The landscape of speech reporting*

Languages offer various ways to report what someone said. There is now a vast but heterogeneous literature on speech report constructions scattered throughout the semantics literature. We offer a bird’s eye view of the entire landscape of reporting and propose a classification along two dimensions: at-issue vs. not-atissue, and eventive vs. non-eventive. This bird’s eye perspective leads to genuinely new insights, for instance on the nature of quotative evidentials and reportative moods, viz., that they are both eventive, and hence semantically more like some types of direct and indirect speech than reportative evidentials and modals are. keywords: reported speech, evidentiality, events, not-at-issue content


Introduction
We often refer to what other people have said. And we do so for a variety of different reasons, and using a variety of different linguistic structures. This ranges from the use of free indirect discourse or direct quotation to vividly represent entire (fictional) dialogues, word for word, to the use of evidential modals or moods to hedge one's commitment to a proposition by indicating that that proposition is based on hearsay evidence. 1 1 early access b. Someone was yelling at me. What the hell did I think I was doing! free indirect discourse c. She told him that he was being paranoid. The speech reporting strategies exemplified in (1) have all received significant amounts of attention, but from different, partly overlapping corners of semantics. Although there are partial attempts at unification among these and other linguistic phenomena -e.g., analyzing free indirect discourse as direct or indirect discourse, and those in turn as attitude reports or demonstration, or analyzing German reportative moods and/or modals as reportative evidentials, and those in turn as conventional implicatures -there is little attention to the general overarching category of speech reporting. 4 In this paper we propose to zoom out and survey the entire landscape of speech reporting. Our first aim is to bring some order to the scattered body of data and theories about these various reporting strategies. We do this by proposing a classification of reports along two dimensions: at-issue vs. not-at-issue and eventive vs. non-eventive. Put in typological terms (e.g. Haspelmath 2010), we use eventivity and at-issueness as the two fundamental semantic comparative concepts and try to classify different strategies of speech reporting in terms of these concepts.
The distinction between at-issue and not-at-issue speech reports has already featured prominently in the (evidentiality) literature. It concerns the distinction between reports where the fact that something is said is at-issue versus reports where not the reportative component but the content of what was said is at-issue. Our second distinction has received much less attention, but is, as we will see, just as important to understand the differences between the various speech reporting options natural languages offer, and the ways these are used in communication. Only some types of reporting introduce a speech event into the discourse record. Only these eventive reports can be used to describe certain specific details about the speech event itself -e.g., focusing on the actual words, the manner of speaking, or the content expressed. By recognizing (and modeling) that the reporting strategies that have received most attention (indirect discourse and reportative evidentials) differ not just in at-issueness but also in this second, independent dimension, eventivity, it becomes clear that there is room for other combinations of these values, such as not-at-issue eventive reports (where we'll classify quotative evidentials and reportative moods).
In section 2 we propose a number of linguistic tests that we use to tease apart the four different combinations (from at-issue eventive to not-at-issue non-eventive).
Here too, distinguishing both dimensions (rather than just at-issueness) brings clarity as to how certain tests that have already been around in the literature relate to each other (and to some new tests). We discuss examples of report strategies exemplifying each combination of values. For instance, direct quotation, as in (1a), is at-issue eventive because, among other things, a quotation describes a specific speech event, and the existence of that speech event can be directly challenged by other participants. On the other end of the spectrum we'll classify, for instance, reportative modals, like Dutch schijnen in (1e), as not-at-issue non-eventive, because they do not introduce or presuppose a specific speech act, and their reportative meaning component, i.e. that the speaker was told Marie was pregnant in (1e), has the status of a not-atissue supplement in that it doesn't need to address the Question Under Discussion (QUD). Between these two extremes we classify for instance German and Ancient Greek reportative moods as not-at-issue eventive. The final cell, that of at-issue non-eventive speech reports, is less densely populated. We discuss why that might be and discuss English according to as a potential inhabitant. In sum, the various report phenomena exemplified in (1) inhabit the following four corners of the landscape of speech reporting: 5 (2) eventive non-eventive at-issue direct discourse (1a), free indirect discourse (1b), (canonical) indirect discourse (1c) according to (1g) not-at-issue reportative moods (1f) reportative evidentials (1d), reportative modals (1e) Our primary aim in this paper is not to introduce a completely new semantic analysis of, say, Gitksan evidential marking, nor to provide new data on, say, Dutch parenthetical indirect discourse, nor to formulate a new linguistic test for not-atissueness -although we do bring up new data here and there, and we do offer some novel modifications and syntheses of semantic theories and tests -but to survey the entire landscape of speech reporting, bringing together previously isolated areas of interest in semantics, investigating what they share, and how exactly they differ. This bird's eye perspective does lead to genuinely new insights, for instance on the nature of reportative moods, which share not-at-issueness with evidentials, but are crucially different from genuine reportative evidentials in being eventive, like direct and indirect discourse.
The paper is structured as follows: in section 2 we introduce the diagnostics for teasing apart the four different combinations, from at-issue eventive to not-at-issue non-eventive speech reports. In section 3 we present a basic formal framework that allows us to talk about eventivity and at-issueness more precisely. In sections 4 to 7 we zoom in on the four cells of the typology and discuss in more detail how various report constructions from different languages behave with respect to our diagnostics, and how best to describe their semantics.

A typology of speech reports: the diagnostic
In this section we introduce a number of diagnostics for at-issueness and for eventivity. In 2.1 we discuss three tests to tease apart at-issue vs. not-at-issue report constructions, i.e., report constructions where the information that someone said something is at-issue vs. not-at-issue. In 2.2 we discuss ways to tease apart eventive and non-eventive reports. As we will see there, we can't always cleanly separate the two dimensions. We'll propose two tests that diagnose the at-issue eventive combination. The only pure diagnostic for eventivity we found is more a heuristic, relying on preferences and/or distributional patterns. Nonetheless, our test battery will allow us to diagnose various reportative constructions according to our proposed four-part typology. 4 early access

At-issue vs. not-at-issue
There are broadly speaking two schools of thought on the nature of the at-issue vs. not-at-issue distinction. The lexical/conventional approach builds the distinction right into to the semantics of certain lexical items or constructions (Potts 2005, Murray 2017). The pragmatic approach on the other hand tries to derive it from the global structure of the surrounding discourse (Simons et al. 2010). For the purposes of this paper, we adopt a mostly lexicalist interpretation of not-at-issueness: some constructions conventionally mark some of their content as not-at-issue. This is not to say that all not-at-issueness is to be lexically or grammatically implemented. In section 6 we'll briefly touch on pragmatic sources of not-at-issueness at work in speech reports in discourse. Nonetheless, we do maintain the following dependency between the two notions of at-issueness: content lexically/grammatically marked as not-at-issue cannot be pragmatically used as at-issue in the discourse-driven sense. This principle will be relevant for the semantic implementation in section 3, but also for our discussion of certain tests in this section.
Diagnostics for establishing that something is at-issue typically involve projection out of embedding, question-answering and challengeability (see Tonhauser 2012, Murray 2010. The last two are both based on the distinctive behavior of not-at-issue content in dialogue. Question-answering is 'backward looking' in this respect (it looks at what could have been the preceding question that the given report gives an answer to), whereas challengeability is 'forward looking' (it looks at what are potential moves to follow a given report) (Koev 2018).

