Type of the Paper (Article

The paper examines distributed extraction coordinations, in which different elements 6 move out of conjuncts of a single coordination, as in Which book and which magazine did Mary buy and 7 Amy read respectively, from a crosslinguistic perspective. A number of properties of such coordina8 tions are discussed, which includes showing that they are also subject to the ATB requirement, 9 which will shed light on the nature of the ATB phenomenon itself. It is also shown that there is a 10 rather strong restriction on distributed extractions which confines such extractions to one context 11 and completely excludes one type of movement, in particular head-movement, from participating 12 in them. The higher coordination is shown to be formed during the derivation and to be semanti13 cally expletive. Distributed extraction constructions are also shown to have consequences for the 14 proper analysis of a number of phenomena, including subject-oriented anaphors, right node rais15 ing, tough-constructions, agreement, and clausal structure. Regarding subject-oriented anaphors, 16 the paper teases apart different approaches to subject-oriented anaphors based on constructions 17 where different elements fill SpecvP and SpecTP (the latter undergoes agreement with T and the 18 former binds subject-oriented anaphors). 19


Introduction
It is well-known that extraction out of conjuncts is disallowed, unless the moving 24 element moves out of each conjunct. This well-known phenomenon is illustrated by 25 (2)-(3). The ban on extraction out of conjuncts, given in (1), is standardly referred to as the 26 Coordinate Structure Constraint (CSC), and the rescuing effect in (3) as 27 across-the-board-movement (ATB). (1) Extraction out of conjuncts is disallowed. 30 (2) *Whoi did you see [enemies of ti] and John? 31 (3) Whoi did you see [friends of ti] and [enemies of ti]? 32 33 Both the CSC and the ATB exception were noted in Ross (1967). (4)  (4) In a coordinate structure, no conjunct may be moved, nor may any element contained 37 in a conjunct be moved out of that conjunct (Ross 1967:98-99) 38 (5) There is an important class of rules to which (4) does not apply. These are rule sche-39 mata which move a constituent out of all the conjuncts of a coordinate structure (Ross 40 1967:107) 41 42 The upshot of (4)-(5) is that extraction of X out of a conjunct is unacceptable unless X is 43 extracted out of each conjunct of the coordination. Postal (1998) provides strong evidence that which book and which magazine undergo sep-51 arate extractions out of the conjuncts in (6), and Zhang (2010) argues that such cases in-52 volve coordination-formation that takes place after (more precisely, through) move-53 ment. 3 These examples violate the CSC ban in (1). They also do not fit the ATB pattern in 54 (3): it is not the case that the moving element is extracted out of each conjunct in (6). (6) in 55 fact appears to involve two separate extractions, of two different elements, out of the 56 conjuncts. One may then expect (6) to be even worse than (2). 57 The goal of this paper is to examine this kind of constructions, which for ease of 58 exposition I will refer to as distributed extractions. Additional cases of distributed ex-59 tractions will be discussed in the effort to examine restrictions on distributed extractions. 60 Evidence will be provided that all these cases involve coordination formation after 61 movement (I will refer to such coordination as late coordination) and the precise timing 62 of (and the operations involved in) this late coordination formation will be discussed. It 63 will also be argued that distributed extractions are actually subject to the ATB require-64 ment, which will shed light on the nature of the ATB phenomenon itself. It will also be 65 shown that there is a rather strong restriction on distributed extractions which confines 66 such extractions to one context and completely excludes one type of movement, in par-67 ticular head-movement, from participating in such extractions. 68 It should, however, be noted that one of the main goals of the paper is descriptive, 69 namely to broaden the scope of the phenomenon empirically. There has been very little 70 discussion of the phenomenon in question outside of English (and outside of construc-71 tions like (6)). In this respect, the paper will bring in additional languages, with con-72 structions which are quite different from (6). The restrictions on distributed extraction 73 coordination established below should also be looked at from this perspective: their goal 74 is to empirically broaden the scope of the phenomenon in question-in this respect the 75 restrictions are actually more important than their deductions. At any rate, one of the 76 main goals of this paper is to prompt further crosslinguistic investigation of the phe-77 nomenon in question as well as several related properties of coordination which are 78 discussed below. 79 Going beyond coordination, the coordination data discussed in this paper will be 80 shown to have consequences for a number of coordination independent mechanisms in 81 that they shed light on how the mechanisms in question should be analyzed. To mention 82 just some of them here: subject-oriented anaphors, right node raising, 83 tough-constructions, and agreement. Regarding subject-oriented anaphors, the paper will 84 tease apart different approaches to subject-oriented anaphors based on constructions 85 where different elements fill the SpecvP and the SpecTP position (the latter undergoes 86 agreement with T). The coordination data discussed in the paper will also be shown to 87 have consequences for determining clausal structure more generally. 88 The paper is organized as follows. In section 2 I present arguments (mostly from 89 Postal 1998, but also new arguments) that (6) involves extraction out of each conjunct. 90 The section will also show that the ATB requirement (more precisely, a reformulated 91 ATB requirement) is operative with such constructions. In sections 3 and 4 I present ad-92 ditional cases of distributed extractions and show that these additional cases are also 93 subject to the reformulated ATB requirement. The possibility of mixing distributed ex-94 traction and traditional ATB in the same sentence is also discussed. Section 5 establishes a 95 new generalization regarding the availability of distributed extractions and section 6 96 discusses islandhood of distributed extraction constructions. Section 7 and 8 examine the 97 exact timing (and the mechanism) of late-coordination formation and explore conse-98 quences of distributed extraction coordinations for other phenomena, like sub-99 ject-oriented anaphors, right node raising, tough-constructions, agreement, and clausal 100 structure. Section 9 is the conclusion, and Appendix discusses a related construction (re-101 ferred to as wh&wh coordination) that also involves late coordination, outlining the 102 range of possible crosslinguistic variation in the relevant domain. It also discusses the 103 ordering of elements involved in late coordination formation.  Postal (1998) gives strong evidence that each wh-phrase is separately extracted from 107 the conjuncts in constructions like (6). A rather strong argument to this effect is provided 108 by the possibility of binding into the individual conjuncts in (7), where which man binds 109 an anaphor in the first conjunct and which woman binds an anaphor in the second 110 conjunct. 111 112 (7) [Which man]i and [which woman]j did respectively the doctor talk to ti about himselfi, 113 and the lawyer talk to tj about herselfj (Postal 1998:161) 114 115 Such licensing is also possible with parasitic gaps, as shown by (8), where the first 116 wh-phrase licenses a parasitic gap in the first conjunct and the second wh-phrase licenses 117 it in the second conjunct.
