk in.

Putting all this together, we get that (29) (a) expresses the proposition that it is compatible
with what is known discourse initially that a wolf walks in, and (b) introduces a possibility
comprising the top-ranked epistemically live worlds in which a wolf walks in, making it available
for subsequent modal anaphora. This is just the desired result.
It is now also easy to see what the updates associated with utterances containing ‘must’
and ‘likely’ should look like. They will exactly parallel the update associated with utterances
containing ‘might’, the only difference being in the truth-conditions. So, the update associated
with ‘likely’ proceeds in exactly the same steps (a.)–(c.). The only difference will transpire in
step (c.), where now the truth-conditional content stored as a value of ‘comp’ of the final output
context, (H), will be the truth-conditional content expressed by an utterance containing ‘likely’,
which, by 2.3, is P (J@EKG,w , G0comp ), i.e. a proposition that requires that given our modal base,
the conditional probability of the prejacent, G0comp , given the appropriately restricted modal
base, J@EKG,w , is greater than .5. So, the update associated with ‘likely’ can be defined as
follows:72
Definition 2.7.
Jlikely(@E, K)K(w, G, H) iff there is a G0 and G00 such that JKK(w, G, G0 ) & G0 ≈
0
G00 &
G000 = G0comp ∩ J@EKG,w & G00 comp
∼ H & Hcomp = P (J@EKG,w , G0comp )

Finally, we need to specify the update associated with a conditional. The update will proceed
in a similar fashion as before, with one difference: now we have to first process the update
associated with the antecedent, and then with the consequent. That is, where G is an input
context, the update first stores the proposition expressed by the antecedent (in G, relative to the
world of evaluation w), as the value of ‘comp’ and introduces the top-ranked epistemically live
worlds in which the antecedent holds as the top-ranked possibility (pushing all other possibilities
one position down). Relative to thus obtained intermediate context (G00 ), it stores the proposition
expressed by the consequent (in G00 , relative to w), as the value of ‘comp’ and introduces the top-
ranked epistemically live worlds in which the consequent holds pushing all other possibilities one
position down in the ranking, resulting in the intermediate context (G0000 ). Lastly, the final output
72 For the definition of an update associated with ‘must’, see § A.4.

22
context (H) differs from the intermediate context G0000 only insofar as it stores the propositional
contribution of the conditional as the value of ‘comp’: as per 2.4, it stores the proposition true at
a world w just in case all the epistemically accessible worlds from w in which the anaphorically
retrieved restrictor and the antecedent hold are such that the consequent holds in them as well.73
More precisely, the update relates an input context G and an output H just in case there are
some contexts G0 , G00 , G000 and G0000 , and:

a0 . G0 is a result of updating G with the update associated with the antecedent (thus storing
the truth-condition expressed by the antecedent in G, as the value of ‘comp’ in G0 ),
b0 . G00 is just like G0 except that it stores the top-ranked epistemically live possibility in which
the antecedent holds, as a novel top-ranked possibility (G000 ), and pushes all other values
one position down,
c0 . G000 is the result of updating G00 with the update associated with the consequent (thus
storing the truth-condition expressed by the consequent in G00 , as the value of ‘comp’ in
G000 ),
d0 . G0000 is just like G000 except that it stores the top-ranked epistemically live possibility in
which the consequent holds, as a novel top-ranked possibility (G0000
0 ), and pushes all other
values one position down, and finally,

e0 . the final output context H is just like G0000 except that it stores the truth-conditional content
expressed by the whole utterance of the conditional as a new value of ‘comp’, which by
2.4, just is Cond(J@EKG,w , G0comp , G000
comp ), i.e. a proposition that requires that in all the
epistemically accessible worlds in which the antecedent (G0comp ) and the restrictor for the
conditional (J@EKG,w ) hold, the consequent (G000 comp ) holds as well.

Putting all this together, where K1 and K2 represent updates associated with the antecedent
and the consequent, respectively, we define the update associated with the conditional as follows:
Definition 2.8.
Jif(@E, K1 , K2 )K(w, G, H) iff there is a G0 , G00 , G000 and G0000 such that
JK1 K(w, G, G0 ) & G0 ≈0
G00 & G000 = G0comp ∩ J@EKG,w & JK2 K(w, G00 , G000 ) & G000 ≈
0
G0000 &
00
G0000 000
0 = Gcomp ∩ J@EK
G ,w
& G0000 comp
∼ H & Hcomp = Cond(J@EKG,w , G0comp , G000
comp )

This completes the specifications of updates associated with modals and conditionals. As
I have shown, we capture both aspects of their behavior—namely, we characterize the truth-
conditions expressed by an utterance containing a modal or a conditional, and the way in which
such an utterance changes the context in which it is uttered. The updates associated with
utterances change the context, by updating the prominence ranking of possibilities that are
candidate restrictors for subsequent modals and conditionals; the choice of a restrictor in turn
affects the truth-conditions of an utterance containing a modal or a conditional. More generally,
the updates associated with utterances affect the way in which the context evolves as the discourse
progresses; the context in turn determines the truth-conditions expressed by the utterances.
Both aspects of interpretation are crucial, and they are interrelated—unless we captured the
change in the context prompted by an update associated with an utterance containing a modal
or a conditional, we would not be able to tell how the modal can make a possibility available
for subsequent anaphora; and unless we calculated in the anaphoric dependency of utterances
73 As before, I abstract away from the contribution of the ordering source.

23
containing modals or conditionals, we would not be able to correctly predict which proposition a
given utterance containing a modal or a conditional expresses, since the anaphorically retrieved
restrictor crucially factors into its truth-conditions.
As we have seen, the updates in 2.5–2.8 all store their corresponding utterances’ truth-
conditional content as the value of ‘comp’ of the output context. But, we also want to characterize
what it takes to assert this content. Minimally, an assertion of an utterance requires that the
proposition expressed holds at the world of evaluation. Plausibly, it also makes the possibility
associated with the asserted content prominent. We can capture this by ensuring that an assertion
promotes the set of top-ranked epistemically live worlds in which the asserted content holds as
a novel top-ranked possibility,74 and requires that the actual world be within that set. We can
introduce a simple assertion update that achieves this effect:
Definition 2.9.
JAssert(K)K(w, G, H) iff there is a G0 such that JKK(w, G, G0 ) & G0 ≈
0
H & H0 = G0comp ∩
J@EKG,w & w ∈ H0 .
We now have almost all the ingredients we need to capture MT. We need to introduce one last
thing—negation. The truth-conditional contribution of ‘not φ’ is simple—it is true (at a context
and relative to a world w) just in case w is a non-φ world. We can let the update associated
with an utterance of ‘not φ’ simply store the complement of the truth-condition expressed by ‘φ’
in the input context (relative to w), as the value of ‘comp’ of the output context. We define this
update as follows:
Definition 2.10.
0
Jnot(K)K(w, G, H) iff there is a G0 such that JKK(w, G, G0 ) & G0 ∼ H & Hcomp = J¬compKG ,w ,
comp
where J¬compKG,w = Dω \JcompKG,w , where Dω is a set of possible worlds provided by the model.

Finally, we can now tackle the task of specifying MT pattern of inference. Prima facie, we
can represent the pattern of the form [‘if φ, ψ’, ‘not ψ’∴ ‘not φ’], as follows. Let Td (φ) and Td (ψ)
stand for whatever updates correspond to φ and ψ. Prima facie, then, the following seems to be
the pattern we are after:
(30) Assert(if(@E, Td (φ), Td (ψ)); Assert(not(Td (ψ))); Assert(not(Td (φ)))75
However, the following is not yet the full logical form of (1)–(3) (indeed, it’s not a fully specified
logical form at all). First, we cannot know whether (30) has the form of MT or not, until we
know what updates Td (φ) and Td (ψ) are. Until this has been specified, the form is incomplete;
not every instance of the schema (30) is an instance of MT. As we have seen, the updates
that constitute a discourse affect which truth-conditions are expressed by it. Whether or not
a discourse that is an instance of (30) will be an instance of MT (partly) depends on whether
the truth-conditional content expressed by ψ in the context obtained after updating with the
antecedent, and the truth-conditional content expressed by ‘not ψ’ in the context obtained after
updating with the big premise, indeed do negate each other.76 Only if this is the case we’ll have
an instance of MT. And whether or not this condition obtains will depend on the respective
74 Note that this will basically be a new set of candidate worlds for the actual world.
75 I use the standard notation, representing the sequencing of updates with a semicolon. Thus, where K1 and
K2 are updates, K1 ; K2 is also an update, that performs the update K1 followed by K2 (Muskens, 1996).
76 I say ‘partly’ because, while validity is primarily a matter of logical form, it is also a semantic notion,

capturing certain semantic patterns. As we shall see in what follows, we will be able to fully capture validity as
a matter of logical form. Moreover, we will be able to prove that all classically valid patterns are associated with
a valid logical form.