In Situ Interpretation Test
To start with interpretation under embedding, we can say that expressions that are syntactically embeddable and that in such environments are also interpreted 'in situ' are at-issue (going back to ideas in Potts 2005). Consider the contrast in (3) ziek. ill. 'If Piet is reportedly ill, then he is ill.' An utterance of (3a) may be perfectly felicitous. It only has a reading where volgens Jan 'according to Jan' is interpreted in situ, that is in the antecedent of the conditional (≈ 'if Jan said that Piet is ill, then he is ill'). By contrast, (3b) does not allow such 5 an in situ reading of the reportative modal schijnen (i.e., something like 'if someone said that Piet is ill, then he is ill').
Three caveats are in order here. First, note that we're not claiming that the absence of in situ interpretation entails that the reportative information is not-at-issue. Other factors may be at play to exclude syntactic embedding (e.g., morphosyntactic reasons for evidentials in Cheyenne, Murray 2010: 66). Our claim is only the other way around: what is interpreted under the scope of operators is at-issue. Thus, as for (3) we can say that the reportative contribution of volgens Jan ('according to Jan') is at-issue. If it is embeddeable but not interpreted in situ, the reportative meaning component is not-at-issue. But if it is not-embeddable the test is inconclusive since this may be for morphosyntactic or other reasons. In (3b), the source of infelicity is not clear. It doesn't 'sound ungrammatical' to us, but when reflecting about the meaning it just doesn't seem to have any sensible reading (hence '#' rather than '*').
Second, in those occurrences where the reportative content is interpreted in situ, it is at-issue. But this does not necessarily mean that it is grammatically/lexically marked as at-issue. In fact, as we will see in section 6, following Simons et al. (2010), everything can be pragmatically made not-at-issue in the right context, so, as we said above, at-issueness can never really be enforced grammatically.
Third, we'll exclude overt speech reports from the class of possible embedding environments for diagnosing at-issueness. Semantic research has shown over and over again that speech reports are special as embedding contexts in that they allow for embedded occurrences and shifted interpretations of more elements than other contexts (e.g., attitudinal particles (e.g., Döring 2013), appositives (e.g., Amaral et al. 2007), indexicals (e.g., Deal 2020), long-distance reflexives (e.g., Solberg 2017)). 6 Although the interpretation of reportative markers embedded in overt speech reports is a very interesting topic (see e.g., Schwager 2010, and also our Anaphoricity Heuristic below), it cannot be used to tease apart at-issue and notat-issue contributions. For this reason we'll stick with embedding in conditional antecedents for this test.
To summarize: In Situ Interpretation Test: If a reporting construction is syntactically embeddable in the antecedent of a conditional and there interpreted in situ, the reportative meaning component it introduces is at-issue. If it is syntactically embeddedable but not interpreted in situ, the reportative meaning component is not-at-issue.
6 See also Krifka 2014 for an account of the special status of speech reports as embedding environments. 6 early access

What Makes You Think Test
Simons (2007) uses question-answering to bring out the 'main point status' of the reported proposition in some report constructions. The idea behind this diagnostic is that in a felicitous discourse the Question under Discussion (QUD) must be addressed by a subsequent discourse move, but the answer cannot be provided by a proposition that is lexically or grammatically marked as not-at-issue (Simons et al. 2010). In other words, B's answer in (4) is felicitous because its main clause, the at-issue contribution, that Mary is ill, gives a partial answer to the question why Mary and Sue aren't here. Answer B' is infelicitous because the information that would answer the question is conveyed by an appositive relative clause, while the main clause content doesn't answer the question.
(4) A: Why are Mary and Sue not here? B: Mary, who likes to come to these meetings, is ill. B': #Mary, who is ill, likes to come to these meetings.
We turn this reasoning into a test to establish whether a component is conventionally marked as not-at-issue. Specifically, we determine the (not-)at-issueness of the reportative proposition by looking at a context where speaker A asks B explicitly about her evidence, thereby making this the QUD. In such contexts the answer in (5B) is infelicitous, even though it entails that someone said that Mary is ill, which would in principle answer the QUD. We take this to show that the reportative component is necessarily not-at-issue. By contrast, the answer in (5B') is felicitous, showing that here the reportative component is at-issue.

(5)
A: What makes you think that Mary is ill? B. #Ze she schijnt seems ziek ill te to zijn. be 'She is ill, reportedly' Dutch B'. Someone said/told me that Mary is ill.
The question in (5) probes for at-issue speech reports, irrespective of whether they are eventive or not. In the literature other questions have been used as well to test at-issueness in a backward looking way (that is, by looking at the kind of questions a construction can answer, or, more generally, by looking at the kind of moves it can follow). In 2.2 we argue that some of these questions ask for answers whose reportative component is both at-issue and eventive and thus do not test for at-issueness per se.
In sum: What Makes You Think Test: If a sentence of the form '[reportative expression] + p' is infelicitous in answering a question of the form What makes you think that p? , the reportative meaning component is lexically/grammatically marked as not-at-issue. Otherwise, it is not lexically/grammatically marked as not-at-issue.

Challengeability Test
Challengeability looks at what are potential next moves, more specifically whether or not the reportative component can be challenged by another discourse participant. 7 In direct and indirect discourse this component is easily challenged. In fact, when responding to a canonical report we can in principle challenge either that such a saying event took place, as in B, or we can challenge the embedded proposition, B': (6) A: She said {"I'm innocent"/ that she was innocent} B: Nonsense, she may be innocent, but she would never say that. B': Nonsense, she's guilty, regardless of what she told you.
By contrast, for reportative evidentials a denial targeting only the reportative proposition is impossible, or at least much more difficult. This is illustrated for the Gitksan reportative gat in (7) Korotkova (2016) argues that the infelicity of denying the reportative component of an evidential report is the result of the subjectivity of the evidence -you cannot usually deny that I heard something because I'm in better epistemic position for such facts. Moreover, Snider (2018) notes that some reporting components would be classified as not-at-issue on the basis of a backward looking test, but as at-issue on the basis of challengeability. Recognizing that challengeability alone may be inconclusive, we choose nonetheless to include it in our test battery for at-issueness. 8 The consultants were first asked to translate the individual English sentences into Gitksan and were then given the dialogues in (7) and asked whether these were ok. As for B', where the evidential proposition is targeted, a consultant remarked "It's fine but you're just contradicting yourself." Abbreviations: DM = determinate marker, DUR = durative, FOC = focus, PN = proper noun connective. 8 early access 'That's not true. You saw John fishing yesterday.' 9 Challengeability Test: If the reportative component of a sentence cannot straightforwardly be challenged without thereby also challenging the reported content, the reportative component is not-at-issue. Otherwise it is at-issue.

Modification Test
Eventivity is characterized by reference to a speech event. We can linguistically probe for the presence of a speech event in the semantics by trying to manipulate it. A first way to do this is by modification.
(8) Two days ago Juan whispered to Toshi, in his deep voice, {that it was over/"It's over"}.
Dutch schijnen and the reportative modal zou do not allow such modification: Although the sentence in (9b) does entail that someone said something, what is located yesterday in (9b) is not a speech event but the (alleged) illness. For (9a) the lack of modifiable speech event results in infelicity due to the fact that in a deep voice cannot plausibly be taken to modify the state of a relationship being over.
Note that this test only goes one way: if we can modify the manner of speaking, or specify a concrete time and place, we must be dealing with an eventive report. But if we can't, that doesn't necessarily mean we're dealing with a non-eventive construction. Independently of eventivity, not-at-issue content is also often considered immune to modification as in (9), so if we find that something can't be modified, this can be blamed on either not-at-issueness or non-eventivity.
Modification Test: If we can modify the manner of speaking, or specify a concrete time and place for the reported speech act, we are dealing with an eventive speech report. Otherwise, the report may be non-eventive or not-atissue.

What Happened Next Test
Next, note that only at-issue eventives can be used in answer to questions that focus on or ask for an event, such as Did anyone dare say anything? and What happened next?. By asking for a (speech) event and letting the answer be the report of a state we can separate at-issue eventive speech reports (yielding felicitous question-answer pairs) from other kinds of speech reports (yielding infelicitous question-answer pairs).

(10)
A: Did anyone dare say anything? / What happened next? B: John said, "It's over". B: #Allegedly, it was over.

Anaphoricity Heuristic
The tests discussed so far divide the total landscape of speech reports into three classes: at-issue eventives, at-issue non-eventives, and not-at-issue reports. Within the not-at-issue class, however, there are also reasons to distinguish between eventives and non-eventives, as we can see when we consider their natural environments. Let's consider the optative mood in Ancient Greek. This mood usually occurs in complement clauses after verbs of saying. In grammars it is indeed described under the heading of indirect discourse, as a special mood occurring therein (e.g., The more elaborate grammars go on to note that the optative also has reportative uses in continuations of indirect discourse, as in (12) (this is still only the beginning, the report continues for a couple of sentences), where the optative-marked verbs are underlined in the translation: 'After this the opportunity to speak was offered to any one who desired it; many spoke to the same effect, saying that Seuthes said things of supreme importance; for the season was winter, and it was not possible to sail back home, if that was what one wished, and impossible also to get along in a friendly country if they had to maintain themselves by purchasing . . . ' Ancient Greek, X. An. 7.3.13 As argued in Bary & Maier 2014, Bary 2018, optatives in such continuations are not syntactically dependent on the verb of saying that introduces the speech report. The presence of the particle gar 'for' indicates that there a new main clause starts.
Out of the blue occurrences of the reportative use of the optative are very rare: Only five examples are known, all from Herodotus (Bary 2018). The fact that the most natural habitat of the Greek optative is the complement clause of an indirect speech report, in combination with the fact that, when not syntactically embedded, out of the blue occurrences are much rarer than occurrences where the reported content is interpreted as further specification of a speech event previously introduced, 11 points in the direction of an eventive character.
Note that a 'continued report' reading is not sufficient for a report to qualify as eventive. Faller (2002) and Murray (2009) already give examples for Cuzco Quecha and Cheyenne respectively in which the content of a sentence with a reportative marker is to be interpreted as part of the content of a speech event referred to earlier.
(13) is an example from Cheyenne: Annie. Annie 'Yesterday I spoke to Dale. [He says that] Annie won.' Cheyenne, Murray 2009: 338 The difference is that while for the Greek (eventive) ones such an anaphoric link is the rule, the reportative markers in Cheyenne and Cuzco Quechua occur happily out of the blue, without picking up a particular speaker for example. We suggest that the interpretation of the sequence of two sentences in (13) is due to a more general and looser pragmatic requirement to establish discourse relations between utterances that succeed each other in a discourse.
Since anaphoric readings are evidently not strictly sufficient for eventivity, we formulate this 'test' as a heuristic.
Anaphoricity Heuristic: If a reportative expression tends to occur in discourse contexts where the reported speech situation has already been explicitly introduced before, we may assume the report is eventive.