118 119 (8) [Which secretary]1 and [which programmer]2 did Jerome respectively fire t1 after 120 finding t1 drunk and hire t2 after finding t2 sober? (Postal 1998:136) 121 122 Another argument comes from cases where the extracted elements contain an anaphor: 123 the anaphor can be bound within the first conjunct or within the second conjunct, as in (9) 124 (for a somewhat different reconstruction effect, see (83)). 4 The binding relations can also 125 be combined, as in (10) (Postal 1998:135) 142 143

The ATB requirement on distributed coordinations
The evidence discussed in the previous section shows that distributed coordination 145 constructions like (6) involve separate wh-movements from each conjunct. As such, they 146 do not fit the traditional ATB-exception-to-the-CSC schema, where the CSC is voided if 147 the moving element moves out of each conjunct. Notice, however, that examples like (6) 148 do actually involve movement out of each conjunct, the difference between (3) and (6) 149 being that in (3) it is the same element that moves out of each conjunct while in (6) 150 different elements move out of the conjuncts. 151 Interestingly, it turns out that the ATB requirement holds for constructions like (6) 152 as well. This is shown by the unacceptability of (12)-(13), which contrast with (14). 153 154 (12) *Which booki and which magazinej did [John buy ti] These data indicate that the ATB requirement is at work in the construction under con-162 sideration: movement still must take place out of each conjunct. This means that the ATB 163 requirement needs to be reformulated: it is not the case that the moving element must 164 move out of each conjunct; rather, movement must take place out of each conjunct. It can 165 be the same element moving out of each conjunct or different elements: as long as there is 166 a gap in each conjunct the ATB requirement is satisfied. I will refer to the cases where 167 different elements move from the conjuncts as non-ATB ATB. 5 168 Not only does the ATB requirement hold for distributed coordination constructions 169 but it in fact holds in the same way as with regular ATB constructions. It is well-known 170 that there is an interpretative parallelism requirement on regular ATB. Thus, both gaps 171 must be either subjects or objects (the requirement is actually more detailed than that, it 172 can also affect two internal arguments and concerns thematic prominence-see Franks 173 1993Franks 173 , 1995 Zhang (2010:193) observes the data in (17)- (19), which indicate that the parallelism re-179 quirement in question also holds for non-ATB ATB. Zhang does not discuss cases involving cross-clausal extraction. With regular ATB, the 186 parallelism requirement in question is relaxed; i.e. it does not hold with cross-clausal 187 ATB, as (20) shows. 188 189 (20) I wonder whoi [John saw ti] and [Peter thinks ti kissed Mary]. 190 191 The same holds for distributed extraction coordinations. 6 (21) Which writeri and which actorj does John adore ti and Peter claim tj will succeed in 193 Hollywood respectively? 194 195 The ATB requirement thus holds in the same way in distributed coordination construc-196 tions as with regular ATB constructions, which further indicates that the former are a 197 type of ATB constructions although they do not involve extraction of the same element 198 (hence the term non-ATB ATB). 199 In the following sections I will present additional cases of non-ATB ATB which are 200 quite different from English examples like (6). We will see that the ATB requirement 201 holds in these cases as well: although different elements are moving out of the conjuncts 202 there must be movement out of each conjunct. The cases discussed in the following sec-203 tions will also enable us to establish additional restrictions on non-ATB ATB.  205 I now turn to a case of distributed ATB in Serbo-Croatian (SC) which has interesting 206 additional properties. SC productively allows left-branch extraction of adjectives (see 207 Corver 1992, Bošković 2005, 2013a, Despić 2011, Talić 2017 It also allows it in distributed coordinations. One difference from English wh-movement 214 involving distributed coordination is that such cases involving adjectival ATB in SC do 215 not require "respectively" (in fact, there is no clear counterpart of respectively in SC; note 216 that in some cases respectively is not needed in English, see (93) It is also possible to have three adjectives in this type of constructions, as in (25) (27) improve if the first 245 two conjuncts are pronounced as a single prosodic unit (followed by a pause), with an-246 other coordinator, as in (28). What is going on here is that suknja i kaput form a coordina-247 tion, which is then coordinated with šareni šešir. In other words, we are not dealing here 248 with a single coordination with three conjuncts, as in (26)-(27), but with two separate 249 coordinations, each of which has two conjuncts: suknja i kaput forms a ConjP that is itself 250 located in the Spec of a ConjP (the head of the second coordination takes šareni šešir as its 251 complement), as shown in (29) This in fact holds for regular ATB as well, as indicated by (30) (assuming the same 268 prosody as in (29), with the first two conjuncts pronounced as a single prosodic unit (with 269 a pause following them); crvene here undergoes regular ATB extraction from the first 270 ConjP-as result, "red" modifies both "skirts" and "dresses"; there is a potential inter-271 fering factor here that is controlled for in fn. 47). The ATB requirement is then still satisfied in (32): (32) is in fact acceptable only on the 309 reading on which there is an AP-gap in the base position of each of the conjuncts in (32). 310 What is particularly interesting about this example is that it involves a mix of non-ATB 311 ATB and regular ATB. Examples like (32) then provide evidence that non-ATB ATB can 312 be mixed with true ATB. 313 Another example of this sort is given in (34), which involves regular ATB between 314 'red skirt' and 'red shirt' (košulja is feminine). Apparently, a traditional ATB dependency can only be formed between contigious NPs 342 here. There can be no ATB between 'red skirt' and 'red hat' given that the adjective needs 343 to agree with the nouns and these nouns have different gender (suknja is feminine and 344 šešir masculine). Also, there can be no ATB between 'white coat' and 'white skirt' since 345 these nouns also have different gender (kaput is masculine and suknja feminine). 346 Interestingly, there can apparently be no ATB between 'white coat' and 'white hat'. There 347 is no gender disagreement issue here since the nouns have the same gender.