24
contexts that determine the truth-conditions expressed by ‘ψ’ and ‘not ψ’, which in turn will
depend on the way the updates that result in these contexts proceed. This is partly a matter
of what updates Td (φ) and Td (ψ) are, but it is also a matter of which discourse structuring
mechanisms organize the premises and the conclusion into a coherent discourse, since we have
seen that those mechanisms also update the context in a way that affects the truth-conditions
of modal discourse.
We can then state the following generalization: whenever the truth-conditional content ex-
pressed by the small premise negates the one expressed by the consequent of the conditional in
the big one, the truth of the big and the small premise together will entail the falsity of the
truth-conditional content expressed by the antecedent of the conditional in the big premise. (For
a proof of the generalization see A.6.1.) Provided that the updates in (30) satisfy the constraint
that the truth-conditional content expressed by the utterances they represent indeed conform
to the pattern of MT, that is, that the proposition expressed by the small premise negates
the proposition expressed by the consequent of the big one, this pattern is valid; whenever the
premises describe a possible update, the conclusion does so as well.77 Only those fully specified
instances of (30) that preserve the adequate form to meaning mapping corresponding to MT will
be genuine instances of MT; and all those are valid. This is exactly the same point, as the point
that we cannot decide what the full logical form of (12)–(14) is until we know how the different
occurrences of the pronoun are resolved. This is hardly a threat to MT.
Note that by characterizing MT in this way, we characterize it as a pattern depending partly
on the truth-conditional content expressed, not merely on the underlying syntactic form, because
only those ways of specifying (30) that ensure that the truth-conditions of the premises and the
conclusion conform to the pattern of MT will count as MT. One might instead maintain that MT
should be characterized exclusively in terms of a unique syntactic form. Yalcin offers this as an
additional argument against MT (Yalcin, 2012). Namely, the standard Kratzerian “restrictor”
analysis of conditionals does not recognize the English conditional as a binary operator, and
crucially according to this analysis, what is in the scope of the negation in (2) would not even
be a constituent of (1) (Kratzer, 1983); thus, Yalcin argues, since there is no single dyadic
operator corresponding to the English language conditional (but rather just a multiplicity of
different dyadic operators that correspond to different modals), MT, which he takes to be a
generalization about this alleged dyadic operator, makes no sense. Since there’s no “stable
notion” of “antecedent” and “consequent”, there is no MT.78
I find this line of argument unpersuasive; that a certain syntactic/semantic analysis eliminates
a unique binary operator corresponding to the conditional should be independent of the fact that
a certain semantic pattern is valid. Even if something like the Kratzerian analysis of a conditional
turns out to be true, that will hardly constitute a demonstration that MT and MP actually do
not exist. Perhaps, we should understand the intuitively valid patterns like MT and MP as
precisely the patterns that reflect the behavior of modals in modally subordinated environments,
but that does not mean that such patterns are merely an illusion. (Perhaps, it is useful to reflect
on the fact that though we can formalize propositional calculus by means of, say, Sheffer stroke,
it would be odd to argue on that basis that in such a system MT is somehow ill conceived.)
What, then, follows for the alleged counterexample we began with? Prima facie, the argu-
ment’s structure is as follows. Where ‘p’ stands for “the marble is big”, ‘q’ for “the marble is
red”, and ‘@E’ as before denotes the top-ranked epistemically live possibility in the context, a
77 Here, I’m appealing to the standard dynamic notion of validity: an inference pattern is dynamically valid just
in case the sequential updates with the premises followed by the update with the conclusion lead to a non-empty
context. (See § A.5 for a precise definition.) What we can prove is that dynamic system embeds classical logic:
whenever the truth-conditions associated with the premises classically entail the ones expressed by the conclusion,
the inference pattern is dynamically valid. See § A.
78 See (Yalcin, 2012).

25
first pass representation of (1)–(3) is as follows:
(31) Assert(if(@E, hcomp := pi, likely(@E, q)); Assert(not(likely(@E, q)));
Assert(not(p)))
However, (31) is still not a full-blown logical form of (1)–(3), since it leaves out some of the
relevant mechanisms that affect the truth-conditions.79 An instance of MT requires that the
truth-conditional content expressed by the consequent of the big premise is a negation of the
truth-conditional content expressed by the small premise, and whether this is the case, depends
on the way the relevant updates affect the context which determines the truth-conditions; in
particular, here it depends on the way the anaphoric dependency (i.e., the value of ‘@E’) of the
modal in the big premise and the small one is resolved. We have seen earlier that discourse
coherence plays a crucial role in resolving modal anaphora; thus, in order to determine whether
(1)–(3) is an instance of MT or not, we need to take into account the contribution of these
mechanisms, which are left out of (31). To get a full-blown logical form (1)–(3) we need to
describe the way that mechanisms of discourse coherence update the context, as well.
I argued that mechanisms of discourse coherence change the context by updating the promi-
nence ranking of possibilities that are candidates for anaphora resolution. Furthermore, I argued
that (1)–(3) is not an instance of MT, because the Elaboration relation between the antecedent
and the consequent in the big premise requires that the modal in the consequent further elabo-
rates on the possibility made prominent by the antecedent, while the Contrast relation between
the small premise and the big premise requires that the modal in the small premise quantifies
over the body of information that both premises are about—i.e., the whole set of epistemically
accessible worlds discourse initially. So far, I have specified the updates associated with modals
and conditionals. Now we need to capture the effects of the mechanisms of discourse coherence
on prominence.
We can capture the effect of these mechanisms on the prominence ranking by representing
coherence relations as contributing prominence-affecting updates. Let us first characterize Elab-
oration. I argued that Elaboration promotes the possibility that is elaborated upon. Here’s one
way of capturing this idea: when an utterance elaborates on a possibility φ, a two-fold contribu-
tion is made—first, the possibility elaborated upon, φ, is promoted to prominence, pushing all
other possibilities one position down in the ranking, and second, it is required that the proposi-
tional content expressed by the utterance in question, stands in an elaboration relation to φ. We
can provisionally characterize an elaboration relation between propositions φ and ψ, Elab(φ, ψ),
by requiring that it holds just in case φ and ψ are centered around the same event or entity, i.e.
just in case the event or scenario described by ψ is a part of the event or scenario described by φ.80
Putting all this together, where φ is a possibility, and K an update representing the utterance
elaborating on φ, we can characterize the update associated with Elaboration as follows:
Definition 2.11.
JElab(φ, K)K(w, G, H) iff there are G0 and G00 such that G ≈
0
G0 & G00 = JφKG,w & JKK(w, G0 , G00 )
& G00 ≈
0
H & H0 = G00comp & Elab(JφKG,w , H0 ).

The contrast relation, as we have seen in § 1, has a different effect on the context than
Elaboration. Its main effect is the following: the two bits of discourse provide contrasting
79 As noted before, once we have the full-blown logical forms we can restore the idea of validity as a matter

of logical form: in particular, a sequence of updates expressing classically valid truth-conditional pattern will be
dynamically valid, as well.
80 Cf. Hobbs (1979), Asher and Lascarides (2003). The provisional characterization suffices, because the exact

characterization of Elaboration is not crucial for us; the only thing that matters is the way in which the relation
affects prominence. Ditto for other coherence relations.

26
information about some body of information regarding some common topic. In turn, this body
of information is made prominent. Thus, to characterize Contrast formally, we need to have a
way of specifying the body of information that a given sentence is about—that is, a body of
information it contributes information relative to. Here’s one way of doing this. Where φ is
a formula, we say that φ is about a set of worlds θ just in case, where G is an input context
to φ, θ = J@EKG,w . The idea is just, once more, that as a discourse progresses we are trying
to narrow down the space of epistemic possibilities—thus, a sentence is just about a currently
top-ranked epistemic possibility. I will use ‘θφ ’ to denote the set of worlds φ is about. Then we
can characterize Contrast as follows:
Definition 2.12.
JContrast(K1 , K2 )K(w, G, H) iff there is a G0 and G00 such that JK1 K(w, G, G0 ) & G0 ≈
0
G00 &
00
G000 = JθK1 KG,w & JK2 K(w, G00 , H) & JθK1 KG,w = JθK2 KG ,w
& Contrast(G00comp , Hcomp )
According to 2.12, when two bits of a discourse contrast with each other, a two-fold contri-
bution is made: first, the body of information they are about is made prominent, and second, the
propositions expressed by them are required to stand in Contrast relation (Contrast(G00comp , Hcomp )),
i.e., to provide contrasting information about this body of information, regarding some common
topic.81
Finally, now that we have specified the ways in which prominence ranking changes as the
discourse evolves, putting all this together, we can return to the original counterexample. Where
p stands for “the marble is big” and q for “the marble is red”, we represent (1) and (2) as follows:
(32) Contrast(Assert(if(@E, hcomp := pi, Elab(w0 , Likely(@E, hcomp := qi)))),
Assert(not(Likley(@E, hcomp := qi))))82
The following are the key steps in (32).83 By 2.8, the conditional update introduces the possibil-
ity corresponding to the set of top-ranked epistemically accessible worlds in which the proposition
expressed by the antecedent holds, i.e. the set of epistemically live worlds in which the marble
is big. The consequent provides an elaboration of this possibility; as a result, this possibility
is promoted to prominence (as per 2.11). Furthermore, it is required that the possibility in-
troduced by the consequent stands in the Elaboration relation to the possibility introduced by
the antecedent, which at this point is the possibility ranked at the position 0 (and, so, denoted
by ‘w0 ’ in (32)). Since the consequent contains an occurrence of the modal ‘likely’, by 2.7, the
proposition expressed by the consequent of the given utterance of the conditional corresponds
to the proposition that the marble is likely red, given the top-ranked possibility, which due to
the effect of Elaboration at this point is the set of epistemically accessible worlds in which the
marble is big. Thus, to put it simply, the consequent expresses the proposition that, for all that is
known, the marble is likely red, given that it is big. By 2.8 again, the whole conditional expresses
the proposition that for all that is known, if the marble is big, then it is likely red, given that it
is big. By 2.9, the assertion update requires that the conditional holds of the actual world and
promotes the set of epistemically live worlds in which it holds. Due to the effect of Contrast, as
81 I represent Contrast as operating on two updates, and Elaboration as operating on a proposition and an

