4.1 Necessitation by Attitude (NBA)

Looking at the instances when harmonizing readings arise, we first note an empirical generalization. In examples (1a), (2a), and (3a), a modal inference seems to arise naturally from the meaning of the attitude report in the antecedent of the conditional, which happens to match in flavor with the modal in the consequent. In (1a), we can infer from the attitude report in the antecedent You want to go to Harlem the teleological necessity that you have to_T go to Harlem. In (2a), we can infer from the attitude report The Provost requests that grades be posted on Monday the deontic necessity that grades must_D be posted on Monday. And in (3a), we can infer from Sherlock thinks that the crime occurred at 6pm the epistemic necessity that the crime must_E have occurred at 6pm. Schematically, the generalization can be stated as follows:

Before we detail our account, two notes are in order about the modal □_F in (20).⁶ First, necessities are expressed in English either by must or have to (should and ought to express weak necessity). In the theoretical literature, these modals are assumed to express a variety of different flavors. But in fact, at least in colloquial North American English, must is almost always used epistemically (Tagliamonte & D’Arcy 2007); deontic uses can sound formal, and bouletic and teleological uses are almost impossible. Meanwhile the semi-modal have to is natural for teleological and bouletic necessities, and for many cases of epistemic modality (see Williams et al. 2025). We use □_F in (20), to state our generalization, which is intended to stand for either must or have to, depending on flavor. Second, one might wonder if the NBA that a verb like want tends to trigger might not better be labeled as bouletic (desire-based), rather than teleological (goal-based). As mentioned in footnote 1, distinguishing bouletic and teleological modality is difficult, and some may question whether genuine cases of the former truly exist.⁷ Whether there is a genuine distinction between bouletic and teleological modality in fact doesn’t matter to us. What matters is that utterances of the form you want to p seem to routinely trigger an inference that p is necessary given what you want, expressible with a necessity modal in English like have to, which we might call either bouletic or teleological.

In all cases where harmonizing goes through, the subject of the attitude report can be viewed as an authority in the relevant modality. With teleological modality, as in (1a), we routinely take people to be authorities on their own goals and desires: obtaining what one wants should fulfill one’s goals. Thus you are an authority for teleological necessities relative to your desires: you are naturally compelled to have your own desires fulfilled. However, you may be less compelled to fulfill someone else’s desires. Thus, when we change the subject from you to Satan in (19a), the NBA inference fails to go through without further assumptions. This is because we are unlikely to infer from Satan wants p that p is teleologically necessary for you, unless we know that Satan’s desires dictate your own. If we enrich the context with the assumption that you are a Satanist, a harmonizing reading becomes possible. We will return to the meaning of want and teleological necessity, and the fairly automatic nature of the inference from the former to the latter in §4.4.

With deontic modality, as in (14a), we take the Provost to be a legal authority on university policies. Thus, we can infer from The Provost requests that p that it is deontically necessary that p: the NBA inference goes through. However, this inference disappears in (18a), when the subject the Provost is replaced with Chad, a student with no authority on university policies.

Lastly, with epistemic modality, as in (16a), we readily take the revered detective Sherlock Holmes to be an epistemic authority, and when we do, we infer from Sherlock Holmes thinks p, that it is epistemically necessary that p: the NBA inference goes through. However, in (17a), this inference disappears when the subject is replaced by Sherlock’s dense acolyte Watson, whose epistemic authority we do not trust.

To sum up, Necessitation by Attitude is an inference that arises naturally from the meaning of an attitude report in the antecedent of a conditional, when the attitude holder is assumed to be an authority in the relevant modality. When the flavor of the modal inference matches the modal in the consequent, we derive a harmonizing reading, as discussed in the next section.