348
The same effect is found in English. Thus, (40), where given the pragmatics of the 349 example regular ATB dependency has to hold between the first and the third conjunct, 350 skipping the second conjunct, is worse than (36)-(38), where this is not the case. This 351 contrast also provides evidence that the English and the SC construction in question 352 should be treated in the same way (given that both exhibit the contiguity effect). We may be dealing here with a locality effect on traditional ATB formation, where it is 358 not possible to skip a potential ATB site. 13 359 Alternatively, this may be related to a general interpretive effect associated with 360 distributed extraction coordinations. Notice first that examples like (6) are not ambigu-361 ous: the first trace must correspond to the first wh-phrase and the second trace to the 362 second wh-phrase. In other words, only a crossing wh-trace dependency is possible here; 363 a nesting dependency, which would give an interpretation where the first trace corre-364 sponds to the second wh-phrase, is disallowed. This is a general property of distributed 365 extraction coordinations. Thus, (41) gives the only possibility for the distributed 366 interpretation of the extracted adjectives in this SC example (ignoring the irrelevant reg-367 ular ATB interpretation, on which each conjunct is red, white, and colorful), where all 368 adjectives have the same gender, and (42) illustrates the same effect for English distrib-369 uted coordination constructions involving three conjuncts, where the indices again indi-370 cate the only possibility for the interpretation of the conjuncts (the parallel behavior of 371 the SC and the English construction under consideration in this respect can be taken as 372 another argument for treating the two in a uniform manner). (43) involves a mixture of crossing and nesting dependencies (the last trace is involved in 388 a nesting dependency). It then seems plausible that it is ruled out due to the general 389 crossing dependencies requirement on distributed coordinations. 390 As for the source of the effect of question, notice that what we are dealing with here 391 is essentialy a matching effect: the order of the conjuncts within the newly formed ConjP 392 must match the order of the conjuncts from which extraction takes place in the original 393 ConjP. Given that in this kind of cases the co-ordination structure is in a sense 394 "re-created" in a higher position, with another ConjP, it seems natural to assume that 395 there should be some parallelism between the two coordinations where the order of the 396 conjuncts in the higher ConjP should correspond to the order of the conjuncts (which 397 contain the relevant gaps) in the lower ConjP, which means that the first conjunct should 398 correspond to the first gap, the second conjunct to the second gap and so on. The result of 399 this is strictly crossing dependencies. Under this approach the ordering effect is essen-400 tially a parallelism effect (see, however, the appendix, where the parallelism effect is 401 deduced from independent considerations). 402 Before concluding this section, one potentially interfering issue should be discussed. 403 Consider (44). Gračanin-Yuksek (2007) and Citko and Gračanin-Yuksek (2013)  Notice also that adjectives involved in the coordination in question must bear the 434 same case, as shown by the unacceptability of (48) (as shown by (47) It is also worth noting here that SC clitics are second position clitics (see Bošković 449 2001 and references therein); as such they are standardly used as a constituenthood test 450 (since they cannot follow more than one constituent). Clitic placement in (49) then con-451 firms that crvene i bijele is a single constituent, which is indeed the case under the coor-452 dination-in-the-moved position analysis.  (51); not surprisingly, they also cannot undergo left-branch extraction 477 as a coordination, as shown by (52). 478 Turning now to non-ATB ATB left-branch extraction, such cases also involve mul-479 tiple LBE. Notice, however, that (24)  What we see at work in (53) is what is at work in (50)-(51). The relevant elements, the 486 demonstrative and the adjective, can undergo left-branch extraction; in fact they can be 487 involved in multiple left-branch extraction, as shown by (50). However, these elements 488 cannot be coordinated, as shown by (51), hence they cannot undergo left-branch extrac-489 tion as a coordination (cf. (52)). The ungrammaticality of (53) is not surprising from this 490 perspective: (53) is ruled out on a par with (51) because one and bijele cannot be coordi-491 nated. That the restriction in question is relevant in (53) is not surprising given that ele-492 ments that undergo non-ATB ATB are involved in a coordination with each other. 