update. This is in line with a more general distinction between two classes of coherence relations, ones that select
their arguments structurally (based on syntactic and structural constraints), and ones that select their arguments
anaphorically (Webber et al., 2003). I presuppose this distinction here without defending it, due to considerations
of space.
82 Note that given this formalization, Contrast will not contribute any asserted content on its own. This is a

welcome conclusion, but it is inessential. We could in principle make discourse relations a part of asserted content,
by imposing additional constraints on the world of evaluation in the specification of the updates associated with
coherence relations.
83 For a detailed derivation, and a proof that (1)–(3) is not an instance of MT, see § A.8.

27
specified in 2.12, the body of information that the first utterance is about (which, given that the
conditional is uttered discourse-initially, as by assumption it is, is just the set of epistemically
accessible worlds discourse initially) is promoted to prominence. Then the modal in the small
premise will be interpreted with respect to this body of information—given all that is known,
the marble is likely red. By 2.10, negation then expresses the complement of this possibility; i.e.
the small premise expresses the proposition that it is not the case that given all that is known
the marble is likely big. The assertion (by 2.9, again) makes sure that this propositional content
holds of the actual world, and promotes the set of epistemically accessible worlds in which the
content holds. Finally, the propositions expressed by the two premises are required to stand in
the Contrast relation—i.e to provide contrasting information regarding a common topic, about
the body of information they are about, i.e. the set of epistemically accessible worlds discourse
initially.
Crucially, (32) guarantees that the proposition expressed by the small premise, and the one
expressed by the consequent of the big premise do not contradict each other. The information
expressed is the following: given all that is known, if the marble is big, then it’s likely red, but,
given all that is known, the marble is not likely red. This pattern does not fit the pattern of the
premises of MT. So, a fortiori, (1)–(3) is not a counterexample to MT.84

2.3 Conclusion to Part II

The formal semantics developed in § 2 demonstrates that the account argued for in § 1.2–§ 1.4
can be given a precise formal characterization. The semantics provided in § 2 models the two
main aspects of the interpretation of modals and conditionals—the way in which utterances
containing these expressions change the context, and the way in which the context determines
their truth-conditions. We have seen that both aspects of interpretation are crucial, and that
they are interrelated. Unless we captured the change in the context prompted by an update
associated with an utterance containing a modal or a conditional, we would not be able to tell
how the modal can make a possibility available for subsequent anaphora; and unless we calculated
in the anaphoric dependency of utterances containing modals or conditionals, we would not be
able to correctly predict which proposition a given utterance containing a modal or a conditional
expresses, since the anaphorically retrieved restrictor crucially factors into its truth-conditions.
Only once both aspects of the interpretation are taken into account can we give a full explanation
of the apparent failure of classically valid patterns of inference.
The formal theory developed in § 2 thus provides a precise theory of context-change, and
captures the way the mechanisms of discourse structure affect the interpretation of modals. As
we have seen, these mechanisms are independently motivated, and their impact is systematic and
rule governed; most importantly, the underlying logic is classical. Thus, the semantics achieves
our two main goals: it preserves the validity of classically valid patterns, while at the same time
explaining away the apparent counterexamples.

References
Asher, Nicholas and Lascarides, Alex. 2003. Logics of Conversation. Cambridge University Press.
84 Recall, the account maintains that (1)–(3) harbors linguistic elements (modals, antecedents of conditionals,

and coherence relations) part of the meaning of which is to change the context in a way that affects the truth-
conditions (by introducing certain possibilities, making them prominent, and demoting others). These elements
are reflected in the logical form of (1)–(3): it is because (1)–(3) harbors these elements that it is not associated
with a valid logical from.

28
Asher, Nicholas and McCready, Eric. 2006. “Modal Subordination in Japanese: Dynamics and
Evidentiality.” University of Pennsylvania Working Papers in Linguistics 12:237–249.
—. 2007. “Were, Would, Might and a Compositional Account of Counterfactuals.” Journal of
Semantics 24:93–129.

van den Berg, Martin. 1996. Some Aspects of the Internal Structure of Discourse: The Dynamics
of Nominal Anaphora. Ph.D. thesis, University of Amsterdam.
Bittner, Maria. 2001. “Topical Referents for Individuals and Possibilities.” In Rachel Hast-
ings, Brendan Jackson, and Zsófia Zvolenszky (eds.), Proceedings of SALT XI, 36–55. CLC
Publications, Cornell University, Ithaca.

—. 2014. Temporality: Universals and Variation. Wiley-Blackwell.

Bledin, Justin. 2014. “Logic Informed.” Mind 123:277–316.
—. forthcoming. “Modus Ponens Defended.” Journal of Philosophy .

Brasoveanu, Adrian. 2010. “Decomposing Modal Quantification.” Journal of Semantics 27:437–

527.
Cantwell, John. 2008. “Changing the Modal Context.” Theoria 74:331–351.
Dekker, Paul. 2011. “Dynamic Semantics.” In Klaus von Heusinger, Claudia Maienborn, and
Paul Portner (eds.), Semantics: An International Handbook of Natural Language Meaning,
Volume 1., 923–945. Berlin: Mouton De Gruyter.
Dorr, Cian and Hawthorne, John. 2014. “Embedding Epistemic Modals.” Mind 122:867–913.
Dowell, Janice L. 2011. “Flexible Contextualist Account of Epistmeic Modals.” Philosopers’
Imprint 11:1–25.

Egan, Andy, Hawthorne, John, and Weatherson, Brian. 2005. “Epistemic Modals in Context.”
In Gerhard Preyer and Peter Georg (eds.), Contextualism in Philosophy, 131–169. Oxford
University Press.
von Fintel, Kai. 1994. Restrictions on Quantifier Domains. Ph.D. thesis, University of Mas-
sachusetts, Amherst.

von Fintel, Kai and Gillies, Anthony S. 2009. “‘Might’ Made Right.” In Andy Egan and Brian
Weatherson (eds.), Epistemic Modality. NY: Oxford Univeristy Press.
Gillies, Anthony S. 2004. “Epistemic Conditionals and Conditional Epistemics.” Noûs 38:585–
616.

—. 2010. “Iffiness.” Semantics and Pragmatics 3:1–42.

Grice, H.P. 1989. “Indicative Conditionals.” In Studies in the Way of Words, 58–85. Cambridge,
MA: Harvard University Press.
Grosz, Barbara J., Joshi, Aravind K., and Weinstein, Scott. 1995. “Centering: A framework for
modelling the local coherence of Discourse.” Computational Linguistics 21:203–225.
Hobbs, Jerry R. 1979. “Coherence and Coreference.” Congitive Science 3:67–90.

29
Holliday, Wesley H. and Icard, Thomas F. III. 2013. “Measure Semantics and Qualitative Se-
mantics for Epistemic Modals.” Proceedings of SALT 23:514–534.
Kaiser, Elsi. 2009. “Effects of Anaphoric Dependencies and Semantic Representations on Pronoun
Interpretation.” In Anaphora Processing and Applications, 121–130. Heidelberg: Springer.

Kaplan, David. 1989a. “Afterthoughts.” In Almog Joseph, Perry John, and Wettstein Howard
(eds.), Themes From Kaplan, 565–614. Oxford University Press.
—. 1989b. “Demonstratives.” In Almog Joseph, Perry John, and Wettstein Howard (eds.),
Themes From Kaplan, 481–563. Oxford University Press.
Kehler, Andrew. 2002. Coherence, Reference and the Theory of Grammar. Stanford, CA: CSLI
Publications.
Kehler, Andrew, Kertz, Laura, Rohde, Hannah, and Elman, Jeffrey L. 2008. “Coherence and
Coreference Revisited.” Journal of Semantics 25:1–44.
King, Jeffrey C. 2014. “The Metasemantics of Contextual Sensitivity.” In Alexis Burgess and
Bret Sherman (eds.), New Essays on Metasemantics, 97–118. Oxford Univeristy Press.
Kolodny, Niko and MacFarlane, John. 2010. “Ifs and Oughts.” Journal of Philosophy 107:115–
143.
Kratzer, Angelika. 1977. “What Must and Can Must and Can Mean.” Linguistics and Philosophy
1:337–355.