Note that we take NBA inferences to be run-off-the-mill, defeasible inferences, based on world knowledge. They are similar to the kind of inference one draws from Al is a basketball player to the assumption that Al is tall. Just like with NBA, such inferences can pragmatically enrich a conditional: when one hears If Al is a basketball player, he needs an extra long bed, one naturally restricts the interpretation of the consequent to worlds where Al is tall, something one might not do if we were to replace basketball player with skateboarder. Finally, we take the NBA inference to be a general inference. Thus, it should hold not just in our world, but in all worlds that we take to be like ours. We will thus assume that the NBA holds in all worlds that are epistemically accessible. This importantly will allow the NBA inference to hold in conditionals, which we assume are implicitly quantified by an epistemic modal NEC, following the literature discussed in §2.⁸

4.2 Flavor Matching

When a modal inference is triggered by Necessitation by Attitude (NBA), a harmonizing reading for a conditional becomes possible if the flavor of the modal in the consequent of the conditional matches the flavor of the modal in the NBA. We refer to this condition as Flavor Matching:

In (1a), the flavor of the NBA modal is teleological, which is the flavor of the modal in the consequent (both modals are anchored to the addressee’s goals). In (2a), both modals are deontic, and in (3a), both modals are epistemic.

Importantly, we do not take the domain of the consequent modal to be inherited anaphorically from the attitude verb in the antecedent, or by any semantic mechanism. Rather, we assume that the domain of the consequent modal is restricted by the prejacent of the NBA-derived modal inference indirectly, from the fact that the two modals happen to quantify over the exact same sets of worlds. We justify this pragmatic proposal over semantic or syntactic alternatives in §5.

We propose that harmonizing arises pragmatically when the modal inference derived from the meaning of an attitude report in the antecedent (NBA) matches in flavor with the modal in the consequent (Flavor Matching), as schematized below. Take a conditional of the form If subject Attitude p, must_F q in (22a). Via NBA, we derive the necessity inference must_F p in all epistemically accessible worlds in which subject Attitude p, as sketched in (22b). Flavor Matching occurs when the NBA modal and the consequent modal match in flavor, as shown in (22c). (22d) shows the meaning of (22a) enriched with the inference in (22b), in boldface. Given that both the NBA modal and the consequent modal quantify over the same set of F worlds, it looks as if the consequent modal is directly restricted by the complement of the attitude verb, viz. the prejacent p of the NBA modal (in boldface). (22d) can be simplified as in (22e),⁹ giving the illusion that the complement of the attitude in the antecedent directly restricts the consequent modal:

The following illustrates harmonizing with the textbook Harlem conditional, repeated in (23a). First, we assume that you are an authority on your own desires. Thus, through NBA, we obtain the modal inference in (23b). Since the teleological flavor of the modal of the NBA inference matches the teleological flavor of the consequent modal, Flavor Matching occurs, as in (23c). This leads to the harmonizing reading in (23d), where the meaning of (23a) used in this context is enriched with the NBA inference (indicated in boldface). As both the NBA and the consequent modals quantify over the same set of worlds, (23d) can be simplified as in (23e), giving the illusion that the complement of want (you go to Harlem), which is the prejacent of the NBA modal, directly restricts the consequent modal:

Harmonizing readings are prevented when either step, NBA or Flavor Matching, fails to hold. As we saw in the previous section, NBA can fail when the attitude holder is not taken to be an authority on the relevant modality. For instance, Satan may not be an authority on your desires, and thus it may not follow that you have to_Tyou go to Harlem from the fact that Satan wants you to go. Harmonizing can also be prevented when Flavor Matching does not hold, for instance, if the consequent modal in (23a) is interpreted deontically, rather than teleologically. In the next section, we show how to derive harmonizing readings in more formal detail.

4.3 Deriving harmonizing conditionals

As in previous accounts, we assume the independently motivated double modal structure for conditionals (Frank 1997, Huitink 2005, von Fintel & Iatridou 2005, Condoravdi & Lauer 2016), according to which conditionals with an overt modal in the consequent involve an additional covert epistemic necessity modal (NEC), which the antecedent restricts. We assume that this double modal structure is always possible, not just with overt root modals (e.g., deontic, teleological) in the consequent, but with overt epistemic modals as well. We however remain agnostic as to whether the covert modal is always present, or only optionally, so long as it is present when harmonizing arises.