493 However, in contrast to (51), where the demonstrative and the adjective are coordinated 494 in their base position and modify the same noun, the demonstrative and the adjective 495 obviously cannot be involved in a coordination in their base position in (53). This is so 496 because of the interpretation of (53), which is "those coats and white dresses"-the 497 demonstrative and the adjective do not modify the same noun in (53), in contrast to (51). 498 The coordination in (53) can then only take place after movement, since the relevant 499 elements are clearly not coordinated in their base-position. The individual movements 500 themselves also must be possible in (53), given that such multiple left-branch extraction is 501 in principle possible, as shown by (50) (see Bošković 2016). (53) is thus ruled out because 502 it involves illicit coordination, where the coordination takes place after movement. The 503 data in question then also provide evidence that we are indeed dealing here with late 504 coordination formation (i.e. non-base coordination). 505 Notice that we also have additional evidence here that non-ATB ATB examples in-506 volving left-branch extraction do not involve a larger coordination with ellipsis in the 507 first conjunct. Under such an analysis we would not be able to appeal to the impossibility 508 of coordination of a demonstrative and an adjective, i.e. the ungrammaticality of (51), 509 since this is not what would be coordinated in (53) under that analysis. 17 510 Another issue that is relevant here is that a clitic (mu) can intervene between the 511 demonstrative and the AP in (50), as shown by (54). Recall that this is not possible with 512 non-ATB ATB constructions, as shown by (46) All this confirms the coordination in the moved position analysis of (24)/(46). Elements 519 undergoing multiple LBE need not move to the same position, hence a clitic can intervene 520 between them, as in (54). Elements involved in non-ATB ATB (as in (46)), on the other 521 hand, are located in the same position, in fact non-ATB ATB involves a coordinated 522 phrase, hence a clitic cannot intervene between the relevant elements, which are coordi-523 nated with each other.
The above data thus provide additional evidence that coordination formation 525 should not be restricted to base-generation (i.e. lexical insertion/external merge), i.e. it 526 should not be restricted in such a way that it can only occur pre-movement. 527 In summary, in this section we have seen another case of non-ATB ATB, which also 528 involves non-base coordination formation and which is also subject to the ATB require-529 ment. We have also seen that the ATB requirement does not apply across ConjPs. 530 Furthermore, we have seen that non-ATB ATB can be combined with traditional ATB 531 and that the crossing dependencies requirement on distributed coordination is maintai-532 ned regardless of whether such constructions involve only non-ATB ATB or a mixture of 533 non-ATB ATB and traditional ATB. Importantly, it is possible to extract the NP from the conjuncts in (55), with a coordination 543 structure recreated in a higher position (for some speakers 'respectively' is optional here, 544 others require it; see also fn. 18). As another parallel to SC, (61) and (58) are actually marginally acceptable if there is a 592 pause following the second conjunct in the lower ConjP (i.e. if the first two conjuncts in 593 the lower ConjP form a separate intonational phase). This is the same prosody as the one 594 discussed above with respect to SC (28). Recall that this prosodic pattern, on which the 595 first two conjuncts in (57) are pronounced as a single prosodic unit, has a different deri-596 vation, on which 'three oranges' and 'five bananas' form a coordination (as reflected in 597 this unit also forming a prosodic unit), which is then coordinated with "two grapes". In 598 other words, on this prosodic pattern we are dealing here with two separate coordina-599 tions, each of which has two conjuncts. The Japanese construction under consideration in this section thus represents another 617 case of non-ATB ATB, where movement takes place out of each conjunct, but it is diffe-618 rent elements that are moving out of the conjuncts. As in the case of non-ATB ATB 619 examples from English and SC discussed above, the ATB requirement holds in this case 620 too: although different elements are extracted, extraction must take place from each 621 conjunct. 622 5. When is non-ATB ATB possible? 623 The above data confirm the existence of non-ATB ATB, where there is movement 624 out of each conjunct but different elements are moving out of the conjuncts. In other 625 words, the ATB requirement should be stated in a such a way that it does not require that 626 the same element moves out of each conjunct but simply that there is movement out of 627 each conjunct. 628 There is another interesting property of non-ATB ATB. All the cases involving 629 non-ATB ATB discussed above involve coordination formation in the moved position. 630 What happens when non-ATB ATB is attempted without coordination formation in the 631 moved position? Consider in this respect (65) (65) involves extraction of different elements from a single coordination without coordi-637 nation formation in the higher position. In English this requires moving wh-phrases to 638 different +whCPs, which in turn brings in a wh-island violation. Still, (65) is clearly much 639 more degraded than typical wh-island violations. 