—. 1981. “Notional Category of Modality.” In Hans-Jürgen Einkmeyer and Hannes Rieser

(eds.), Words, Worlds, and Contexts: New Approaches in Word Semantics, 38 – 74. Berlin:
de Gruyeter.
—. 1983. “Conditionals.” In Anne M. Farley, Peter Farley, and Karl-Erik McCullough (eds.),
Papers From the Parasession on Pragmatics and Grammatical Theory, 1–15. Chicago: Chicago
Linguistic Society.
—. 1991. “Modality.” In Arnim von Stechow and Dieter Wunderlich (eds.), Semantics: An
international handbook of contemporary research, 639–50. Berlin: de Gruyter.
—. 2012. Modals and Conditionals. Oxford: Oxford University Press.

Lewis, David K. 1979. “Scorekeeping in a Language Game.” Journal of Philosophical Logic

8:339–359.
MacFarlane, John. 2014. Assessment Sensitivity: Relative Truth and Its Applications. Oxford
Univeristy Press.

May, Robert. 1977. The Grammar of Quantification. Ph.D. thesis, MIT.

McGee, Van. 1985. “A Counterexample to Modus Ponens.” Journal of Philosophy 92:462–471.
Moss, Sarah. 2015. “On the Semantics and Pragmatics of Epistemic Modals.” Semantics and
Pragmatics 8:1–81.

Muskens, Reinhard. 1996. “Combining Montague Semantics and Discourse Representation.”

Linguistics and Philosophy 19:143–186.

30
Partee, Barbara. 1973. “Some Structural Analogies between Tenses and Pronouns in English.”
The Journal of Philosophy 70:601–609.
—. 1984. “Nominal and Temporal Anaphora.” Linguistics and Philosophy 7:243–286.
Roberts, Craige. 1989. “Modal Subordination and Pronominal Anaphora in Discourse.” Lin-
guistics and Philosophy 12:683–721.
—. 1996. “Information Structure: Towards an integrated formal theory of pragmatics.” In
Jae Hak Yoon and Andreas Kathol (eds.), OSUWPL Volume 49: Papers in Semantics, 91–
136. The Ohio State University Department of Linguistics.
Rooth, Mats. 1992. “A Theory of Focus Interpretation.” Natural Language Semantics 1:75–116.
Schlenker, Philippe. 2013. “Temporal and Modal Anaphora in Sign Language (ASL).” Natural
Language and Linguistic Theory 31:207–234.
Sidner, Candice. 1983. “Focusing in the Comprehension of Definite Anaphora.” In Michael Brady
and Robert C. Berwick (eds.), Computational Models of Discourse, 267–330. MIT Press.
Smyth, Ron. 1994. “Grammatical Determinants of Ambiguous Pronoun Resolution.” Journal of
Psycholinguistic Research 23:197–229.
Stalnaker, Robert. 1978. “Assertion.” Journal of Semantics and Syntax 9:315–332.
Stanley, Jason. 2000. “Context and Logical Form.” Linguistics and Philosophy 23:391–434.
Stanley, Jason and Szabó, Zoltán Gendler. 2000. “On Quantifier Domain Restriction.” Mind
and Language 15:219–261.
Stojnić, Una. 2016a. “Content in a Dynamic Context.” Manuscript, Rutgers University .
—. 2016b. Context-Sensitivity in A Coherent Discourse. Ph.D. thesis, Manuscript, Rutgers
Universtiy.
Stojnić, Una, Stone, Matthew, and Lepore, Ernest. 2013. “Deixis (Even Without Pointing).”
Philosophical Perspectives 27:502–525.
—. 2014. “Discourse and Logical Form.” Manuscript, Rutgers University.
Stone, Matthew. 1997. “The Anaphoric Parallel Between Modality and Tense.” IRCS Report
97–06, University of Pennsylvania .
—. 1999. “Reference to Possible Worlds.” RuCCS Report 49., Rutgers University, New Brunswick
.
Veltman, Frank. 1985. Logics for Conditionals. Ph.D. thesis, University of Amsterdam.
Webber, Bonnie L., Stone, Matthew, Joshi, Aravind, and Knott, Alistair. 2003. “Anaphora and
Discourse Structure.” Computational Linguistics 29:545–587.
Willer, Malte. 2010. “New Surprises for the Ramsey Test.” Synthese 176:291–309.
Wolf, Florian, Gibson, Edward, and Desmet, Timothy. 2004. “Discourse Coherence and Pronoun
Resolution.” Language and Cognitive Processes 19:665–675.
Yalcin, Seth. 2007. “Epistemic Modals.” Mind 116:983–1026.

31
—. 2010. “Probability Operators.” Philosophy Compass 5:916–937.
—. 2011. “Nonfactualism about Epistemic Modality.” In Brian Weatherson and Andy Egan
(eds.), Epistemic Modality, 295–332. NY: Oxford Univeristy Press.
—. 2012. “A Counterexample to Modus Tollens.” Journal of Philosophical Logic 41:1001–1024.

A Appendix: Formal definitions

In this appendix, I first provide a dynamic language that formalizes my approach to modality
with anaphora (§ A.1–§ A.5), and then prove that the system preserves the validity of MP and
MT (§ A.6). Then I go on to provide a translation from a fragment of English into this language
and into a simple classical language and prove that our dynamic interpretation preserves classical
interpretation (§ A.7–A.7.4). Finally, I prove that Yalcin’s counterexample is not an instance of
MT (§ A.8).

A.1 Syntax:
In this section I specify the expressions of the language. We first start by listing the basic
vocabulary:
• Propositional expressions: the elements of the set C of constants (p, q, r...), and the elements
of the set V of variables (comp and wn for n ∈ N).
• Unary predicates: P , Q, R
• Unary operator: @
• Update expressions: K, H

• Connectives: ∧, ¬
• Identity: =
The following are atomic formulae (atoms) in our language:

• Propositional expressions are atoms.

• @P is an atom, where P is an unary predicate.

• Nothing else is an atom in our language.

These are the conditions in our language:

• All atoms are conditions.

• φ = ψ is a condition, where φ, ψ are conditions. (Stands for identity.)

• ¬φ, where φ is a condition. (Stands for negation.)

• φ ∧ ψ is a condition where φ, ψ are conditions. (Stands for conjunction.)

These are the update expressions:

32
 • hcomp := φi is an update expression, where φ is an atom.
• If φ is a condition, then [φ] is an update expression.
• K; K 0 is an update expression, if K is an update expression and K 0 is an update expression.
• might(φ, K) is an update expression, if φ is a condition and K an update expression.
• must(φ, K) is an update expression, if φ is a condition and K an update expression.
• likely(φ, K) is an update expression, if φ is a condition and K an update expression.
• if(φ, K1 , K2 ) is an update expression, if φ is a condition and K1 and K2 are update
expressions.
• and(K1 , K2 ) is an update expression, if K1 and K2 are update expressions.
• not(K) is an update expression, if K an update expression.
• Assert(K) is an update expression, if K is an update expression.
• Elab(φ, K) is an update expression, if φ is a condition and K an update expression.
• Contrast(K1 , K2 ) is an update expression, if K1 and K2 are update expressions.

A.2 Models:
I define frames and models in the usual way:
• A Frame is a tuple F = hDw , Dt = {0, 1}, R, Pi such that Dt is a domain of truth values
(Dt = {0, 1}), Dw is a finite domain of possible worlds, Dt ∩ Dw = ∅, with R, a (transitive
and reflexive) accessibility relation defined over Dw , and P, a probability measure over Dw ,
that maps each subset of Dw to [0, 1], satisfying the following constraints:
i P(Dw ) = 1
ii P(p ∪ q) = P(p) + P(q), when p and q are disjoint subsets of Dw .
iii P is a regular probability measure: if p 6= ∅ then P(p) > 0.
• A Model is a pair M = hF, Ii, where F is a frame and I an interpretation function,
which assigns to each propositional constant p a subset of Dw and each predicate constant
P a set of subsets of Dw .

A.2.1 Truth-conditional contributions of modals and conditionals:

Let us define some meta-language abbreviations that will help us state the truth-conditions asso-
ciated with updates associated with modals and conditionals. These correspond to propositions
expressed by modals and conditionals.
• Where R is the accessibility relation, and P the probability measure over sets of possible
worlds provided by the model:
Definition A.1. (Definition 2.2 in the text)
M (p, q) := {w | ∃w0 : wRw0 &w0 ∈ p & w0 ∈ q}—might q, relative to some possibility, p.
Definition A.2.