A formal framework for representing events and not-at-issue content
Throughout the paper we'll make our proposal explicit by offering logical forms representing the compositional contributions of the relevant constructions. We'll rely on a traditional, two-stage setup, where natural language sentences are translated into formulas, which in turn can be interpreted in a model. The resulting interpretation can be used to update a context. For the formal language we use a standard higher-order, intensional, typed lambda calculus, with indexicals. Basic types are e (entities), r (times), v (eventualities), s (possible worlds), and t (truth values). We'll use explicit quantification over individuals, eventualities, and times, but not possible worlds -for intensional constructions we use ∧ ϕ, a term of type st to denote the possible worlds proposition expressed by ϕ.
We'll denote the translation of a natural language expression into formal meta-language with or T. For example, T (happy) = happy et . Subscripts indicate the type of an expression, but are typically omitted. Expressions in the formal language are model-theoretically interpreted relative to a context c (an agent-world-time triple), an index w (world-time pair), and an assignment function f : E.g., happy(i) c, f w = 1 (i being a constant denoting the agent parameter of c) iff i c, f w ∈ happy c, f w iff agent(c) ∈ {x x is happy in w}. From here on we'll ignore the assignment function f whenever there are no variables to interpret.
Note that happy(i) is of type t and thus refers to a truth value when interpreted in a model, relative to a context and an index. Given this Montague/Kaplan framework we can now define the proposition expressed by an utterance of a sentence (a term of type t) in context c by abstracting over the index: happy(i) c = {w ∈ W happy(i) c w }. To avoid confusion later on it may be worth stressing here that terms of type st by contrast refer to propositions, relative to a context and index: ∧ happy(i) c w = λ w. happy(i) c w (= the function that maps worlds where the first-person agent of c is happy to 1 and other worlds to 0). Such st terms don't correspond to utterances or assertions as they are not 'truth-evaluable' (relative to context and index). Put differently, there is no such thing as the proposition expressed by an utterance of a type st term in a context c. Our analysis in section 5.3 and beyond will rely on this fundamental distinction between type t terms (corresponding to assertions) and type st terms (corresponding to sentential complements in intensional constructions).

Contentful events
In Neo-Davidsonian event semantics, clauses and adverbial modifiers alike express properties of events, which can be combined via the familiar Predicate Modification rule (Heim & Kratzer 1998): If T (α) = λ xϕ and T (β ) = λ yψ (and x doesn't occur free in ψ), then For example: In the Neo-Davidsonian paradigm the standard arguments of the verb (agent, experiencer, theme, goal, etc) are introduced via (covert) thematic role operators (which we'll sometimes leave out of our syntactic LF representations): At the end of the compositional translation of a clause we will now usually end up with a property of events, as in (15) (and (16)). We then automatically apply Existential Closure to turn that event property into a truth-evaluable event quantification (of type t): For our purposes, an important application of event semantics involves the representation of attitude and speech reports. Kratzer (2006) argues against the traditional Hintikka-style approach to attitude verbs as intensional operators, in favor of an approach where the attitude verb introduces an eventuality and the complement serves as a modifier. Simplifying her approach, in line with subsequent applications (e.g., Brasoveanu & Farkas 2007, Hacquard 2010, Maier 2017, we'll assume that some eventualities (belief states, speech events, but not walking or touching events) themselves have propositional contents. We posit the operator CONTENT, relating an event and its propositional content: 10 The syntactic LF of an attitude or speech report construction then looks like (19) (where we may think of the optional complementizer that as the overt realization of CONTENT):

(19) [ EXPERIENCER Mary ] thinks [ CONTENT [ I'm crazy ]]
Translating this into the proper logical form is straightforward. Following Heim & Kratzer's (1998) Intensional Functional Application rule, the clause I'm crazy, of type t (20a), gets lifted to the propositional type st in order to fit the argument slot of CONTENT, yielding an event property (20b) that modifies the verb (20c) via Predicate Modification (20d). The rest is as before: the subject term is analyzed as another event property, further modifying the thought event, and finally we apply Existential Closure to get the actual truth conditions in (20e): In words, the report expresses that Mary was the agent of some thought event that had as its content the proposition that I'm crazy. In the following sections we will adopt this general event-based approach to model various forms of direct and indirect speech reports, and later reportative moods.

At-issue vs. not-at-issue content
The distinction between at-issue and not-at-issue content is cashed out in terms of multiple meaning dimensions (Bach 1999, Potts 2005, Geurts & Maier 2013. 11 (21) Mary, who likes to come to these meetings, is ill.

≈
Mary is ill Mary likes to come to these meetings Potts uses these two-dimensional forms to model the projection behavior of what he calls conventional implicatures, i.e., the supplemental content on the second dimension. The idea is that linguistic operators (negation, quantifiers, modals etc.) are semantically redefined in such a way as to only operate on the first dimension, leaving the second dimension unaffected. This familiar setup indeed semantically derives universal projection of any meaning components that are conventionally marked as not-at-issue. The next step is to describe how these two meaning dimensions derived for a whole sentence interact with the surrounding discourse structure and the common ground when uttered in a specific discourse context. We want to follow a standard strategy for interpreting appositives and evidential meaning dimensions generally, viz. that they impose a non-negotiable, forced "pre-update" of the common ground (Murray 2009, Koev 2013, AnderBois et al. 2015, Griffiths 2015. Based on ideas from Farkas & Bruce (2009) and Inquisitive Semantics (Groenendijk & Roelofsen 2009), and going back to Stalnaker (2002), at-issue content is analyzed as an "update proposal" that addresses the question under discussion and can be accepted or rejected by other discourse participants. To make this formally precise we need a structured discourse model that distinguishes updates from proposals and further keeps track of questions under discussion.
Simplifying somewhat the familiar structured discourse model of Farkas & Bruce (2009) we take a discourse context to be a quadruple c,C, Q, P , where c is a Kaplanian context (specifying the current speaker s c and the time t c and world w c of her utterance), C is a set of worlds or indices (the context set, representing what is considered common ground), Q a stack of QUDs (formulas representing the open questions under discussion), and P a stack of proposals (formulas representing the update proposals currently being negotiated for inclusion in the common ground). 12 With Farkas and Bruce we assume some basic stack operations: push(X, x) (push an element x on top of stack X) and pop(X) (remove the top element from the stack), and we use top(X) to refer to the top element of the stack. In our discourse model, asking a question amounts to pushing (the logical form of) the question onto the QUD stack, (22a), and asserting amounts to pushing the logical form of the assertion onto the proposal stack, (22b). We assume that a discourse move of asserting is only allowed if it addresses the question at the top of the QUD stack (i.e., it contextually entails a full or partial answer to the top question, see e.g., Simons et al. 2010 for details).
Whenever the proposal stack is non-empty, the further discourse moves of accepting and rejecting become available. 13 Acceptance is a crucial move to meet the ultimate goal of cooperative conversation, as formulated by Stalnaker (1970): increase the amount of information in the interlocutors' common ground, i.e., reduce the context set. Rejection, as defined in (23b), does not affect the common ground, nor does it settle any QUD, it just means that we replace the top of the proposal stack with its negation. After a rejection, the discourse may develop into different directions. We don't want to go into the fine details of dialogue management here. We also don't have anything new to offer with regard to the question of how the abstract conversational moves of ACCEPT and 12 The Farkas and Bruce model keeps track of the proposals made by each of the participants separately (in the form of the individual, public discourse commitments taken on by each speaker). We simplify by just keeping track of what is being proposed. 13 By popping the question stack we are assuming that a question that has been addressed counts as answered and can be moved off the table. This is of course an oversimplification of the discourse dynamics of complex hierarchies of QUD's and partial answers proposed by various participants. See e.g., Roberts 2012. 14 Recall from our remarks about the Montague/Kaplan system in section 3 that if the current proposal top(P) is of type t, top(P) c is the possible worlds proposition (more precisely: a set of indices) expressed by an utterance of that proposal in context c.
REJECT are realized linguistically, or paralinguistically by speakers. We'll make the simplifying assumption that ACCEPT is a default move, i.e., if nobody objects at the end of a turn, ACCEPT is automatically applied.
Our main concern here is to use this general discourse framework to capture the interpretation of a 2D logical form. Implementing the not-at-issue-as-pre-update and at-issue-as-proposal ideas outlined above means updating the context set with the not-at-issue component, and adding the at-issue component on top of the stack of proposals.
To illustrate our basic 2D semantics and pragmatics, let's apply it to the appositive example in (21), repeated in (25a) with 2D logical form in (25b).
(25) a. Mary, who likes to come to these meetings, is ill.
b. ill(mary) like(mary, meetings) Let's assume (25a) is uttered in a context c,C, Q, P , where the top QUD in Q is the question why Mary is not at the meeting. Let's further assume that, in this context, this question would be fully answered by the proposition that Mary is ill. This means we can apply the update schema from (24) to derive the following update: (26) c,C, Q, P + ill(mary) like(mary, meetings) = c,C ∩ like(mary, meetings) c , Q, push(P, ill(mary)) Since the proposition that Mary is ill is now the top update proposal, other participants may challenge it, with REJECT . If nobody challenges the proposal, or it's actively endorsed, as the asserter's turn ends we apply our default ACCEPT as defined in (23a), in order to achieve a further common ground update.
The derivation captures the information-structural intuitions typically associated with appositives: the at-issue component, in this case the host clause contribution, has to address the top QUD; if it does, the not-at-issue component updates the common ground, and the at-issue component is up for discussion, only turning into an additional common ground update if unchallenged (or explicitly endorsed by the other participants).