22 The fact that the contrast between 640 (65) and (6) is stronger than a typical wh-island violation suggests that coordination 641 formation in the moved position is necessary for non-ATB ATB. 642 Consider also (66)-(67), which also involve non-ATB ATB without coordination 643 formation in the moved position. Both examples are unacceptable. Furthermore, (67), 644 where movement does not take place out of each conjunct (hence it is not in accordance 645 with the ATB requirement), is even worse than (66), where movement does take place out 646 of each conjunct (in ATB fashion) (in (67) there are two moved elements and two gaps, 647 while in (66) there are two moved elements and three gaps; the example mixes non-ATB 648 ATB and ATB). The contrast between (66) and (67) parallels in the relevant respect the contrast between 656 SC (26) and (27), indicating that the ATB requirement still holds in such cases. Both ex-657 amples are, however, unacceptable. What seems to be going on here is that performing 658 non-ATB ATB without coordination formation in the moved position leads to a violation, 659 call it a violation of requirement X (to be discussed more below): X is violated in both (65) 660 and (66). The reason why (67) is even worse is that it violates X as well as the ATB re-661 quirement that there needs to be movement out of each conjunct of a coordination. 662 Notice now that in (65), the wh-phrases that are moving out of the coordination are 663 interpreted in different SpecCPs (i.e. different clauses). It is not out of question that this is 664 the source of the ungrammaticality of (65); i.e. it may be that for some reason wh-phrases 665 undergoing this kind of extraction must be interpreted in the same SpecCP, in which case 666 (65) would not necessarily show that non-ATB ATB requires coordination formation in 667 the moved position. This potentially interfering factor cannot be controlled for in English, 668 but it can in SC, SC being a multiple wh-fronting language. Let us then test the possibility 669 of non-coordinated non-ATB ATB with multiple wh-fronting in SC. The relevant exam-670 ples are given below. (68), involving non-ATB ATB without higher coordination, is un-671 acceptable. (69), its counterpart involving coordination in the higher position, is clearly 672 better than (68) The interfering factor noted above with respect to English (65) also does not arise with 681 respect to SC non-ATB ATB constructions discussed in section 3. These constructions also 682 require coordination formation in the moved position, as shown by the contrast in (70) The data discussed in this section thus indicate that non-ATB ATB requires coordination 697 formation in the moved position, i.e. the elements undergoing non-ATB ATB must par-698 ticipate in a coordination in their final position. 699 Recall now the example noted in fn. 5, repeated here, which is unacceptable alt-700 hough, just like (14), it involves extraction (of different elements) from each conjunct. The 701 issue here is that, in contrast to (14), which involves wh-movement out of each conjunct, 702 (72) involves wh-movement out of the second and third, and head-movement out of the 703 first conjunct. If the ATB requirement simply requires that there is movement out of each conjunct, 709 there is then no violation of the ATB requirement here. The ungrammaticality of (72) can, 710 however, now be accounted for independently of the ATB requirement. We have seen 711 above that when different elements are extracted out of conjuncts of a single ConjP they 712 must participate in a coordination in the higher position. This is not the case with did in 713 (72). The unacceptability of (72) then follows independently of the ATB requirement.
But there is a more general issue here. In English, distributed coordination is also 715 possible with A-movement, as in (73) (respectively is not needed in (73)). 716 717 (73) The dogs and the roosters barked and crowed all night. (Zhang 2010:170) 718 719 Japanese, however, does not allow constructions like (73) on the relevant reading (the 720 distributed interpretation is difficult to obtain in (74), where the pragmatically implausi-721 ble regular ATB reading where (the) dogs and (the) birds were both barking and flying is 722 strongly preferred), whereas SC patterns with English in allowing them (see (75)), which 723 can be taken to indicate that distributed coordination can be more restricted with A-than 724 with A'-movement, given that such constructions clearly involve the former. 25 725 726 (74) Inu-to tori-ga hitobanzyuu hoe-te ton-da. There may then be something more general about head-movement that disallows dis-739 tributed coordinations involving head-movement. Interestingly, Kayne (1994) argues 740 that head coordination is quite generally disallowed (see his work for evidence to this 741 effect and discussion how traditional head coordination constructions should be treat-742 ed 26 ). If distributed extractions require that extracted elements be coordinated, as argued 743 above, and if head coordination is quite generally disallowed, as Kayne (1994) argues, it 744 then follows that distributed coordinations with head-movement, as in (76), will be dis-745 allowed. The impossibility of distributed extraction involving head-movement can in fact 746 be taken as another argument for the proposed coordination-in-the-moved position re-747 striction on non-ATB ATB. 748

749
In this section I briefly note a locality effect associated with late coordination for-750 mation. SC allows extraction of conjuncts, as in (77)  While it is not trivial to implement this formally, intuitively it seems clear what is going 768 on here: ConjP that is formed after movement, i.e. ConjP not located in the base position, 769 is an island (such ConjP would in fact be a barrier in Chomsky's 1986 Barriers system). 770 In fact, not only conjunct extraction, but extraction out of a conjunct is also disal-771 lowed from a coordination formed by movement. This is shown by (79), involving ATB 772 wh-movement out of a late-formed ConjP located in SpecCP, which is clearly worse than 773 simple extraction out of interrogative SpecCP, as in (80) Late-formed coordinations are apparently islands, disallowing any kind of extraction, 780 even extractions that are in principle possible out of regular (i.e. base-generated) coor-781 dinations.