Assuming a double modal structure, the Harlem conditional has the LF in (24a), where the antecedent restricts NEC. We get the truth conditions in (24b), assuming the lexical entry for want in (24c):¹⁰

Through Necessitation by Attitude (NBA), we derive the teleological necessity inference in (25). Since the NBA holds across all epistemically accessible worlds, it will hold in the worlds w’ quantified over by nec:

Given that the NBA derives a teleological necessity, Flavor Matching holds: both the NBA modal and the teleological modal in the consequent quantify over worlds compatible with your goals in the same world of evaluation w’, namely, the worlds quantified over by the epistemic modal NEC in which the antecedent holds. Thus, we obtain that in all relevant teleological worlds, you go to Harlem (from NBA), and in all of these worlds, you take the A train (from the consequent). (26b) shows the truth conditions in (24b) enriched with the NBA inference in (25) (in boldface), and simplified in (26c):

These enriched truth conditions give the illusion that the complement of the attitude verb directly restricts the consequent modal. However it does so indirectly, via the additional modal inference, derived pragmatically from the meaning of the attitude report in the antecedent, together with uncontroversial background assumptions (e.g., going to Harlem is achievable, etc.). Thus harmonizing solves the non-compositionality problem of anankastic conditionals pragmatically. It also solves the conflicting goals problem in the process, by having all of the worlds quantified over by the consequent modal be worlds in which you go to Harlem, thereby preventing any Hoboken worlds.

Note that we can derive the same harmonizing reading with other desire attitudes, as in (13), using the same mechanism. Harmonizing should occur whenever the NBA inference You have to_T go to Harlem can be inferred from You {intend/hope/would like/wish} go to Harlem.

The harmonizing reading of the deontic conditional repeated in (27a) works the same way. Again, we assume a double modal structure, such that the antecedent restricts nec, as shown in (27b), with the standard truth conditions in (27c), using the denotation for request in (27d):¹¹

Through NBA, we derive the deontic necessity inference in (28) (again we assume that the NBA inference holds across all epistemically accessible worlds):

The flavor of both the NBA modal and the consequent modal is deontic: the modals quantify over the same sets of worlds, and Flavor Matching obtains. (29a) shows the truth conditions in (27c) enriched with the (boldfaced) NBA inference in (28b), and simplified in (29b). Here again, the complement of the attitude report in the antecedent seems to restrict the consequent modal, but it does so via the addition of the defeasible NBA inference.

We predict that similar harmonizing readings can arise with other attitudes used to report rules, e.g., if we replace request with verbs say, mandate, or declare, via the same mechanism. We should get a harmonizing reading whenever the NBA inference Grades must_D be posted on Monday can be inferred from attitude or speech reports like If the Provost {mandates/says/declares} that grades are posted Monday….

We finally turn to harmonizing readings with epistemic conditionals, as in (3a), repeated in (30a) below. Here again, we assume a double modal structure for conditionals, so that (30a) has the LF in (30b). This assumption is crucial for us, so that the NBA inference can be relative to the same worlds as the consequent modal.¹² For concreteness, we assume that epistemic modals like must take an epistemic modal base and an optional stereotypical ordering source Kratzer (1981), and that think has the lexical entry in (30d). Given all this, the truth conditions for (30a) are given in (30c):

Here again, the NBA inference goes through (in all epistemically accessible worlds), assuming that we take Sherlock to be an epistemic authority:

And here again, we can enrich the truth conditions in (30c) with the NBA inference in (31), as shown in (32a). Given Flavor Matching, as the NBA modal and the consequent modal quantify over the same worlds (i.e., epistemic worlds anchored to the worlds quantified over by NEC), (32a) can be simplified as in (32b), once again giving the illusion that the complement of the antecedent’s attitude verb directly restricts the consequent modal.

We predict that similar harmonizing readings can arise with other attitude reports (e.g., when we replace think in (30) with verbs like say, believe, or assume), using the same mechanism. We should get a harmonizing reading whenever the NBA inference The crime must_E have occurred at 6pm can be inferred from attitude or speech reports like Sherlock {said/believes/assumes} that the crime occurred at 6pm….

In sum, we propose that anankastic conditionals, and harmonizing conditionals more generally, are the result of a modal inference derived from reasoning about the meaning of an attitude verb in the antecedent, a modal inference whose flavor matches the modal in the consequent. Our pragmatic account derives these readings without postulating any lexical idiosyncrasies for particular attitude verbs, or types of modality. It captures the context-sensitive nature of the availability of the readings, which, as we saw, seems to depend on how easily the attitude holder can be viewed as an authority in the relevant modality. In the next subsection, we show how to avoid harmonizing for regular want conditionals, and then justify further a pragmatic view of harmonizing over a semantic, syntactic or lexical alternative in §5.