33
N (p, q) := {w | ∀w0 : wRw0 , if w0 ∈ p then w0 ∈ q}—must q, relative to some possibility p.
Definition A.3. (Definition 2.3 in the text)
P (p, q) := {w | P({w0 | wRw0 & w0 ∈ p & w0 ∈ q}) / P({w0 | wRw0 & w0 ∈ p}) > .5}—probably
q, given p.
Definition A.4. (Definition 2.4 in the text)
Cond(p, q, r) := M (p & q, r) = {w | ∀w0 : wRw0 , if w0 ∈ p & w0 ∈ q, then w0 ∈ r}—if q, r,
relative to p.

A.3 Describing operations on stacks (sequences of worlds) and sets of

stacks.
Here I define operations on stacks and sets of stacks, which I will use to define the semantics for
our language later on. Formally, a stack is just a function from a finite convex subset of N plus
comp to a set of worlds plus ⊥, where ‘⊥’ denotes an undefined value.85 (I’ll assume that ‘comp’
is a designated position on the stack. Where the stack is intended to model prominence ranking,
‘comp’ is not affecting the prominence ranking, as described in § 2.)

• Where m ∈ N, and i is a stack, im is the mth member of the stack if m is within the domain
of i, and im = ⊥ otherwise. (icomp is the member of the stack stored at the designated
position comp.)— Selecting a member of the stack.
– Where G is a set of stacks (i.e. a ‘context’), g a stack, and u a world, Gm =
th
S
g∈G {u|g m 6= ⊥ & gm = u}, for m ∈ N or m = comp.—Getting the m element in
the set of stacks G.

• For m, n ∈ N, and a stack i, im,n is a stack j defined on the set {0, ..., n − m} ∪ {comp}
such that for k ∈ N, jk = i(m+k) if j is defined on k, and jcomp = icomp .
S
– Where G is a context, and g and j are stacks, Gm,n = g∈G {j|j = gm,n } and for
H = Gm,n , Hcomp = Gcomp .

• For m ∈ N, and a stack i, im ... is the stack j defined on the set {k ∈ N | i is defined
at(m + k)} ∪ {comp} such that, for k ∈ N, jk = i(m+k) and jcomp = icomp .
S
– Where G is a context, and g, j are stacks, Gm... = g∈G {j|j = gm... } and for H =
Gm... , Hcomp = Gcomp .

• If i is a stack with a finite domain with maximal element k − 1 then for a stack j, i + j
is a stack h where, for x ∈ N, hx = ix if i is defined at x, and hx = j(x−k) otherwise (and
hcomp = icomp ).
• Where u is a world and i is a stack, u, i is a stack j, such that j0 = u, and for all n ∈ N,
such that n > 0, jn = i(n−1) if i is defined on n, and jn = ⊥ otherwise and jcomp = icomp .—
Appending to a stack.
S
– Where G is a context, u is a world, and g, j are stacks, Gu... = g∈G {j|j = u, g} and
for H = Gu... , Hcomp = Gcomp .

• g[n]g 0 iff gm = gm
0
for m 6= n (where m, n ∈ N ∪{comp}).
85 A set of numbers S is convex just in case if x ∈ S, y ∈ S and x < m < y then m is in S.

34
 • G∼
n
G0 iff {g 0 |g[n]g 0 , g ∈ G} = {g 0 |g[n]g 0 , g ∈ G0 } (where n ∈ N ∪{comp}).

• G≈
n
G0 iff {g0 ,n + gn+1... |g ∈ G0 } = G and Gcomp = G0comp .

A.4 Semantics:
The Interpretation of Atoms: The interpretation of an expression e, relative to the interpre-
tation function I a context G, and a world w:

• JpKG,w = I(p), if p ∈ C.
– Constants.

• Jwm KG,w = Gm , if wm ∈ V and m ∈ N

– Variables.

• JcompKG,w = Gcomp
– A designated position on the stack.

• J@P KG,w = ∅ if G0 = ⊥, J@P KG,w = G0 , if G0 ∈ I(P ), and J@P KG,w = J@P KG1... ,w
otherwise.
– Find the top ranked entity in G, satisfying P .

The Interpretation of Conditions:

• Jφ = ψKG,w = Dω , if JφKG,w = JψKG,w ; Jφ = ψKG,w = ∅, otherwise.

– Identity.

• J¬φKG,w = Dω \ JφKG,w .
– Negation.

• Jφ ∧ ψKG,w = JφKG,w ∩ JψKG,w .

– Conjunction.

The Interpretation of Update Expressions

• Jhcomp := piK(w, G, H) iff G comp

∼ H & Hcomp = JpKG,w

• J[φ]K(w, G, H) if and only if H = G and w ∈ JφKG,w

• JK; K 0 K(w, G, H) iff ∃G0 : JKK(w, G, G0 ) and JK 0 K(w, G0 , H)

• The following are updates that describe how propositional content (A.2.1) in context is de-
termined. Where p is a proposition (an anaphorically retrieved restrictor) and ‘@E’ denotes
the top-ranked proposition that is the subset of the epistemically accessible worlds:86
86 Forgenerality, I let the restrictor in the definition be any proposition p. However, as argued above, epistemic
modals and conditionals select the top-ranked possibility in a given context (‘@E’) as their restrictor.

35
 • Jmight(φ, K)K(w, G, H) iff there is a G0 and G00 such that JKK(w, G, G0 ) & G0 ≈
0
G00 &
G000 = G0comp ∩ J@EKG,w & G00 comp
∼ H & Hcomp = M (JφKG,w , G0comp )

• Jmust(φ, K)K(w, G, H) iff there is a G0 and G00 such that JKK(w, G, G0 ) & G0 ≈
0
G00 &
G000 = G0comp ∩ J@EKG,w & G00 comp
∼ H & Hcomp = N (JφKG,w , G0comp )

• Jlikely(φ, K)K(w, G, H) if and only if there is a G0 and G00 such that JKK(w, G, G0 ) &
G0 ≈0
G00 & G000 = G0comp ∩ J@EKG,w & G00 comp
∼ H & Hcomp = P (JφKG,w , G0comp )

• Jif(φ, K1 , K2 )K(w, G, H) iff there is a G0 , G00 , G000 and G0000 such that JK1 K(w, G, G0 ) & G0 ≈
0
00
G00 & G000 = G0comp ∩J@EKG,w & JK2 K(w, G00 , G000 ) & G000 ≈
0
G0000 & G0000 000
0 = Gcomp ∩J@EK
G ,w

& G0000 comp

∼ H & Hcomp = Cond(JφKG,w , G0comp , G000
comp )

• Jand(K1 , K2 )K(w, G, H) iff there is a G0 , G00 , G000 and G0000 such that JK1 K(w, G, G0 ) & G0 ≈
0
00
G00 & G000 = G0comp ∩J@EKG,w & JK2 K(w, G00 , G000 ) & G000 ≈
0
G0000 & G0000 000
0 = Gcomp ∩J@EK
G ,w

& G0000 ∼ H & Hcomp = G0comp ∩ G00comp

arg 0

• Jnot(K)K(w, G, H) iff there is a G0 such that JKK(w, G, G0 ) & G0 ∼ H & Hcomp =

comp
0
J¬compKG ,w

• JAssert(K)K(w, G, H) iff there is a G0 such that JKK(w, G, G0 ) & G0 ≈

0
H & H0 = G0comp ∩
J@EKG,w & w ∈ H0

In order to define the truth-conditions for updates associated with coherence relations, we assume
the following abbreviations:
Definition A.5. Elab(φ, ψ) iff φ and ψ are centered around the same event or entity, i.e. iff the
event or scenario described by ψ is a part of the event or scenario described by φ.
Definition A.6. A formula, φ, is about of body of information θ iff, where G is the input context
to φ, θ = J@EKG,w , where ‘E’ is a predicate denoting the property of being an epistemically ac-
cessible proposition, and thus, ‘@E’ denotes the top-ranked epistemically accessible proposition.
I use ‘θφ ’ to denote the body of information that φ is about.

Definition A.7. Contrast(φ, ψ) iff φ and ψ describe contrasting information about some body
of information regarding a common topic.

• JElab(φ, K)K(w, G, H) iff there are G0 and G00 such that G ≈

0
G0 & G00 = JφKG,w &
JKK(w, G0 , G00 ) & G00 ≈
0
H & H0 = G00comp & Elab(JφKG,w , H0 ).

• JContrast(K1 , K2 )K(w, G, H) iff there is a G0 and G00 such that JK1 K(w, G, G0 ) & G0 ≈
0
G00
00
& G000 = JθK1 KG,w & JK2 K(w, G00 , H) & JθK1 KG,w = JθK2 KG ,w
& Contrast(G00comp , Hcomp )

36
A.5 Truth, validity, entailment.
• K is true, relative to a context G, a world w, and a model M, if there is some H, s.t.
H 6= ∅ and JKK(w, G, H). K is false (relative to a context G, a world w, and a model M,)
otherwise.