783
While the primary goal of this paper is not to provide a full analysis of distributed 784 extraction coordinations-it is simply premature to do that before the empirical domain 785 of the phenomenon is properly determined (the main goal of this paper being to make a 786 contribution to that end)-in this section I will nevertheless address the issue of how late 787 coordination formation is to be implemented, focusing on its timing. 788 Zhang (2010) discusses examples like (6) and argues that they involve coordination 789 formation through movement. More precisely, she argues that the higher ConjP is 790 formed through sideward movement (see Nunes 2004). 27 On this analysis, the higher 791 ConjP of (6) (the relevant steps of the derivation of (6) are outlined in (81)) is formed not 792 by regular (i.e. upward) movement but by sideward merger of the relevant elements into 793 ConjP (see (81b)), which is introduced into the structure directly in its final position, the 794 interrogative SpecCP (see (81c)). While the analysis captures the most prominent property of distributive extraction coor-802 dination, namely that it involves late-coordination formation, it faces issues with some of 803 the data discussed above. Recall that distributed extraction coordinations exhibit island 804 effects, as illustrated again below with an inner island effect (cf. also SC (31) and Japanese 805 (63)- (64)). Under this analysis we cannot capture such islandhood effects, since the 806 wh-phrases do not undergo movement out of the island. The wh-phrases that participate in late coordination license parasitic gaps within their 825 initial conjuncts here. As is well-known, a wh-phrase in situ cannot license a parasitic 826 gap: a parasitic gap is licensed by a moved wh-phrase that c-commands the parasitic gap. 827 Under Zhang's analysis, there is never a c-command relationship between the moved 828 wh-phrases and the parasitic gaps which they license in (84). 829 These facts indicate that some regular (i.e upward) movement must be involved in 830 the derivation of distributed extraction coordinations. Under Zhang's analysis there is no 831 regular movement, as a result of which the coordination is formed (i.e. integrated into the 832 structure) in the final position. While this captures the late-coordination formation re-833 quirement, it essentially does it too late. However, while the above facts indicate that 834 regular movement must also be involved in the derivation of distributed coordinations 835 (note that sideward movement obviates island effects, see Nunes 2004 and the discussion 836 below) they do not necessarily mean that Zhang's sideward-movement analysis is fataly 837 flawed. The late-formed ConjP can still be formed through sideward movement, as long 838 as this ConjP is introduced into the structure earlier, not in the final position (e.g. within 839 the same phase as the original ConjP, but this will be revised below), in which case the 840 late-formed ConjP would be moving out of the island in (82), and the movement would 841 bring the anaphor close enough to John in (83) to satisfy Condition A during the deriva-842 tion. (This would still leave (84) unnacounted for; I will return to that example be-843 low-see the discussion of (119), which resolves the issue in question.) The modification 844 of Zhang's analysis, on which distributed coordinations involve a combination of side-845 ward movement and regular movement, as a result of which the higher coordination is 846 formed earlier than on Zhang's analysis (though it is still formed during the derivation) is 847 in the spirit of the well-known fact that in ATB constructions, there cannot be an island 848 boundary between the edge of the second conjunct and the original extraction site within 849 that conjunct, which under Nunes's sideward-movement analysis means that the rele-850 vant element needs to get to the conjunct edge, i.e. 'close' to its sideward movement site 851 in the first conjunct. It is then not that surprising that the newly formed ConjP, which is 852 also formed through sideward merger from the original ConjP under Zhang's analysis, 853 cannot be indefinitely far from the original ConjP, which means that it should be intro-854 duced into the structure earlier, not in the final position. according to which extraction of different elements from a coordination is possible only if 860 they are themselves later involved in a coordination. The derivationally-formed coordi-861 nation is essentially semantically expletive, the elements involed in our derivational-862 ly-formed coordination are not interpreted as coordinated. E.g. the interpretation of (6) is 863 'which book did John buy and which magazine did Bill read', there is no coordination of 864 which book and which magazine in the interpretation of this example. Similarly, there is no 865 coordination of the dogs and the roosters in the interpretation of (73), whose interpretation 866 is that 'the dogs barked and the roosters crowed'. The same holds for any of the SC ex-867 amples discussed above; thus, there is no coordination of 'to whom' and 'for whom' in 868 the interpretation of (69). In fact, SC examples like (23)- (24) are ambiguous: on one 869 reading, the adjectives are interpreted as coordinated (on that reading, (23) is interpreted 870 as 'she is selling red and white skirts, and she is selling red and white coats'). This is not 871 the case on the other reading, which we have been focusing on above, on which (23) is 872 interpreted as 'she is selling red skirts and white coats'. The ambiguity is easily captured 873 if on the latter reading the higher coordination is indeed expletive since the adjectives are 874 not interpreted as coordinated on that reading. 