Before we move on, we would like to note a parallel between the NBA inference at the heart of our pragmatic proposal, and parenthetical readings that arise with attitude reports. As first noted by Urmson (1952), attitude reports can sometimes give rise to so-called parenthetical readings, where the speaker seems to endorse the attitude, and somehow proffer its content (Hooper 1975, Simons 2007). (33a), for instance, can be used to indirectly assert (33b). This indirect assertion arises when the speaker takes the attitude holder, Mom, to be an authority on when bedtime is, and endorses her assertion, thus indirectly proffering that it is time for bed. As with the NBA, the endorsement is pragmatic, and the proffering indirect.

As with the NBA, the availability of such parenthetical readings is sensitive to the context, and notably to the reliability of the subject of the attitude holder. Replacing Mom with an individual who the speaker does not take to be a trusted authority, such as Bozo the Clown, seems to prevent the indirect assertion:

4.4 want vs. teleological have to and non-anankastic readings of want conditionals

The pragmatic nature of our proposal for anankastic conditionals was initially motivated by the fact that harmonizing goes beyond want and teleological modality. However, focusing on the original cases with want, one might wonder why the anankastic reading is so robust that it seemed to warrant syntactic or lexical treatments in the previous literature. How systematic are NBA inferences from you want p to have to_Tyou p? And if they are systematic, how do we obtain regular, non-anankastic readings of want conditionals, like the one in (4), repeated in (35) below? Recall that when (35) is uttered as advice to a recovering alcoholic, the intended meaning is not that drinking requires that you think about your family, but rather that the desire to drink does: in fact, the advice seems to be that you should think about your family in order not to drink. In this section, we briefly review meaning differences between want and teleological have to, and show how to derive non-anankastic readings for conditionals like (35).

We believe that the robustness of harmonizing readings in textbook anankastic conditionals is due to the quasi-automaticity of the NBA inference. Indeed, teleological have to and want seem almost synonymous: both express a strong preference for a state of affairs. However, their meanings differ in at least one important way: want is parasitic on belief, teleological have to is not. While the precise semantics for want is subject to great debate in the literature (see e.g., Heim 1992, von Fintel 1999, Villalta 2008, Condoravdi & Lauer 2016), most would agree that it expresses a preference of the subject for a state of affairs, which has to be compatible with the subject’s (sometimes fallible) beliefs. Thus, we can say that Al wants to sell her cello, even if she doesn’t own a cello, but is under the misimpression that she does. Neither teleological nor bouletic have to (if distinct), however, rest on beliefs. In the Kratzerian tradition, all root modals are relative to a circumstantial modal base, which picks out worlds compatible with relevant facts in the world of evaluation, and quantifies over the worlds among those that best fit a salient set of priorities (like laws, goals, or desires—often the subject’s). Teleological have to thus lacks want’s subjectivity: we cannot say that Al has to sell her cello, if she doesn’t own one, regardless of what Al believes.

According to Phillips-Brown (2019), the doxastic nature of want turns out to be fatal for Condoravdi and Lauer’s account, according to which want can be used to express two readings: pure desire, and effective preference, a special kind of preference that guides actions. The latter is what they assume is involved in anankastic conditionals, to guarantee that the expressed desire outranks other conflicting ones. Condoravdi and Lauer propose that an agent’s preferences (e.g., desires) are arranged in a preference structure, where they are ranked by how important they are to the agent. These preferences can be inconsistent, thus accommodating the intuition that we sometimes want incompatible things. However, to act on a desire or preference, one has to consolidate all of one’s desires into an action-guiding set of preferences that is free of contradictions, which they refer to as an effective preference structure. A consistency condition requires that one’s effective preferences cannot be inconsistent with respect to one’s beliefs. For example, one cannot effectively want to go to Harlem and effectively want to go to Hoboken, while believing that one can’t do both. Anankastic readings, they say, arise with effective preference want, which guarantees that the stated desire (e.g., going to Harlem) outranks all others that conflict with it (e.g., going to Hoboken). Phillips-Brown (2019) shows that this account makes the wrong predictions in a scenario where it is compatible with your beliefs that one can both go to Hoboken and to Harlem, but not in reality. In such a scenario, it is still judged true that If you want to go to Harlem, you have to take the A train, but Condoravdi and Lauer predict that it should be false, as going to Hoboken should not be ruled out. Our analysis does not run into this problem, as it does not rely on the wanter’s beliefs. A harmonizing reading arises through the inference that you have to_Tyou go to Harlem, where have to’s modal base is relative to relevant facts in the world of evaluation, not the wanter’s beliefs. The harmonizing reading thus comes out as true, regardless of what your beliefs are.