• K is valid iff it’s true in all models.

• K1 entails K2 iff for any model M, any context G, and any world w if there is a G0 such
that G0 6= ∅ and JK1 K(w, G, G0 ), then there is a G00 such that G00 6= ∅ and JK2 K(w, G0 , G00 ).

A.6 Modus Ponens (MP) and Modus Tollens (MT):

We want to prove that MP and MT are preserved within our system. In order to do that, we
need to have a way of characterizing MT and MP correctly.

A.6.1 MT
Let us start with MT. We need to have a way of individuating instances of MT, first. Let us
begin with the following discourse:
1. Assert(if(@E, K1 , K2 )); Assert(not(K3 ))

• where ‘@E’ denotes the top-ranked epistemically accessible proposition.

What we want to show is that when the propositions expressed by K2 and not(K3 ) in their
respective contexts contradict each other, then if the proposition expressed by if(@E, K1 ,
K2 ) and not(K3 ) both hold (relative to the world of evaluation w), then the proposition cor-
responding to the intersection of the truth-conditional contribution of K1 and J@EKG,w , where
G is the input context, will be false (relative to the world of evaluation w). (Note that for our
proof it does not really matter what ‘@E’ is; the proof goes through regardless of what we take
the anaphoric restrictor to be. I use ‘@E’ because, as I argue in the paper, epistemic modals
and conditionals select top-ranked epistemic possibility in the context as a restrictor; but this is
orthogonal to our proof.)
Proof.
2. JAssert(if(@E, K1 , K2 )); Assert(not(K3 )))K(w, G, H) iff there is a G0 , such that
JAssert(if(@E, K1 , K2 ))K(w, G, G0 ) and JAssert(not(K3 ))K(w, G0 , H).
3. Take such a G, G0 and H.
4. JAssert(if(@E, K1 , K2 ))K(w, G, G0 ) iff there is a G00 , such that
Jif(@E, K1 , K2 )K(w, G, G00 ) & G00 ≈
0
G0 & G00 = G00comp ∩ J@EKG,w & w ∈ G00 .

5. Take such a G00 .

6. Jif(@E, K1 , K2 )K(w, G, G00 ) iff H 0 , H 00 , H 000 and H 0000 , such that JK1 K(w, G, H 0 )&
H0 ≈0
H 00 & H000 = Hcomp
0
∩ J@EKG,w & JK2 K(w, H 00 , H 000 ) & H 000 ≈ 0
H 0000 & H00000 =
00
000
Hcomp ∩ J@EKH ,w
& H 0000 comp
∼ G00 & G00comp = Cond(J@EKG,w , Hcomp
0 000
, Hcomp ).

7. Take such H 0 , H 00 , H 000 , and H 0000 .

37
 8. JAssert(not(K3 ))K(w, G0 , H) iff there is a G000 such that Jnot(K3 )K(w, G0 , G000 ) &
G000 ≈
0
H & H0 = G000
comp ∩ J@EK
G,w
& w ∈ H0 .

9. Take such G000 .

10. Jnot(K3 )K(w, G0 , G000 ) iff there is a G0000 such that JK3 K(w, G0 , G0000 ) & G0000 ∼ G000
comp
0000
& G000
comp = J¬compK
G ,w
.
000
11. Now, what is left to prove is that when Hcomp ∩ G000
comp = ∅ &
G,w 0 000 000 0
w ∈ Cond(J@EK , Hcomp , Hcomp ) & w ∈ Gcomp , then w ∈ / Hcomp ∩ J@EKG,w (i.e.
when the truth-conditional contribution of the small premise negates the truth-
conditional contribution of the consequent, and both premises hold at a world w,
then the antecedent does not hold at w).
000
12. Per reductio, suppose that Hcomp ∩G000
comp = ∅ & w ∈ Cond(J@EK
G,w 0
, Hcomp 000
, Hcomp )
000 0 G,w
& w ∈ Gcomp , and also w ∈ Hcomp ∩ J@EK .

13. Then, w is such that ∀w0 : wRw0 , if w0 ∈ J@EKG,w & w0 ∈ Hcomp

0
then w0 ∈ Hcomp
000
,
000
and also w ∈ Gcomp .
14. Since, w ∈ G000 000 000
comp , and Hcomp ∩ Gcomp = ∅, then w ∈
000
/ Hcomp .
0
15. Suppose that w ∈ J@EKG,w & w ∈ Hcomp . (Since R is reflexive and transitive, we
know that wRw.)
000
16. Then, w ∈ Hcomp .
17. Contradiction!
000
18. So, if Hcomp ∩ G000
comp = ∅ & w ∈ Cond(J@EK
G,w 0
, Hcomp 000
, Hcomp ) & w ∈ G000
comp , then
0 G,w
w∈/ Hcomp ∩ J@EK .

• Note that whenever the proposition expressed by not(K3 ) negates the one expressed by K2 ,
and the one expressed by not(K4 ) negates the one expressed by the proposition expressed
by the intersection of the truth-conditional contribution of K1 and J@EKG , w, where G is
the input context for the conditional, given (§ A.4) and (§ A.5), Assert(if(@E, K1 , K2 ));
Assert(not(K3 )) will entail Assert(not(K4 )).

A.6.2 MP
As with MT, we need to have a way of individuating instances of MP, first. Let us begin with
the following discourse:

1. Assert(if(@E, K1 , K2 )); Assert(K3 )

What we want to show is that when the propositions expressed by K3 and the one corre-
sponding to the intersection of the propositions expressed by K1 and ‘@E’ are identical, if the
proposition expressed by if(@E, K1 , K2 ) and K3 hold (relative to a world of evaluation w), the
proposition corresponding to the truth-conditional contribution of K2 will be true (relative to a
world of evaluation w, as well).
Proof.

38
 2. JAssert(if(@E, K1 , K2 ); Assert(K3 ))K(w, G, H) iff there is a G0 such that
JAssert(if(@E, K1 , K2 )K(w, G, G0 ) and JAssert(K3 )K(w, G0 , H).
3. Take such a G, G0 and H.
4. JAssert(if(@E, K1 , K2 )); K(w, G, G0 ) iff there is a G00 such that
Jif(@E, K1 , K2 )K(w, G, G00 ) & G00 ≈
0
G0 & G00 = G00comp ∩ J@EKG,w & w ∈ G00 .

5. Take such a G00 .

6. Jif(@E, K1 , K2 )K(w, G, G00 ) iff H 0 , H 00 , H 000 and H 0000 such that JK1 K(w, G, H 0 )&
H0 ≈0
H 00 & H000 = Hcomp
0
∩ J@EKG,w & JK2 K(w, H 00 , H 000 ) & H 000 ≈ 0
H 0000 & H00000 =
00
000
Hcomp ∩ J@EKH ,w
& H 0000 comp
∼ G00 & G00comp = Cond(J@EKG,w , Hcomp
0 000
, Hcomp ).

7. Take such H 0 , H 00 , H 000 , and H 0000 .

8. JAssert(K3 )K(w, G0 , H) iff there is a G000 such that JK3 K(w, G0 , G000 ) & G000 ≈
0
H&
H0 = G000
comp ∩ J@EK
G,w
& w ∈ H0 .
9. It is easy to show given § A.4 that if J(K3 )K(w, G0 , G000 ), whatever form K3 has, its
truth-conditional bit will be stored as a value of G000comp .

10. Suppose J(K3 )K(w, G0 , G000 ).

11. So, now we just have to prove that when G000 comp = J@EK
G,w 0
∩ Hcomp , and
G,w 0 000 0 000
w ∈ Cond(J@EK , Hcomp , Hcomp ), and w ∈ Gcomp , then w ∈ Hcomp (i.e. when the
truth-conditional contribution of the small premise and the antecedent are identical,
and both the antecedent and the small premise hold at w, then the consequent holds
at w, as well).

12. Per reductio, suppose that G000

comp = J@EK
G,w 0
∩ Hcomp , and
G,w 0 000 000 000
w ∈ Cond(J@EK , Hcomp , Hcomp ) and w ∈ Gcomp , and also w ∈
/ Hcomp .

13. Then w is such that ∀w0 : wRw0 , if w0 ∈ J@EKG,w & w0 ∈ Hcomp

0
then w0 ∈ Hcomp
000
,
000
and also w ∈ Gcomp .
000
14. By (12), w ∈
/ Hcomp .

15. Then, by (14) and reflexivity of R, ∃w0 : wRw0 , and w0 ∈ J@EKG,w & w0 ∈ Hcomp
0

and w0 ∈ 000
/ Hcomp , namely, w.

16. Contradiction!
17. So, if G000
comp = J@EK
G,w 0
∩ Hcomp 0
and w ∈ Cond(J@EKG,w , Hcomp 000
, Hcomp ) and
w ∈ G000
comp , then w ∈ H 000
comp .