875 That the higher coordination is not itself interpreted (see below for additional evi-876 dence to this effect) suggests that it is present for a formal reason. Sideward merger in 877 fact provides a straightforward formal reason for that. It seems safe to assume that it is 878 not possible to move different elements out of a single ConjP. Sideward merger is the 879 mechanism that makes it possible to get around that restriction. As noted above, side-880 ward merger was originally employed by Nunes (2004) to get around islandhood/locality 881 effects: sideward merger out of a context that would induce a locality/islandhood effect 882 voids that effect. Sideward merger is then also what makes it possible to get around the 883 restriction on moving more than one element out of a single coordination. The relevant 884 elements are sideward merged before they move out of the ConjP. The derivationally 885 formed coordination is then introduced into the structure. But crucially, the relevant 886 constructions then never involve regular movement of different elements out of a single 887 coordination (despite appearances). Sideward merger is then needed to get around the 888 restriction in question, and derivational coordination formation is exactly what provides 889 the needed sideward merger mechanism. 890 Notice that a single element can move out of a ConjP-this is in fact what happens 891 with regular ATB. This is not surprising. ConjP is traditionally considered to be an island, 892 this is in fact what the ban on extraction out of coordinations implies. In the phasal 893 system, it is then natural to assume that ConjP is a phase, given that phases have a 894 potential for inducing locality violations (Bošković 2017 and Oda in press in fact propose 895 that ConjP is a phase). 28 There are a number of analyses of regular ATB where movement 896 takes place only out of the first conjunct. This is e.g the case with the often assumed null 897 Operator movement analysis (see e.g. Munn 1992Munn , 1993, on which a null Op moves 898 within non-initial conjuncts but there is no movement out of these conjuncts: movement 899 takes place only out of the initial conjunct. The same holds for Nunes's (2004) sideward 900 movement analysis, where XP participating in an ATB construction is merged in its the-901 ta-position in the second conjunct, then re-merged in the theta-position in the first con-902 junct, undergoing movement only from that conjunct. 29 Under both of these analysis of 903 regular ATB, upward movement out of ConjP takes place only out of the initial conjunct. 904 In (3), it takes place from SpecDP of the first conjunct, where who is located prior to the 905 extraction out of ConjP. Under Chomsky's (2000Chomsky's ( , 2001 approach to the PIC, who at the 906 edge of the first conjunct is actually located at the edge of ConjP. 30 If two different ele-907 ments were to undergo movement out of the same ConjP, the second one could not get to 908 the edge of ConjP (the element at the edge of the second conjunct is not at the edge of the 909 ConjP). Sideward merger into another ConjP is then what enables this element to get out 910 of the problematic ConjP without undergoing actual movement out of it. 31 To summarize, under the suggestion made here the derivational ConjP formation 912 provides a formal mechanism that makes it possible for more than one element to get 913 dislocated from a single ConjP. This also deduces the restriction established above that 914 different elements that move out of a single ConjP must themselves get coordinated. It 915 also captures the semantic explitiveness of late coordination formation, given that it is 916 present for a strictly formal reason. 917 The semantic explitiveness of late coordination formation enable us to capture an-918 other property of late coordination formation constructions. And is not the only coordi-919 nator in English. What is interesting is that even when a different coordinator is used in 920 the lower, semantically contenful coordination, and is used in the higher, semantically 921 expletive coordination. Thus, while the literature discusses only cases with the coordi-922 nator and, it is possible to use the disjunctive or in the lower position of the constructions 923 under consideration; still, only and is possible in the higher, non-interpreted coordina-924 tion, or is not possible, as shown by (85)  What is essentially going on here is that the most neutral coordinator is used in the der-931 ivationally-formed coordination, even if a different element is used in the lower position 932 (note that this shows that we are not simply dealing here with coordinator copying), 933 which is not surprising if the higher coordination is not interpreted (i.e. if it is indeed 934 semantically expletive). 935 Also relevant here are the SC examples in (86)-(87) (noted by Ksenia Zanon, p.c., for 936 Russian). Repetition of the coordinator i 'and' in SC brings in an additional meaning, as 937 indicated by the rough translation of (86). Importantly, this kind of coordination cannot 938 be used in derivationally-formed coordination, as shown by (87). 32 Note also that (87) (88) indicate that what is present in 961 the θ-position of the relevant conjuncts (given the predicate-internal subject hypothesis) 962 is not the you and me ConjP: only you is present in the θ-position of the first conjunct and 963 only me is present in the θ-position of the second conjunct, given that each conjunct 964 agrees separately in (88), in contrast to (89). These examples thus confirm that elements 965 involved in distributed extraction coordinations start the derivation separately, as ex-966 pected given the interpretation of such constructions. (For additional arguments to this 967 effect based on binding, see (100), (101), and (103) below; the data in question quite con-968 clusively show that the conjuncts start the derivation separately.) 969 970 (88) He wants you and me to respectively go out of your mind and (go) out of my mind. 971 (89) cf. You and I are going out of our/*my/*your mind(s). (Postal 1998:161) 972 973 In (88)  This means that late coordination must be formed before subject-verb agreement is de-980 termined here. In light of this I will use such constructions as a diagnostic for determin-981 ing when exactly the derivationally-formed coordination is inserted into the structure. I 982 will consider the constructions discussed in this section under Chomsky (2000Chomsky ( , 2001 983 approach to agreement, where agreement is established through the Agree relation 984 holding between a probe and a goal, leaving it to the reader to verify that the conclusions 985 reached below can also be maintained under e.g. Chomsky (1995) approach, where 986 agreement is established in a Spec-Head relation (though with somewhat different deri-987 vations). Under the Agree analysis, when the relevant agreement relation is established 988 ConjP must be located lower than T, so that T can probe it (which means T must 989 c-command it).