While want and teleological have to differ in subjectivity, the inference from the former to the latter always seems possible. If so, how do we prevent a harmonizing reading for (35)? Ultimately, we believe that the lack of harmonizing in (35) stems from a lack of Flavor Matching, and not necessarily from a lack of NBA. Indeed, a teleological necessity inference seems possible in (35) from You want to drink to You have to_Tyou drink: given your desires for immediate pleasure, it is teleologically necessary that you drink. However, the consequent modal seems to express a teleological necessity relative to a different set of your (longer term) goals or desires: desires to be healthy, to have a good relationship with your family, etc. It seems that both teleological have to and want can be relative to a host of different desires, even within a single individual, and that such desires can even be expressed within the same utterance. In a context in which I am conflicted between these two kinds of desires (immediate vs. long term), I can say: I want to drink and I want to not drink, or I have to_Tme drink and I have to_Tme not drink. Given the fine-graininess of the desires that both want and teleological have to can express, we take it that the lack of harmonizing in (35) is due to a lack of Flavor Matching. The NBA modal and the consequent modal are relative to different desires of yours: the former, relative to your desires for immediate pleasure, the latter, to your longer term desires. The modals then quantify over different worlds, and we do not get a harmonizing, anankastic reading.¹³

5.1 A lexical alternative?

First, let’s consider a lexical approach, which would essentially generalize Condoravdi & Lauer (2016)’s account of want to other attitude predicates: various attitude verbs could have an effective reading, where the attitude complement would somehow be taken to hold in the relevant modality. Request or ask could, for instance, have readings that express requirements that trump others, verbs like think might have a true belief reading, etc. It is unclear to us whether such an account could be made for all attitudes. But even if it could, we believe that a lexical implementation would still both over- and undergeneralize. First, a lexical approach would have to be supplemented by some pragmatic principles, to explain why effective readings depend on the reliability of the subject of the attitude verb in the relevant modality. This could well end up recapitulating the NBA, making the purported systematic ambiguity redundant. More decisively, a lexical approach would undergenerate, in that it would have to assume a fairly strict mapping between attitude meaning and modal flavor of the harmonizing modal. However, this mapping seems to be fairly flexible, and largely determined by context. For instance, the use of want with a subject considered a legal authority can trigger a harmonizing deontic reading, as illustrated in (36a). This reading is derived straightforwardly in our pragmatic account, given that in such a context, the NBA inference in (36b) seems warranted:

5.2 A syntactic alternative?

Harmonizing could also be implemented syntactically, by having a syntactic counterpart of the attitude verb’s complement restrict the consequent modal, for instance, via a covert given clause. Such an account could be viewed as an extension of von Fintel & Iatridou (2005)’s treatment of anankastic conditionals beyond teleological modality.

Von Fintel and Iatridou observe that anankastic conditionals like the Harlem sentence can be paraphrased using a purpose clause as in (37a), and propose that an elliptical purpose clause is also present in anankastic conditionals, as an argument of a teleological modal, whose content is supplied by the if-clause, as in (37b). They propose that teleological modals take two ordering sources, a primary ordering source (the designated goal), supplied by the elliptical purpose clause, which takes precedence over any other goals, and a secondary ordinary ordering source. The designated goal guarantees that the worlds quantified over by the modal are worlds in which you go to Harlem:

While this purpose clause analysis is tied to teleological modality, one might propose an extension of this idea, by syntacticizing the NBA inference in a covert given-clause, as sketched in (38). The content of the modal inference would have to be inherited from the antecedent through ellipsis, requiring a linguistic antecedent (underlined in (38)):