• Note that whenever the proposition expressed by K3 is identical to the one corresponding
to the intersection of the propositions expressed by K1 and @E, and the one expressed by
K2 identical to the one expressed by K4 , given (§ A.4) and (§ A.5),
Assert(if(@E, K1 , K2 )); Assert(K3 ) will entail Assert(not(K4 )).

39
A.7 Relation between the dynamic (A.7.2) and classical (A.7.3) inter-
pretations
I shall now prove that my dynamic interpretation preserves a classical one. To this end, I shall
first give a dynamic translation for a fragment of English, specifying the updates associated
with utterances containing modals and conditionals. Then I will give a classical translation for
the same fragment, and prove that the dynamic interpretation preserves the truth-conditions
assigned by the classical interpretation. For ease of comparison between the two translations, I
shall avail myself of abstract level of logical forms (ALFs) for the relevant fragment of English.
The reader should bear in mind that we do not have to take a stand on the existence of a level
of representation corresponding to ALFs. This level of representation is merely a dispensable
convenience that helps us compare the two interpretations in a streamlined way.

A.7.1 Abstract Logical Forms (ALFs) for a Fragment of English:

Terms:

• Propositional constants from our base language in § A.1 (set C).

Atoms:
• All terms are atoms, and nothing else is an atom.
ALFs:

• Atoms are ALFs.

• If φ and ψ are ALFs, then might(φ, ψ) is an ALF. (Stands for “it might be the case that
ψ, given the restrictor φ”.)
• If φ and ψ are ALFs, then must(φ, ψ) is an ALF. (Stands for “it must be the case that ψ,
given the restrictor φ”.)
• If φ and ψ are ALFs, then likely(φ, ψ) is an ALF. (Stands for “it’s likely the case that ψ,
given the restrictor φ”.)
• If φ, ψ and γ are ALFs, then if (φ, ψ, γ) is an ALF. (Stands for “given the restrictor φ if
ψ, then γ.”)
• If φ and ψ are ALFs, then and(φ, ψ) is an ALF. (Stands for “φ and ψ.”)
• If φ is an ALF, then not(φ) is an ALFs. (Stands for “Not φ.”)
• If φ is an ALF, then Assert(φ) is an ALF. (Assertion operator—makes sure that the
proposition φ is asserted.)

A.7.2 Dynamic Interpretation:

I first give a translation of the relevant fragment of English into our dynamically interpreted
language (A.7.2.1), and then a translation of the same fragment into a classically interpreted
language (A.7.3)

40
A.7.2.1 Dynamic Translations
In this section, I provide a translation of the relevant fragment of English, into our dynamically
interpreted language defined in § A.1–§ A.4. I’ll assume the ALFs for the relevant fragment of
English defined in A.7.1, (e.g. might(φ, ψ) for “it might be the case that ψ”, where the modal
is anaphorically dependent on JφKG,w , for an input context G.)

(Base case, where Td (φ) is a translation of a formula φ into our dynamic system.)
• If φ is an atom, then Td (φ) = hcomp := φi.
(Recursive case)
• Td (might(φ, ψ)) = might(φ, Td (ψ))
• Td (must(φ, ψ)) = must(φ, Td (ψ))
• Td (likely(φ, ψ)) = likely(φ, Td (ψ))
• Td (if (φ, ψ, γ)) = if(φ, Td (ψ), Td (γ))
• Td (and(φ, ψ)) = and(Td (φ), Td (ψ))
• Td (not(φ)) = not(Td (φ))
• Td (Assert(φ)) = Assert(Td (φ))
Next, I shall introduce the classical interpretation of modals and conditionals in A.7.3, and then
go on to prove that our dynamic interpretation of modals and conditionals (A.7.2.1) preserves
the classical interpretation in (A.7.4).

A.7.3 Classical Interpretation

I will now introduce a classical toy modal language, and provide a translation of the relevant
fragment of English into this modal langauge, so I can compare our dynamic interpretation and
classical interpretation. Assume the following modal language:

Terms:
• Propositional terms: propositional constants (p, q, r).
Atomic formulae:
• All terms are atoms, and nothing else is an atom.
Now we introduce well-formed formulae:
• All atoms are well-formed formulae.
• If φ and ψ are formulae, then ♦(φ, ψ) is a well-formed formula.
• If φ and ψ are formulae, then (φ, ψ) is a well-formed formula.
• If φ and ψ are formulae, then φ → ψ is a well-formed formula.
• If φ and ψ are formulae, then φ ∧ ψ is a well-formed formula.
• Nothing else is a well-formed formula.

41
A.7.3.1 Classical Semantics:
I now define classical semantics for the simple modal language. I assume models are defined as in
§ A.2. (Assuming the definition of models in § A.2, and sets of sequences in § A.3), where R is an
accessibility relation provided by the model, and φ a restriction on the domain of quantification
of a modal:
• JpKG,w = w ∈ I(p)
• J♦(φ, ψ)KG,w = {w|∃w0 : wRw0 & w0 ∈ JφKG,w , w0 ∈ JψKG,w }
• J(φ, ψ)KG,w = {w|∀w0 : if wRw0 & w0 ∈ JφKG,w then w0 ∈ JψKG,w }
• Jφ ∧ ψKG,w = JφKG,w ∩ JψKG,w
• J(φ ∧ ψ) → γKG,w = {w|∀w0 : if wRw0 & w0 ∈ JφKG,w ∩ JψKG,w then w0 ∈ JγKG,w }

A.7.3.2 Classical Translation:

Assuming the same ALFs for the relevant fragment of English as in A.7.1, I now define the
following translations of the relevant bits of the fragment into our classical language.
Where Tc (p) is a translation of a formula p into classical system:
• Where φ is an atom Tc (φ) = JφKG,w
• Tc (might(φ, ψ)) = J♦(φ, ψ)KG,w
• Tc (must(φ, ψ)) = J(φ, ψ)KG,w
• Tc (if (φ, ψ, γ)) = J(φ ∧ ψ) → γKG,w
This translation does not capture the systematic effects of context on the interpretation of
modals: it doesn’t control for where the restrictor comes from. But, note that we don’t have to
care about what the restrictor φ is. We can simply assume that it is the domain of all possible
worlds (Dw ), the set of epistemically live worlds, or any other proposition. So long as all the
anaphoric restrictors are denoting the same proposition, what the restrictor is won’t matter,
because by resolving the restrictor always to the same proposition we are making sure that the
domain of quantification for modals is held constant throughout, as in a simple classical modal
logic.

A.7.4 Proof.
• We want to prove that our dynamic interpretation (A.7.2) preserves the classical inter-
pretation (in A.7.3). In particular, we prove that for any Td (p), if JTd (p)K(w, G, H), then
Hcomp = JTc (p)KG,w , where JTc (p)KG,w is the corresponding translation of the formula p in
classical system; Assert(Td (p)) guarantees that JTc (p)KG,w is true at the actual world. We
do a proof by induction.
* Base case. First prove that for an atom p, and translation Td (p), if JTd (p)K(w, G, H), then
Hcomp = JpKw .

Proof.
2. By (A.7.2.1), Td (p) = hcomp := pi.
3. By (§ A.4), we have Jhcomp := piK(w, G, H) iff G ∼ H & Hcomp = JpKG,w .
comp

42
 4. Suppose Jhcomp := piK(w, G, H).
5. By (2)–(4), Hcomp = JpKG,w , and JpKG,w iff w ∈ I(p). Thus, Hcomp = JTc (p)KG,w ,
which we were set to prove.

* Recursive case.