early access
This concludes our presentation of a general framework for representing contentful events and not-at-issueness. In the remainder of the paper we plan to apply this framework to capture the semantics of the four types of speech reports from our typology in such a way as to properly predict the various test behaviors identified in the previous section.

At-issue eventive reports: direct and indirect speech
In section 2 we proposed a number of tests to diagnose whether a linguistic report construction is eventive, and whether the report component is at-issue. On the basis of these tests, canonical direct, indirect, and free indirect discourse come out as at-issue eventive reports. In this section we propose a uniform neo-Davidsonian semantics for these report types and show that it captures both the eventivity and the at-issueness we observe in the tests.

A uniform semantics for direct and indirect speech
Our semantics of canonical report constructions starts from the idea, introduced in section 3.1, that speech report verbs, like attitude verbs, introduce events that can have a propositional content. One difference between speech and attitudes is that speech events not only have contents but also concrete linguistic forms. Following Maier (2017) we use this to provide a uniform semantics of direct and indirect speech reports: while an indirect speech report complement specifies the content of a speech event, a direct report specifies its form. The event-based approach has also proven particularly well-suited for modeling less canonical report types that we'd classify as eventive, such as free indirect discourse which has been argued to be some kind of mixture of form and content (Maier 2017). 15 Working this out compositionally, we need first of all a way to refer to a linguistic form (and/or demonstration, depending on your basic theory of quotation). In our formal language we introduce a new type u (à la Potts 2007) and use Quine hooks to indicate (pure) quotation in the formal type-theoretic metalanguage, i.e.: If σ is a string of letters (or signs, or phonemes), then σ is of type u and σ = σ .
Strings of letters are now part of our ontology. We assume that each natural language expression corresponds to such a string and that at any point in the semantic derivation we have access to that string. We then introduce an operator QUOT as the quotational analogue of CONTENT, i.e., relating an event and its form (given as an expression of type u). In words: the direct speech report in (31) says that, within the contextually relevant conversation that Mary had, there was an event e, which was a saying event with Mary as agent, and with a phonological surface form approximated by the sequence of letters You're crazy.

Evaluating the event semantics for direct and indirect speech reports
Let's take a closer look at the various diagnostic tests and see how our semantics predicts the at-issue eventive behaviors of canonical direct and indirect speech reports.

Eventivity characteristics
In diagnosing eventivity, the first test we discussed involved modification of the speech event as a sign of eventivity. Direct and indirect speech reports clearly allow such modification, as shown in example (8), repeated below in (32).
Two days ago Juan whispered to Toshi, in his deep voice, that it was over.

19
The current semantics correctly predicts this behavior. We analyze the speech verb as an event property which can be further modified through eventive Predicate Modification: Next is the What Happened Next Test. Our semantic analysis of canonical reports, combined with our general pragmatic account of structured context updates, correctly predicts that such reports should be felicitous as answers to questions like What happened next? or Did anyone dare say anything?.
(34) A. What happened next? B. Sue told May that she's a genius.
On our discourse model (more specifically see definition (22)), in interpreting this dialogue A's question gets pushed on top of the QUD stack. The report is treated here as a simple unidimensional assertion, which, still following (22), can be felicitously used to update a context quadruple if it addresses that top QUD. Since the occurrence of a certain speech event fully answers the question what happened, the dialogue is correctly predicted to be pragmatically felicitous.

At-issueness characteristics
In section 2.1 we introduced three tests to detect at-issueness. We'll go through those now. First, direct and indirect speech obey In Situ Interpretation. In (35), the existence of a saying event doesn't project, but is interpreted as the antecedent of the conditional: If Mary said {that I'm crazy/"You're crazy"}, then maybe I am crazy.
Our semantics predicts this lack of projection from clausal embedding, because the existential event quantification is introduced by Existential Closure, which is triggered at the level of the smallest clause containing the verb. We thus predict the following logical form: (36) (∃e[agent(e) = mary ∧ say(e) ∧ content(e) = ∧ crazy(i)]) → crazy(i) 20 Second, we've already observed in (5) that indirect speech can be used to answer a What Makes You Think question: (37) A: What makes you think that Mary is ill? B: John told me that she is (ill).
B's answer is perfectly felicitous, which means that it must address the top QUD, i.e. A's question. On our semantics, B's answer contributes a certain speech event (with the content that Mary is ill), which directly addresses A's question. We thus correctly predict that the dialogue is felicitous. Thirdly, we have seen in (6), repeated in (38), that (in)direct speech comes out as at-issue in the Challengeability Test: we can deny just the existence of the speech act, without thereby denying the reported content.
A. She said that she was innocent. B. Nonsense, she may be innocent, but she would never say that. This is as predicted by our event semantics and discourse model. With a simple, one-dimensional speech report, A proposes to make it common ground that some salient 'she' said something, i.e. the existence of such a speech act gets put on top of the current proposal stack, and can then be either accepted or rejected. B's 'Nonsense' indicates that this proposal is in fact being rejected, meaning it gets taken off the table, to be replaced by its negation.
5 Not-at-issue non-eventive reports: reportative evidentials Reportative evidentials are constructions that mark some content as based on hearsay evidence. In this section we propose a semantics that treats this type of evidential as a not-at-issue non-eventive speech reporting construction. This semantics extends beyond reportative evidential morphemes in the strict sense of Aikhenvald (2004) to include also uses of lexical and other non-functional markers with a reportative meaning component which according to our diagnostics fall in the same corner of the speech report landscape. This includes some English adverbials and Dutch and German modals.

A 2D semantics for reportative evidentials and modals
Based on some of our at-issueness tests, we've already suggested in section 2 that the reportative component of constructions like in (38c) and (39) is conventionally marked as not-at-issue. The reported content on the other hand is at-issue. Hence, the starting point of our semantic analysis is to put the reported content, often referred to as the scope proposition in the evidentiality literature, on the first dimension of a 2D logical form and the reportative component on the second (see McCready 2008 for discussion of such a 2D analysis of evidentiality): Allegedly, Mary is ill.

Mary is ill I've been told that Mary is ill
To make this precise and compositional, we propose the following lexical interpretation of allegedly and other reportative markers (Gitksan gat, Dutch schijnen): The predicate HEARSAY in the second dimension is a traditional Hintikka-style propositional attitude operator, indicating that the actual speaker has hearsay evidence for a proposition from some source: 16 (42) p ∈ HEARSAY c w iff for all w compatible with the information the agent of c heard in w, w ∈ p.
Our proposal to use a Hintikka-style intensional operator does not entail that events play no role in a deeper conceptual analysis of evidentiality. One might well want to further analyze HEARSAY in terms of the existence of an event of someone hearing that p, or of some speaker saying that p (Ozyildiz et al. 2018). One should then be careful to ensure that such an event-based decomposition does not lead to unwanted speech event modifications, what-happened-next-answers, and especially anaphoric binding tendencies. We will not pursue this route here but opt for a traditional Hintikka analysis as the most economical analysis to get all the observed test behaviors right.
In the following we illustrate our semantics by focusing on (clause-initial) allegedly as a concrete example. Just as in eventive indirect discourse, we use Intensional Function Application to fit the sentential type t argument into a type st argument slot (by adding ∧ ). (43) T (Allegedly, Mary is ill) = ∧ ill(mary) HEARSAY( ∧ ill(mary)) We deliberately do not 'extensionalize' (with ∨ ) this proposition to get a regular assertive sentence contribution (of type t) in the first dimension of (43). Instead, in section 5.3 we will slightly modify our pragmatic treatment of 2D logical forms from section 3.2 in order to turn these intensional 2D logical forms to our advantage in dealing with a phenomenon known in the evidentiality literature as reportative exceptionality. But first, in the next subsection, we evaluate the compositional, syntax-semantics part of our proposal thus far.