It seems, however, that harmonizing readings can arise even in cases where the prejacent of the NBA modal (the modal in the given-clause in this syntactic implementation) lacks a true syntactic antecedent. Consider a scenario where both speaker and addressee know that Fletcher is a pathological liar, who only says the opposite of what is true. Fletcher claims that Al stole your money. In this scenario, the conditional in (39) seems to have a harmonizing reading, where the negation of the proposition expressed by the attitude verb’s complement seems to restrict the consequent modal, though there is no linguistic constituent Al didn’t steal your money that could serve as antecedent for the covert clause.¹⁴

This reading is, however, predicted by our pragmatic account, given that we can derive the NBA inference in (40) from the meaning of the antecedent together with the contextual assumptions given:

Note that in this case again, we find a parallel in behavior with parenthetical uses of attitude reports, suggesting that parenthetical uses and NBA inferences are made of the same pragmatic cloth. Indeed, if it is common ground that Fletcher is a pathological liar, a speaker could use an utterance of Fletcher said that it’s raining to proffer that it is not raining. We take these kinds of examples to support the view that harmonizing readings are pragmatic in nature, and cannot be straightforwardly derived syntactically.

5.3 A semantic alternative?

A final potential implementation for harmonizing conditionals is a semantic one, where harmonizing readings would arise via modal subordination (e.g., Roberts 1989, 2020, Brasoveanu 2010), where the domain of the consequent modal would be determined anaphorically from the content of the attitude in the antecedent.

Stone (1999) and Sudo (2014) argue that attitude verbs like want can introduce a set of worlds that a subsequent modal can then be anaphoric to. This anaphoric relation could hold between the attitude in the antecedent and the modal in the consequent, assuming some kind of donkey-anaphoric treatment: roughly, (1a) would express that you take the A train in all of those worlds compatible with your desires in which you go to Harlem.

Assuming that anankastic conditionals result from modal subordination would easily extend to the other harmonizing cases that we have seen. However, it would also generate harmonizing readings where they do not seem to be attested. Indeed, we would expect that a modal could be anaphoric to any attitude report, including in cases where the speaker doesn’t take the subject to be an authority in the relevant modality, as it isn’t clear why speaker endorsement should matter for anaphoricity. Tricky cases of undergeneration discussed for the syntactic approach, such as (39a), would also be problematic for a semantic approach.

To sum up, we have seen that harmonizing readings can be found in all modal flavors and with a wide range of attitude verbs in the antecedent. We have also seen that their distribution is highly context-sensitive, in a way that semantic, lexical, or syntactic accounts can not derive straightforwardly. We conclude that there is nothing semantically nor syntactically special about anankastic conditionals and their kin: harmonizing is a pragmatic phenomenon.

6.1 Near-anankastic conditionals

As discussed in Condoravdi & Lauer (2016), the problems with anankastic conditionals are present in cases beyond want conditionals expressing a means-to-an-end relation. Consider for instance the example in (41):

A consequent-focused conditional like (41) doesn’t express that spending five days is necessary for going to Disneyland; instead, it is about what you have to do in worlds where you do go to Disneyland. Such examples are problematic for von Fintel & Iatridou (2005), as a purpose clause paraphrase (To go to Disneyland, you have to spend at least five days there) doesn’t seem to capture the meaning of (41).

Our proposal accounts for these conditionals straightforwardly: (41) gets the LF in (42a), and the standard truth conditions in (42b). We assume that the NBA inference in (42c) goes through.

Given the NBA inference, (42b) gets the enriched truth conditions in (43a) (with the NBA enrichment in boldface), and given Flavor Matching between the NBA modal and the consequent modal, (43a) gets simplified as in (43b):

6.2 Partial harmony

In this section, we turn to a couple of potential counterexamples to our account, where there seems to be a mismatch between the flavor of the NBA modal and that of the consequent modal, which should prevent harmonizing, but where a harmonizing reading does, in fact, seem to be possible.