Proof.
• IH: Assume that for a formula φ of the depth k or less, if JTd (φ)K(w, G, H), then Hcomp =
JφKw , where JφKw is the corresponding classical interpretation of the formula φ.
• Consider a formula φ of the depth k + 1. We prove that the IH holds for the possible ways
of constructing φ:
i Let φ =might(χ, ψ). Then, by A.7.2.1, Td (φ) = might(χ, Td (ψ)). Suppose that
Jmight(χ, Td (ψ))K(w, G, H). We know by (§ A.4) that Jmight(χ, Td (ψ))K(w, G, H)
iff there is a G0 and G00 such that JTd (ψ)K(w, G, G0 ) & G0 ≈
0
G00 & G000 = G0comp ∩ J@EKG,w
& G00 comp
∼ H & Hcomp = M (JχKG,w , G0comp ). Take such a G0 and G00 . We have that
JTd (ψ)K(w, G, G0 ); thus, by IH, G0comp = JψKG,w . Then, since Hcomp = M (JχKG,w , G0comp ),
given Definition (A.1) and (A.7.3.2), Hcomp is equivalent to JTc (might(χ, ψ))KG,w , by simple
math.
ii Let φ = must(χ, ψ). Then, by A.7.2.1, Td (φ) = must(χ, Td (ψ)). Suppose that
Jmust(χ, Td (ψ))K(w, G, H). So, by (§ A.4), there is a G0 and G00 such that JTd (ψ)K(w, G, G0 )
& G0 ≈0
G00 & G000 = G0comp ∩ J@EKG,w & G00 comp ∼ H & Hcomp = N (JχKG,w , G0comp ).
Take such a G0 and G00 . Since JTd (ψ)K(w, G, G0 ), by IH, G0comp = JψKG,w . Then, since
Hcomp = M (JχKG,w , G0comp ), given Definition (A.2) and (A.7.3.2), Hcomp is equivalent to
JTc (must(χ, ψ))KG,w , by simple math.
iv Let χ = if(χ, ψ, γ). Then, by A.7.2.1, Td (χ) = if(χ, Td (ψ), Td (γ)). Suppose that
Jif(χ, Td (ψ), Td (γ))K(w, G, H). So, by (§ A.4), we know that there is a G0 , G00 , G000 and G0000
such that JTd (ψ)K(w, G, G0 ) & G0 ≈0
G00 & G000 = G0comp ∩ J@EKG,w & JTd (γ)K(w, G00 , G000 ) &
00
G000 ≈
0
G0000 & G0000 000
0 = Gcomp ∩J@EK
G ,w
& G0000 comp
∼ H & Hcomp = Cond(JχKG,w , G0comp , G00comp ).
Take such a G0 , G00 , G000 and G0000 . Since JTd (ψ)K(w, G, G0 ), by IH we know that G0comp =
JψKG,w and since JTd (γ)K(w, G00 , G000 ), we know by IH that G000 comp = JγK
G,w
. Then, since
G,w 0 000
Hcomp = Cond(JχK , Gcomp , Gcomp ) given Definition (A.4) and (A.7.3.2), Hcomp is equiv-
alent to JTc (cond(χ, ψ, γ))KG,w , by simple math.

A.8 Yalcin’s Counterexample:

I shall now formally represent Yalcin’s counterexample in our system and prove that the coun-
terexample is not an instance of MT (and, thus, is not a counterexample). Recall the two
premises of the counterexample:
(1) If the marble is big, then it’s likely red.
(2) The marble is not likely red.
Let I(p) be the proposition that the marble is big, and I(q) be the proposition that the marble
is red. Then we represent the two premises of the counterexample as in (32), repeated below:

43
 (32) Contrast(Assert(if(@E, hcomp := pi, Elab(w0 , Likely(@E, hcomp := qi)))),
Assert(not(Likley(@E, hcomp := qi))))
Let ‘K1 ’ stand for ‘if(@E, hcomp := pi, Elab(w0 , Likely(@E, hcomp := qi)))’ and ‘K2 ’ for
‘not(Likley(@E, hcomp := qi))’. Then, we show that (32) is not an instance of MT.
Proof. First, we shall calculate the truth-conditions expressed by the consequent of the big
premise, then, we shall calculate the truth-conditions expressed by the small premise, and then
we show that the two do not contradict each other.

1. JContrast(Assert(K1 ); Assert(K2 ))K(w, G, H) iff there is a G0 and G00 such that

JAssert(K1 )K(w, G, G0 ) & G0 ≈
0
G00 & G000 = JθAssert(K1 ) KG,w & JAssert(K2 )K(w, G00 , H)
00
& JθAssert(K1 ) KG,w = JθAssert(K2 ) KG ,w &
Contrast(G00comp , Hcomp ). (By (§ A.4).)
2. Take such a G, G0 , G00 andH. Then JAssert(K1 )K(w, G, G0 ) and
JAssert(K2 )K(w, G00 , H). (By (1).)
3. First we calculate the truth-condition of the consequent of the big premise.
JAssert(K1 )K(w, G, G0 ) iff there is a G000 such that JK1 K(w, G, G000 ) & G000 ≈
0
G0 &
G00 = G000
comp ∩ J@EK
G,w
& w ∈ G00 . (By (§ A.4).)

4. Take such G000 . JK1 K(w, G, G000 ) iff there is a H 0 , H 00 , H 000 and H 0000 such that
Jhcomp := piK(w, G, H 0 ) & H 0 ≈
0
H 00 & H000 = Hcomp
0
∩ J@EKG,w &
JElab(w0 , Likely(@E, hcomp := qi))))K(w, H 00 , H 000 ) & H 000 ≈
0
H 0000 & H00000 =
00
000
Hcomp ∩ J@EKH ,w
& H 0000 comp
∼ G000 & G000
comp = Cond(J@EK
G,w 0
, Hcomp 000
, Hcomp ).
(By (§ A.4) and the definition of K1 .) Note that by § A.4 and A.7.2.1, Hcomp
000
stores
the truth-conditions of the consequent of the big premise.

5. Take such H 0 , H 00 , H 000 and H 0000 .

JElab(w0 , Likely(@E, hcomp := qi))))K(w, H 00 , H 000 ) iff there are some J and J 0
00
such that H 00 ≈ 0
J & J0 = Jw0 KH ,w & JLikely(@E, hcomp := qi)K(w, J, J 0 ) &
00
J0 ≈
0
H 000 & H0000 = Jcomp
0
& Elab(Jw0 KH ,w
, H0000 ). (By (§ A.4).)

6. Take such J and J 0 . JLikely(@E, hcomp := qi)K(w, J, J 0 ) iff there is a J 00 and

J 000 such that Jhcomp := qiK(w, J, J 00 ) & J 00 ≈
0
J 000 & J0000 = Jcomp
00
∩ J@EKJ,w &
J 000 comp
∼ J 0 & Jcomp
0 00
= P (J@EKJ,w , Jcomp ). (By (§ A.4).)

0
7. Thus, by (A.3), Jcomp = {w|P({w0 | wRw0 & w0 ∈ J@EKJ,w & w0 ∈ Jcomp
0
}) /
0 0 0 J,w
P({w | wRw & w ∈ J@EK }) > .5}.
0 000
8. By 5 and the definition of ‘≈
n
’, Jcomp = Hcomp .

000
9. By (4)–(6), and (A.7.2.1), Hcomp = {w|P({w0 | wRw0 & w0 ∈ JpKG,w ∩ J@EKG,w
& w ∈ I(q)}) / P({w | wRw & w0 ∈ JpKG,w ∩ J@EKG,w }) > .5}. These are the
0 0 0

truth-conditions expressed by the consequent of the big premise. Now we calculate

the truth-conditions of the small premise.

44
 10. From (2) we have: JAssert(K2 )K(w, G00 , H).
By (§ A.4), JAssert(K2 )K(w, G00 , H) iff there is a I such that JK2 K(w, G00 , I) &
00
I≈0
H & H0 = Icomp ∩ J@EKG ,w & w ∈ H0 .

11. Take such I. By definition of K2 and (§ A.4), we have: JK2 K(w, G00 , I) iff there
is a I 0 such that JLikley(@E, hcomp := qi)K(w, G00 , I 0 ) & I 0 ∼ I & Icomp =
comp
0
J¬compKI ,w
.
12. Take such I 0 . Then, by (§ A.4), JLikley(@E, hcomp := qi)K(w, G00 , I 0 ) iff there is a
00
I 00 and I 000 such that Jhcomp := qiK(w, G00 , I 00 ) & I 00 ≈
0
I 000 & I0000 = Icomp
00
∩J@EKG ,w
00
& I 000 comp
∼ I 0 & Icomp
0
= P (J@EKG ,w 00
, Icomp )

13. By (11), (6), and (A.3), the truth-condition expressed by the small premise is as
00
follows: Dw \ {w|P({w0 | wRw0 & w0 ∈ J@EKG ,w & w0 ∈ Icomp00
}) / P({w0 | wRw0
00
& w0 ∈ J@EKG ,w }) > .5}, where Dw is the domain of possible worlds from the
model.

14. By (1), (A.6), (§ A.4) and (A.7.2.1), we have: Dw \ {w|P({w0 | wRw0 & w0 ∈
J@EKG,w & w0 ∈ I(q)}) / P({w0 | wRw0 & w0 ∈ J@EKG,w }) > .5}., where Dw is the
domain of possible worlds from the model. This is the truth-condition expressed by
the small premise.
15. From (9) and (13), we see that the truth-conditions expressed by the big premise
and the consequent of the small one do not contradict each other.87 Hence, (32)
does not correspond to the premises of an instance of MT.

87 Yalcin’s scenario with an urn with 100 marbles can be used to generate a model in which both propositions

are true. In particular, suppose the domain of worlds Dw is partitioned according to a color-size distribution:
into big and blue, small and blue, big and red and small and red worlds. Where I(p) is the proposition that
the marble is big, I(q) the proposition that the marble is red, I(r) the proposition that the marble is blue and
I(s) the proposition that the marble is small, let the probability measure assign the following probabilities:
P(I(p) ∩ I(q)) = .3, P(I(p) ∩ I(r)) = .1, P(I(s) ∩ I(q)) = .1, and P(I(s) ∩ I(r)) = .5.

45