At-issueness characteristics
The Pottsian derivation of 2D logical forms for complex sentences forces projection of the reportative meaning component, which correctly predicts the not-at-issue behavior diagnosed by the In Situ Interpretation Test of section 2. Although many of the evidential-type constructions mentioned above are syntactically restricted to main clauses, so the test doesn't apply, the constructions that can be embedded do show projection, or rather, they don't show in situ interpretation. In (3) we saw an example with the Dutch reportative modal schijnen embedded in the antecedent of a conditional. In (44) we see a similar example with the reportative marker ku7 from St'át'imcets that does not allow the sensible reading you'd get with an in situ interpretation of the reportative ('if you hear that Sonja arrives, tell me'). 17 (44) *lh-t'íq-as HYP-arrive-3CONJ ku7 REPORT k DET Sonja, Sonja sqwal'-en-ts-kál'ap tell-DIR-1SG.OBJ-2PL.SUBJ 'If [reportedly] Sonja arrives, tell me.' St'át'imcets, Matthewson et al. (2007) We confirmed the not-at-issue status of Dutch schijnen with the What Makes You Think Test in section 2. We can see the same pattern for the reportative enclitic gat in Gitksan, an evidential language in the strict sense of Aikhenvald (2004). 18 17 Note that the semantics we proposed for the hearsay operator in (42)  The fact that (45B) does not provide a felicitous answer to (45A) shows that the reportative proposition is not-at-issue. 19 This pattern is predicted by our 2D formal analysis in combination with our dialogue model. In a dialogue like (45), A's question gets added to the QUD stack, and then B's assertion gets analyzed as a 2D logical form, of which the second dimension, the reportative proposition, updates the context set, while the first, the scope proposition, gets put on top of the proposal stack, but only on the condition that it addresses the top QUD. In other words, the scope proposition has to address A's question on pains of infelicity. Since A's question in (45) is not about the scope proposition p but about the speaker's evidence for p, we correctly predict infelicity.
The Challengeability Test, finally, also shows not-at-issueness. In section 2 we showed the relevant data for Gitksan in (7). Similar observations have been made about, for example, Cuzco Quechua (Faller 2014), Cheyenne (Murray 2009), and Turkish and Bulgarian (Korotkova 2016). The pattern extends to the Dutch reportative schijnen, where an interlocutor may challenge the reported proposition, as in B's objection in (46) Our semantics and pragmatics predicts this behavior: A's assertion gets analyzed as a 2D logical form, so the second dimension, the reportative proposition, pre-updates the common ground, while the scope proposition (that she's sick in (46)) gets put on the proposal stack for discussion. B's answer indicates a rejection of the proposal, taking the proposition that Anne is ill from the table. The pre-update, with the proposition that someone told A about Anne's illness, survives B's rejection, which is what we observe in (46). The infelicity of the B' answer is accounted for because a rejection can only target what's on the table.

Eventivity characteristics
We now turn, more briefly, to the eventivity tests. First, in evidential reports the manner of speaking or hearing can't be modified abverbially, as witnessed by (9) where the Dutch translations of 'in a deep voice' and 'yesterday' couldn't modify the reportative components introduced by reportative schijnt and zou, respectively. This follows directly from our decision to use a Hintikka-style hearsay operator rather than a Davidsonian semantics with contentful events for evidentials.
In (10) we also saw that evidential reports with allegedly combined with a stative scope proposition are indeed unable to answer 'eventive questions' like What happened next?. This also follows directly from the absence of a speech event in the logical form: The fact that p holds in all worlds compatible with what I heard is a fact, or at best a state, but not an event that could count as an answer to What happened?. If, as in the example in (10), the scope proposition p itself is also stative, we correctly predict the answer to be infelicitous.
As for the Anaphoricity Heuristic, finally, note that the evidential report constructions discussed here occur happily out of the blue, that is, without the earlier introduction of a speech event or a contextually salient conversation to connect to. Example (47), where the question could be a discourse-initial move, shows this point: The Hintikka-style hearsay operator in our semantics indeed does not come with any anaphoric dependencies or presuppositions with respect to earlier conversations. We return to this point when we contrast it with reportative moods in section 6.3.

The pragmatics of reportative evidentials: Hedged commitments and reportative exceptionality
Non-reportative evidential markers such as direct or inferential ones entail that the speaker is in some sense committed to the scope proposition, as witness the fact that she can't follow the evidential with a denial of it's scope content (see e.g.,  Matthewson et al. 2007: ex. 28 Now compare that with the canonical, eventive report constructions of section 4, where the speaker expresses no commitment to the reported content whatsoever: (50) John said {"I didn't do it"/that he didn't do it}, but he clearly did.
When I report John's words in this way I clearly do not implicate my own belief in or acceptance of their truthfulness. The vast majority of instances of what we are calling not-at-issue non-eventive speech reports, including most so-called reportative evidentials and reportative modals, seem to fall somewhere in between on the scope commitment scale:  (51) is not primarily interested in reporting the words of some film critic, rather, they're telling you that it's probably a good movie, and using the reportative marker mainly to add some not-at-issue evidential information, viz., that they don't have this from personal experience but from hearsay. 26 At the same time, adding the reportative does tend to signal a weakening of the speaker's commitment to the truth value of the scope proposition and allows her to maintain plausible deniability. This weakening effect can sometimes be brought out with a follow-up where the reporter explicitly further distances herself from the scope proposition: of 'Anneloes is at home, I am told, but I don't believe it.' Dutch, Koring 2013 This weakening of 'scope commitment' in reportative evidential constructions has been dubbed 'reportative exceptionality' in the evidentiality literature (cf. AnderBois 2014 for a detailed overview), referring to the fact that this pattern is not found in any other types of evidentiality marking. A crucial puzzle in the evidentiality literature thus consists in reconciling this behavior with the information structuring diagnosed above: If the at-issue proposition contributed by a report is the scope proposition, how come the speaker is not committed to that?
Existing accounts differ on whether or not speaker commitment to the scope proposition is built into the semantics. On the one side, we have AnderBois (2014), who, from a crosslinguistic perspective, argues that scope commitment is part of the semantics of evidentials, analyzing exceptions in terms of a pragmatic perspective shift of the sort proposed by Harris & Potts (2010). On the opposite side, Faller (2002) argues for Cuzco Quechua that the reported content is not proposed to be added to the common ground at all: it is merely 'presented' rather than asserted. Murray (2014Murray ( , 2017 offers a hybrid account for Cheyenne, distinguishing between what is on the table for discussion (viz., the plain scope proposition), and what eventually updates the common ground (viz., a suitably modalized version thereof).
Synthesizing these three ideas, we suggest the following picture: With a not-atissue non-eventive construction, the speaker herself initially only presents, as Faller puts it, a main content, which we already model by generating a proposition of type st rather than t on the first dimension of the logical form (see (43)), and hence on the proposal stack in our dialogue model from section 3.2. The idea there was that the top proposal is what's on the table for discussion, leading either to rejection or acceptance. But what we accept (or reject) cannot, for logical type reasons, be the term of type st that we have as our at-issue component. We have to first bring it down to something 'assertable', something that can be judged true or false and that can hence be judged as addressing the QUD or not -in other words, something of type t. Inserting a default ∨ operator to 'fix the type mismatch' would entail full commitment to the scope proposition. As we just saw in (52), that would be too strong for our not-at-issue non-eventive reporting constructions. 20 We suggest that what we accept or reject in case of a 'propositional proposal' (type st) is a modalized version of the scope proposition. Using Kratzer's (1981) standard framework for representing modality, we take the relevant modal to be an epistemic necessity modal, formally represented as MUST, an operator of type (st)t. We assume that the context c supplies an epistemic modal base f c for this modal, giving us the set of worlds epistemically accessible from w ( f c (w)). 21 In our dialogue model we now extend the definition of ACCEPT from (23a)  Let c,C, Q, P be a context in the sense of our dialogue model where top(Q) is the 20 St'át'imcets ku7 from (49) is a notable exception here. To deal with this datum we propose to follow a suggestion from Kratzer (2012: chapter 2) (in turn based on the data and analysis of Rullmann et al. 2008) to modify the lexical reportative contribution of ku7. Instead of introducing a 'weak' hearsay modal on the second dimension of the logical form, this marker lexically introduces a 'strong' modal, presupposing a circumstantial modal base (≈ given these rumors I heard, ϕ must be true). 21 f c is not to be confused with the (here suppressed) assignment f , and note that ϕ in (53) is of type st. 22 Note that in our 2D update definition in (24) we also have a felicity condition that says that a proposal must address the top QUD. For a formal definition of the 'address' relation we referred to Simons et al. 2010. We must now extend that definition to cover terms of type st rather than just plain assertions of type t. Here's a simple proposal: a term ψ of type st addresses * a question ϕ? iff ∨ ψ (i.e. the assertion that ψ is true) addresses the question. 28 early access question why Mary is not at the meeting. The first step in the update process is as before: (56) c,C, Q, P + ∧ ill(mary) HEARSAY( ∧ ill(mary)) = c,C ∩ HEARSAY( ∧ ill(mary)) , Q, push(P, ∧ ill(mary)) When we apply ACCEPT we effectively bring down the intensional type of the proposal by means of the MUST operator so that we can use it to update the common ground: (57) ACCEPT ( c,C ∩ HEARSAY( ∧ ill(mary)) , Q, push(P, ∧ ill(mary)) ) = c,C ∩ HEARSAY( ∧ ill(mary)) ∩ MUST( ∧ ill(mary)) , Q, P We've thus updated the context set with the information that the speaker has heard that Mary is ill, presented the proposition that she is ill as a proposal, and eventually, after acceptance by the discourse participants, further updated the context set with the information that, for all the speaker knows, Mary must be ill (where the speaker's epistemic modal base for this last modal hedging is determined by the context c). (52), like Faller's Cuzco Quechua example, translated below (with reportedly representing the reportative evidential):