First, Condoravdi & Lauer (2016) present an example where an anankastic reading of a want conditional seems to be available, but where the consequent modal is taken to be deontic:

Under our account, we should only derive a harmonizing reading if an NBA inference goes through, and the flavor of the NBA modal matches that of the consequent modal. However, the only possible NBA inference for (44) seems to be teleological, not deontic. The attitude report You want to use the exemption now seems to trigger the inference that it is teleologically necessary for you to use the exemption, not that you are legally obligated to do so.

Importantly, we take NBA inferences to be just one kind of possible pragmatic enrichments, not the only one. In this particular example, we can easily assume that anyone would want to take advantage of any tax exemption that they qualify for. Given that the context presupposes that there is an exemption that can be used (given the use of the definite), it is fairly natural for the speaker to assume that if you want to and can use an exemption, then you will use it. We can thus interpret (44) with the enrichment in boldface in (45), and get an anankastic interpretation:

Another tricky example comes from a footnote in Condoravdi & Lauer (2016; 8:38), where it is used to show that on a restrictor analysis, the problems with anankastic conditionals are present even when the antecedent is about the facts that influence the contents of the modal base, and not the preferences in the ordering source. This problem is left as an open issue in Condoravdi & Lauer (2016). An adaptation of this example is given in (46), assuming a context where Al has been bragging about how much money she made this year:

In this example, we seem to get a harmonizing reading, where the complement of the attitude verb seems to restrict the modal base of the consequent modal: Al will have to pay a lot of taxes if she made a lot of money. Our account, however, only derives harmonizing readings when the flavor of the NBA modal matches the flavor of the consequent modal. The consequent modal in (46) is deontic, but while we can plausibly derive the epistemic NBA inference in (47b) from (47a), we cannot get the deontic one in (47c):

Intuitively, it seems that what is going on in (46) is that we take as fact that Al made a lot of money, and we take this fact for granted for the interpretation of the consequent modal. We tentatively propose that harmonizing can occur here, if we assume the following NBA inference from a belief report to a circumstantial (C) necessity:¹⁵

Since deontic modals take a circumstantial modal base, the modal inference would deliver that in all worlds quantified over by the modal’s modal base, Al made a lot of money, and in the legally best of these worlds, she pays a lot of taxes, as desired.

Assuming the LF in (49a), and the NBA inference in (48), we obtain the enriched truth conditions in (49b), with the enrichment in boldface. Under the hypothesis that the NBA modal and the consequent modal quantify over the same circumstantial worlds (further ordered by a deontic ordering source, in the case of the latter), we can simplify (49b) as (49c):

6.3 Embedding

We conclude this paper by briefly turning to matters of embedding. What happens when the attitude report that triggers the NBA inference, Subject ATTITUDE p, is itself embedded within the antecedent? Can we still get a harmonizing reading then? On our view, the prediction is that we can, provided that the antecedent triggers an NBA inference, from Subject ATTITUDE p to □_F p in the relevant worlds, and that F matches the flavor of the consequent modal. Since this inference is pragmatic, the question of whether it arises will depend on various background assumptions, pertinent not only to the embedded ascription, but also to the predicates that embed it. In this section, we briefly discuss some examples, and conclude that the facts seem to be as we predict them to be.

First, it appears that harmonizing readings are possible when a neg-raising attitude verb in the antecedent appears under negation. For instance, a natural interpretation of (50a) below, is the harmonizing reading that not drinking requires that you think of your family. This, we believe, is due to the NBA inference in (50b), where the negation is interpreted in the scope of the necessity modal:

Similarly, we can get a harmonizing reading for (51a) as in (51b), thanks to the NBA inference in (51c), where the negation is again interpreted in the scope of the modal:

We can also get a harmonizing reading when an attitude report is embedded under another, through a chain of NBA inferences. Thus, from (52a), we may infer (52b), if we take Sherlock to be an epistemic authority, and Watson a truthful speaker, via the NBA in (52c), assuming the NBA holds in the relevant worlds:

But while harmonizing is possible in some cases where the attitude report is deeply embedded inside the antecedent, we predict that harmonizing will fail whenever the relevant NBA fails to arise. (53), for instance, doesn’t get the reading in (52b), because the NBA inference in (52c) is not triggered.

We leave other more complex cases of embedding to future discussion.