Now let's return briefly to the explicit disavowal examples in
(58) They left me money, reportedly. But no, they didn't leave me one Sol.
Following the procedures outlined above we predict -assuming that a default acceptance occurs after both sentences of (58) -three subsequent updates to the common ground: the speaker heard that they left her money, for all she knows they left her money, and they didn't leave her any money. The latter two come close to a Moore paradox. After proposing that she believes they left her money, how can she go on to propose that they didn't? Our Dutch example from (52) brings out the apparent inconsistency even more clearly (again, translating the reportative evidential with reportedly): (59) Anneloes is at home, reportedly, but I don't believe it.
Here the speaker proposes that Anneloes must (in the epistemic sense) be at home, and then proposes she doesn't believe Anneloes is at home.
The key to avoid paradox and inconsistency in these cases lies in the contextdependence of the modal base for our Kratzerian epistemic modal MUST. In (59) this modal picks out a different epistemic modal base from the explicit believe verb in the continuation. Perhaps the most plausible reading here ties the implicit modal to hearsay-based beliefs: for all I knew from hearsay sources, she's at home, but for all I really believe, upon careful reflection, pulling data from any and all sources available to me now, she is not. In such a reading the implicit MUST picks out a rather specific subset of epistemically accessible worlds, rendering the accepted modalized assertion very weak, perhaps even as weak as the pre-update with HEARSAY that is already added to the common ground. The explicit 'believe' verb then picks out a different subset of the agent's epistemically accessible worlds, rendering this apparent disavowal non-contradictory. To account for unmodalized disavowals like (58) we'd have to first spell out the Moore-like paradox more precisely, appealing to some epistemic norm of assertion, and then allowing that norm to pick out a different modal base than the MUST operator. We leave the details of this interface of our semantics with speech act theory to a future occasion.
Finally, note that the idea of picking two distinct epistemic bases in (59) is corroborated by observations from Koring (2013) and AnderBois (2014) that in cases like these we tend to find particles and the like to bring out some kind of contrast, to the point where a plain disavowal would be slightly degraded: 6 Not-at-issue eventive reports: Parentheticals, moods, and quotatives We have covered the two extremes of our classification: at-issue eventive and not-atissue non-eventive. These cover the most discussed classes of report constructions, viz., (in)direct discourse and reportative evidentials. We've proposed a semantics built around contentful speech events for the first, and a two-dimensional semantics with an intensional HEARSAY operator for the second. In the latter, we started by distinguishing a reportative component from the reported content, and treating the first as a not-at-issue supplement. Eventive reports can be decomposed into the same two propositional components: someone said that p, and p itself. In canonical direct and indirect discourse in narrative contexts, the reportative component tends to be 30 at-issue. But that need not always be the case, and in fact in this section we review various ways in which languages mark an arguably eventive reportative component as not-at-issue.

Parenthetical reports
First, consider the information structural effects of parentheticalization. In general, as we already saw in section 3.2, parentheticalization marks something as not-at-issue (Potts 2005). So we expect that the familiar parenthetical realizations of eventive reports should be classified as not-at-issue eventive.
Take indirect discourse, whose parentheticalized variant is occasionally discussed under the header of inverted indirect discourse (Verhagen 2012) or syntactically parenthetical indirect discourse (Hunter 2016).

(61)
Mary doesn't want to come to my party tomorrow, Sue told me last week.
Note the unshifted tomorrow and my, which shows this is neither direct nor free indirect discourse. Without going into the syntactic/semantic details we'll assume something like a Pottsian 'comma feature' at LF to split the assertion into two dimensions, with which we compositionally derive a 2D logical form like (62) (Potts 2005), simplifying the contribution of the temporal elements: When we feed this into our discourse model we get a pre-update with the second dimension, followed by a proposal to update with the first -or rather a 'presentation' of the first, as it's an st term. If the proposal is not rejected, a suitably modalized version, with our epistemic, context-dependent MUST operator, gets added to the common ground. In the end we are thus left with only worlds where there was an event of Sue telling the speaker Mary wasn't coming and, as far as the speaker knows (relative to some contextual modal base), Mary isn't coming.
To evaluate this proposal for analyzing parenthetical eventives as not-at-issue eventive reports we'll no longer go through all our tests but just pick a few representative ones and see how the semantics predicts the observed behavior.
Parenthetical indirect reports are just as easily modifiable as their canonical counterparts. This is fully explained by our event-based analysis. However, the two 31 early access differ in their ability to directly answer a question about a saying event: 23 (63) A: What happened next? Did anyone dare say anything? B: Mary said in her thick New York accent that John is guilty. B':??John is guilty., Mary said in her thick New York accent.
On our analysis, the B' answer would amount to an update with Mary saying that John is guilty, followed by a proposal that John is guilty. But a proposal has to address the question to be felicitous. In this case, being guilty doesn't address the question whether someone said something, so we correctly predict the infelicity of the B' answer.
In a similar vein, we correctly predict the impossibility of using these constructions to answer the What makes you think? question (cf. (5)):

(64)
A: What makes you think that Mary is ill? B. #She is ill, John said.
Although here, in contrast to (63), it is not required that the answer is eventive, the evidence that is asked for does need to be provided as at-issue information, i.e. at the first dimension. Since in the logical form predicted for (64) this information would be at the second one (cf. (62)), infelicity is the result. The In Situ Interpretation Test is inconclusive for parenthetical reports. Example (65) does not have a reading where parenthetical John said is interpreted in the antecedent of the conditional: 24 (65) #If Peter is ill, John said, then he is ill.
'If Peter is ill according to John, then he is ill.' We can handle this impossibility with our 2D logical forms: the reportative component is on the second dimension and semantic operators only target the first. In light of the discussion about commitment in 5.3 it is worth noting that our account predicts an epistemically hedged commitment for parentheticalized indirect reports, as opposed to a total lack of commitment for canonical indirect reports. As pointed out above, applying our discourse model to an eventive 2D logical form like (62) amounts to an update with an event of telling the speaker that Mary isn't coming, followed by an update to the effect that for all the speaker knows, it must be true that Mary isn't coming. Following the final remarks from section 5.3 it is important to note that this 'must' can be interpreted rather weakly, depending on the context. If there's some reason to assume that the speaker doubts the reliability of Mary's testimony, the second update could just mean 'as far as the speaker knows from hearsay, John did it', rendering the actual commitment expressed by the speaker with this hedged update rather trivial. In sum, though we predict a hedged commitment for parenthetical reports, as for evidentials, the context dependence of the modal hedging can account for cases where the speaker is known to doubt (or explicitly disavows) the content of the report.

Pragmatic backgrounding in indirect speech
Sometimes a not-at-issue interpretation arises through global discourse structure and coherence assumptions rather than being forced by a syntactic device like parentheticalization. This happens for instance in canonical indirect discourse in discourses like (66) (e.g., Simons 2007, Hunter 2016 A: Why is John not at the meeting? B: Mary said that he is ill. We saw in (10) that an indirect discourse report can be used as answer to the question What happened next?, indicating that the reportative component is at-issue eventive. Cases like (66), however, show that it can also be the information in the complement that answers the question while the reportative component merely signals that the speaker has heard this information from Mary. This shows that a canonical indirect discourse complement can be made at-issue by discourse structure, resulting in a discourse-driven demotion of the reportative component to not-at-issue status. How best to model this kind of information structuring in the semantics/pragmatics interface remains an open question. One perhaps overly simplistic option would be to stick with our two-dimensional not-at-issue eventive logical forms and then posit a structural ambiguity or movement in the syntax/semantics interface. Alternatively, we can assume that under discourse structural pressure, we insert the Pottsian comma feature alluded to above as a 'coercion operator' to split the logical form into two dimensions. In any case, the effect will be that we select a two-dimensional logical form when discourse coherence demands it (i.e., when a one-dimensional logical form fails to provide a coherent discourse interpretation). 25 Concretely, in (66), the one-dimensional, at-issue eventive logical form would be incoherent because the occurrence of a speech act by Mary does not address the question about John's presence at the meeting. So we instead select the twodimensional logical form (whether through movement at LF or 'comma coercion', or otherwise): (67) ∧ ill(john) ∃e[say(e) ∧ agent(e) = mary ∧ content(e) = ∧ ill(john)] Our dialogue model indeed predicts that this discourse is felicitous, as the proposal that John is ill would address the QUD. Unless explicitly rejected by other speech act participants, this would eventually result in an updated context where (i) there was a speech act by Mary with the content that John is ill, and (ii) given some contextual epistemic modal base, for all the speaker knows, John must be ill.
Note that this eventive logical form accounts well for the possibility of modification (e.g., by adding yesterday to (66)). Some other tests are more difficult to apply, the reason being that the interpretation with the reportative component at-issue is always available as well. As a result, in the What Happened Next, What Makes You Think, Challengeability, and In Situ Interpretation tests, we'll automatically fall back on this at-issue reading when the one with the reporting component not-at-issue yields infelicity (and on a coercion approach, the latter one wouldn't even arise in the first place).

Reportative mood
There may be yet other sources of not-at-issueness in eventive reports. Consider for instance the Ancient Greek optative mood in examples like the Xenophon passage in (68) (full gloss in (12) The first optative (underlined) is in the complement clause under the verb of saying elegon. The following optatives are no longer part of the same sentence (as shown by the presence of the particle gar, see Bary 2018). Still, the optative morphology marks these clauses as continuations of the explicit report. Below, we first establish the not-at-issue eventive nature of this reporting strategy on the basis of the Anaphoricity Heuristic and In Situ Interpretation Test. The findings will drive our proposal for a not-at-issue eventive logical form, which we'll extend to the German reportative 34 mood, and evaluate briefly by examining the predicted behaviors on the other tests. As mentioned in section 2.2, out of the blue reportative uses of the optative are very rare. The most natural habitat of the Greek optative is the complement clause of an indirect speech report. Unembedded occurrences typically offer more content of a specific earlier introduced speech event (as in (68)), which we take as evidence for eventivity, as per the Anaphoricity Heuristic.
For the In Situ Interpretation Test we've been looking at embedding in conditionals. The final optative in (68) is a case in point. We see that the reportative contribution of this optative, embedded in a conditional, percolates up: the conditional clause as a whole is interpreted as the content of a speech report. Since this seems to be the only possible interpretation in other cases of antecedent-embedded reportative optatives as well, we can conclude on the basis of the In Situ Interpretation Test that the reportative meaning component is conventionally marked as not-at-issue.
The two test results above thus point towards a not-at-issue eventive analysis for the ancient Greek optative. More specifically, we translate the optative morphology as an operator that is raised above the clause, and that compositionally introduces a second dimension with a saying event, presupposing a specific conversation, modeled with a free variable E ranging over conversations (which we won't analyze further but which may be thought of as sequences of utterances). Applying our now familiar semantics and pragmatics for 2D logical forms, (70) will give the right results for the examples above: (i) the saying event is not-at-issue (as witness the projection out of the conditional in (68)), and (ii) it is anaphorically dependent on a salient conversation in the context (partially provided by a preceding at-issue eventive report in (68)).
In addition to the Ancient Greek optative, we will also consider data from the German reportative subjunctive (Fabricius-Hansen & Saebø 2004, Sode 2014, Kratzer 2016, which closely resembles the Greek optative for instance in forcing free-standing sentences to be interpreted as continuations of previous indirect speech reports, as in (71)  Augen. eyes 'He said she's pretty. She has green eyes, he said." German, Jäger 1971 For German as for ancient Greek out of the blue cases are dispreferred (Fabricius-Hansen & Saebø 2004:226), in stark contrast to, say, reportative modals and evidentials which tend to be completely fine out of the blue.
As we'll verify with other tests below, our at-issue eventive analysis in (69) together with our discourse framework will make the right predictions for unembedded occurrences. However, one of the striking features of reportative moods is that their reportative contribution seems to evaporate when embedded in indirect speech. Not only do we never get an in situ interpretation (leading to a double report reading *they said that someone said that Seuthes said things of supreme importance for (68) and *he said that someone said that she was pretty for (71)) but there is no projection either. Instead, the reportative meaning component of an embedded optative/subjunctive mood seems to be canceled by (or 'filtered by', 'bound by', 'concordant with') the overt reporting verb. On the current analysis we get for the first sentence of (71) The somewhat peculiar logical form in (72c) does not quite capture the concord intuition, i.e. that the subjunctive gets filtered out by the overt saying verb, or put differently that the subjunctive here functions as a device to signal to the interpreter that we're in the scope of an indirect discourse construction. In fact, when we consider what happens when we embed (72) as a whole under negation or in a conditional (or under another speech or attitude verb), the predictions of (72b) would start to falter. To remedy this, there are roughly two options explored in the literature: (i) Fabricius-Hansen & Saebø (2004) and Bary & Maier (2014), who have tried to give a purely (dynamic) semantic account of this cancelation effect whereby the reportative component is treated as a presupposition that literally gets bound (or satisfied) by the at-issue speech report; and (ii) Schlenker (2003) andvon Stechow (2002), who have analyzed this cancelation at the level of morphosyntactic feature agreement and deletion. In principle we could implement either option in our current framework, but a detailed implementation and comparison would take us too far afield. We end this subsection again with a closer look at the various diagnostic tests and see how our semantics predicts the behavior that we see. First, we have to 36 realize that much of our diagnostic manual is difficult to apply to an older stage of a language such as Ancient Greek: for example, as expected, we indeed do not find examples where the reportative contribution is challenged. But as always, the fact that something does not occur does not mean that it is impossible -especially since challenging is a dialogue move, and all we have is written texts. Nonetheless, statistical data from corpus analysis may provide some indirect evidence. In particular, it has been observed that discourse-coherence relations within long stretches of indirect discourse with the optative such as (68) (which continues even longer than given here) are always established at the level of the reported content, without discourse interference from a speech event (see Haug et al. 2018 for a similar phenomenon in Latin). We can see this as the result of the not-at-issueness of the reportative contribution, because only at-issue propositions are available for attaching further discourse relations.
As for the Modification Test, our not-at-issue eventive analysis predicts that modification should be impossible. This prediction is borne out, but again, we can never be sure that what is not found is impossible.
We are in a better position to apply our tests to the German subjunctive. The eventive, not-at-issue analysis correctly predicts the behavior on the What Happened Next Test: German subjunctive marking on a stative predicate cannot be used to answer a question that asks specifically about a speech event: We can conclude that the behavior on our tests is as predicted by the not-at-issue eventive analysis in (69).

Quotative evidentials
According to the evidentiality literature there are evidentials that involve a direct quote, the so-called quotatives. 26 On our analysis in section 5 the type of constructions that are commonly called reportative evidentials are not-at-issue non-eventive reports. That means that their semantics doesn't introduce a speech event of which the form can be specified, making the notion of a quotative evidential rather puzzling. A closer look at a sample of the languages that are claimed to have such quotatives, however, strongly suggests that they inhabit a different corner of the reportative landscape than (other) reportative evidentials.
The fact that the form of a speech event is specified already indicates the presence of a speech event variable in the logical form. The Modification Test confirms that quotatives are indeed eventive. 27 Consider first Plains Cree, an Algonquian language, spoken mostly in Saskatchewan and Alberta (Canada). Plains Cree is said to have a quotative evidential marker itwê in addition to reportative, nonfactual and dubitative evidentials (Blain & Déchaine 2007). An example is given in (75) (1) itwêw. QUO '. . . , "Well, no," he said, "I am not going to engage in that," he said.'

Plains Cree
When we look at the larger context in which the example occurs, it becomes clear that we're really dealing with a description of a speech event, including reference to a specific source, time and conversation: (76) Today, for example, when I spoke to this one [in the audience], I asked him about these six things, whether he was going to engage in them; and when I asked him the first time, for example, he said "Well, no, I am not going to engage in that." Kâ-Nîpitêhtêw 1998: 52 We showed in section 5 how not-at-issue non-eventive speech reports are used to specify the source of evidence for a main proposition. This is natural in a way: Being non-eventive they are less useful for vividly recounting a specific dialogue. Eventive reports can in principle be used for both: they can be used both to provide evidence for the main point and to help tell a story by describing specific speech events. However, as soon as only the form of a speech event is specified (and not the content), as is the case with quotatives, the contribution to the discourse can no longer be an evidential one, as there is no proposition introduced that evidence could be provided for. Note in this respect that in no sense does the actual speaker of (75) use itwê to provide evidence for the scope proposition (given by the reported words). In other words, so-called quotative evidentials are not evidentials in any semantic or intuitive sense.
Note also that morphosyntactically, the quotative is again the odd one out in the evidential paradigm of Plains Cree: whereas the other evidentials in this language are particles, itwê is a verb (Blain & Déchaine 2007: 261), as it can be inflected for first-, second-, or third-person, and tense/aspect. It also seems to be the only evidential form that allows recursive embedding (John said: "Mary said '. . . ' ").
We take these facts about Plains Cree to suggest that we're really dealing with an eventive report construction, with the person, tense and quote each specifying features of a reported speech event -on a par with the English direct speech construction in section 4.
In languages where the quotative marker is not a verb but a clitic or particle, quotatives also behave rather differently from other evidentials. In the evidential systems of these languages, quotatives are the only ones that can occur with imperatives (see e.g., Boye 2012:204-206 on Kannada). The incompatibility of evidentials with imperatives is expected since the notion of providing evidence to a scope proposition is not applicable in the case of non-propositional speech acts. 28 An eventive analysis, by contrast, allows quotatives with any speech act type, just as we can quote questions and imperatives in direct speech in English. On the basis of these data we suggest that quotatives are best analyzed as eventive, like direct speech constructions but unlike evidentials (in line with Boye's (2012) considerations to exclude quotatives as evidential markers). 29 We conclude that the type of constructions that are sometimes called quotatives live in a different side of the speech reporting landscape than reportative evidentials, viz., the eventive side. Unfortunately we do not currently have the necessary data 28 Propositional analyses of imperatives exist, most notably that of Kaufmann (2012). Such accounts then would need to explain why it is that only quotatives and not (other) reportative evidentials are compatible with imperatives. 29 Korotkova's (2017) also sets what she calls 'quotative evidentials' apart from other kinds of evidentials. Her definition of quotatives however differs from ours in that in her terminology they do not necessarily involve a direct quote but they express a 'relayed speech act'. Indeed, there seems to be a difference in this respect between e.g., Mbyá je which features prominently in her account (using data from Thomas 2014) on the one hand, and Plains Cree itwê and Kannada anta on the other, in terms of the behavior of indexicals (see Thomas 2014: section 3.3 for the relevant Mbyá data). The analysis proposed in the present paper only holds for quotatives in our stricter sense.
to determine whether these quotative markers are at-issue or not-at-issue (but we imagine that not-at-issue behavior is what made Blain & Déchaine 2007 and others treat itwê as a quotative evidential in the first place). It also remains to be seen whether the various strategies labeled as quotatives in this sense (conceptualized as 'evidentials' that specify the form of a reported speech act, see footnote 26 for other uses) form a homogeneous class or differ in fundamental respects from each other.
7 At-issue non-eventive reports: according to We conclude with a brief examination of the most elusive category predicted by our typology: at-issue non-eventive reports. This cell has fewer inhabitants than the others, which is probably to be expected from a functional, communicationtheoretic perspective: if the reporting component (and not the content of the report) is what is at-issue in a given speech report, one can expect that there is some interest in the speech event (e.g. its form, and how it relates temporally to other events described), resulting in the choice for an eventive reporting strategy. Why else would the reporting component be at issue? That said, the two notions are not contradictory and one context where we would expect the combination to be felicitous is in answer to the question What makes you think (that Mary is ill)?, where the (at-issue) answer could theoretically just be that the speaker has hearsay evidence for this. In section 1 we've already suggested that English according to, as in (77), may belong to this category. (77) According to Theresa May, any extension of Article 50 will have to be a short, one-time-only deal.
The What Makes You Think Test unfortunately doesn't give a clear result: (78) A: What makes you think that Mary is ill? B. John said/told me that Mary is ill. B'. ?According to John Mary is ill.
Although dispreferred over indirect speech as in (78B), according to is better in such contexts than e.g., Dutch schijnen (see (5)). The other tests give a clearer picture. Take the In Situ Interpretation Test. We can see that, just like Dutch volgens ('according to') in (3), the reportative meaning component of according to doesn't project by noting that a conditional like (79) has a non-tautological reading: If Piet is ill according to Jan, then he is ill.
In this, according to patterns with canonical indirect discourse ('if Jan says Piet is ill, 40 then he is ill') rather than with not-at-issue speech reports (e.g., the nonsensical or at least near tautological 'if Piet is ill, reportedly, then he is ill'). This in situ behavior indicates that we're dealing with an at-issue reporting strategy. Next, consider the Anaphoricity Heuristic to detect eventivity. It seems that an according to construction is perfectly fine out of the blue, as in a newspaper headline or opening paragraph. It requires no overt speech report construction already established in the discourse record. It thus patterns with non-eventives, like evidentials and modals, rather than with eventives like direct and indirect discourse and reportative moods.
The Modification Test points in the same direction. Speech act modifiers are clearly out, supporting a non-eventive analysis (given that we've ruled out not-atissueness already): 30 (80) According to Sumo {*loudly/*over the phone} the movie is terrible.
A semantics for according to that predicts precisely these at-issue non-eventive behaviors is rather straightforward. In fact, it's just an application of the traditional Hintikka analysis of propositional attitude reporting as intensional operators. All we need is a generalization of the HEARSAY operator that takes a source argument (cf. p, x ∈ HEARSAY c w = 1 iff all w compatible with what the speaker heard from x in w, p (w ) = 1 In sum, our typology predicts four classes of reports. For three of these it's been easy to find examples from different languages that meet the criteria. One class, at-issue non-eventives, is clearly less densely populated, although the predicted semantic profile is quite clear and not contradictory. We've tentatively argued here that English according to fits the description although the test results are less clear here than in the other classes, pointing to a need for further research.

Conclusion
We often refer to what other people have said. And we do so for different reasons and using a variety of different constructions. In this paper we've proposed a classification of speech reporting strategies along two dimensions: at-issueness and eventivity.
The first distinction has already featured prominently in the (evidentiality) literature. It concerns the distinction between reports where the fact that something is said is at-issue versus reports where not the reportative component but the content of what was said is the primary piece of information.
The second distinction of our classification is just as important to understand the differences between the various speech report options natural languages offer and the ways these are used in communication. Only some types of reporting introduce a speech event into the discourse record. It is only these eventive reports that can be used to describe certain details about the speech event itself -e.g., the actual words, or the manner of speaking. This aspect makes eventive reports especially useful for vivid characterizations of dialogues between characters in a story. By contrast, non-eventive speech reports are more at home in everyday communication, where it can be useful to specify what kind of evidence we have for certain claims -whether to convince the hearer of the primary content or rather to distance ourselves from it.
These two distinctions result in four classes of speech reports. In this paper we have proposed a number of tests to determine to what class a given reporting strategy belongs. We then proposed matching semantic analyses for various constructions belonging to each class, and developed a very general pragmatic dialogue/discourse model à la Farkas & Bruce (2009) to model how these logical forms can be used to update a discourse context (containing the common ground, a stack of QUDs and a stack of proposals). Finally, in each class, we verified that these semantic analyses in combination with this general pragmatic framework correctly predict the behavior on the tests.
Sketching the landscape of reports along two comparative concepts has brought order to the scattered body of data and theories about the various reporting constructions. In addition to the two well-known classes of reporting constructions -at-issue eventives (direct and indirect discourse) and not-at-issue non-eventives (reportative evidentials) -we also identified report markers that express the two other possible combinations of features. While according to is arguably best understood as an at-issue non-eventive report marker, reportative moods have the opposite combination of features. By charting the landscape of speech reports, we've brought out new insights on the nature of so-called quotative evidentials (as constructions that are eventive and not properly deserving the label 'evidential') and German and Ancient Greek reportative moods (as report markers that have their not-at-issueness in common with evidential markers, but share their