990
Consider (90) in light of this. The relevant part of (90) can be derived as in (92) (only 991 the relevant elements are shown in the structures below): we have a vP&vP coordination 992 in the lower position, with the subjects still not being conjoined with each other at this 993 point. Another ConjP (what I have referred to above as late/derivationally-formed ConjP) 994 is then formed through sideward movement (92b). Given that this ConjP must be higher 995 than the θ-positions of the relevant elements, as discussed above (cf. (88)), and that it 996 must be below T so that T can probe it, there must then be a phrase between vP and TP, 997 with the late-formed ConjP (see (92b)) introduced into the Spec of this phrase (see (92c); I 998 leave open the identity of this phrase, referring to it as XP below). T then probes the 999 late-formed ConjP (92d), before the latter moves to SpecTP (92e Additional structure then needs to be present between T and vP so that the higher ConjP 1010 can be inserted into the structure outside of the lower ConjP but still below T (92c). This is 1011 straightforward in examples like (90), involving an auxiliary. It also needs to be the case 1012 in examples like (93). 34 I take this not to be an issue, given that many authors have any-1013 way argued for additional structure between TP and vP even for examples like (93) (see 1014Bošković 2015, Cinque 1999, Collins 2005, Merchant 2013, de Swart 1998, Ramchand and 1015 Svenonious 2013, Tenny 1992, among others). What is particularly interesting about this example is the discrepancy between agreement 1025 and interpretation within the conjuncts: what is interpreted in the relevant θ-position of 1026 the first conjunct is John, and what is interpreted in the relevant θ-position of the second 1027 conjunct is Mary. Yet, the agreement within the conjuncts is with John and Mary. 36 Let us 1028 see how this mismatch can be captured. 1029 First, the lower coordination here must be on a higher level than in (90)-it cannot 1030 be a vP&vP coordination given that the auxiliary is present inside each conjunct. The 1031 auxiliary is plural (although what is interpreted as the subject of each conjunct is singu-1032 lar), which means that the auxiliary agrees with the late-formed ConjP. So, what has to 1033 happen here is that the auxiliary agrees with the late-formed ConjP, just as in (90), but the 1034 auxiliary must be within the lower ConjP, in contrast to (90). 1035 Note first that, quite independently of the issues under consideration here, there are 1036 two ways of analyzing such constructions, as noted in Bošković (2020a). If only phrases 1037 can be coordinated, the subject and the auxiliaries cannot be located in the same phrase 1038 here, given that the subject is outside of the coordination and the auxiliaries are inside of 1039 the coordination-such examples would then provide evidence for a return to split Infl. 1040 (94) would then involve TP&TP coordination, with the subject undergoing 1041 ATB-movement out of each conjunct to the Spec of a higher projection, which for ease of 1042 exposition I refer to is as YP. 37 Alternatively, if traditional bar-level coordination is al-1043 lowed, (94) can involve T'-coordination, with the subject undergoing ATB-movement 1044 from each conjunct to SpecTP. The choice between the two analyses is immaterial here, I 1045 will adopt the former for ease of exposition (the reader should bear in mind though that 1046 both analyses are compatible with the discussion below). 1047 Consider then (94). The derivation will proceed similarly to (90), as shown in 1048 (95)-(98): John and Mary are inserted in their θ-positions, i.e. the positions where they are 1049 interpreted, separately (95). Late coordination is then formed (96), and inserted into 1050 SpecXP of each conjunct (97). Since XP is located lower than the auxiliary, each auxiliary 1051 will probe this ConjP, resulting in plural agreement on the auxiliary. The late-formed 1052 ConjP then undergoes ATB movement out of the coordination ( (98) The example, which shows a mismatch between agreement and interpretation in the 1063 second conjunct, can then be accounted for.  The dual behavior with respect to binding is easily captured under the current analysis: 1084 when the anaphor is located low in the structure, namely below the position in which the 1085 relevant elements are base-generated (i.e. inserted prior to derivational coordination 1086 formation), the relevant elements function as binders separately (this is the case with 1087 (100)). On the other hand, when the anaphor is inserted high in the structure, where only 1088 the derivationally-formed coordination is higher than the anaphor, the relevant elements 1089 function as the binder together (this is the case with (99)). 1090 Furthermore, (101) (due to Steven Franks, p.c.) confirms that the conjuncts must 1091 start separately. In (101), they induce a blocking effect for binding separately, as the 1092 simplified structure in (102)  This is a rather rare mismatch, which can help tease apart different approaches to sub-1120 ject-oriented anaphors. What functions as the binder of the subject-oriented anaphors in 1121 (103) The agreement/semantics mismatch constructions also have ramifications for an 1132 important question, where is agreement (i.e. what is its locus). It is standardly assumed 1133 that although agreement surfaces on the verb, it is not actually on the verb, i.e. its source 1134 is somewhere else (but see Lasnik 1995a for a proposal that in V-raising languages, the 1135 agreement is actually on the verb). It is certainly not a priori clear what the source of 1136 agreement is, whether the verb has it to start with (the verb is in fact where the agreement 1137 surfaces), or it is somewhere higher up (like on T/I). Semantics/agreement mismatches 1138 discussed above provide evidence for the latter (including for V-raising languages). If the 1139 source of agreement is the verb itself, not a higher head like T/I, given the standard as-1140 sumptions regarding the locality of agreement and the VP-internal subject hypothesis, 1141 we would expect the verb to show singular agreement in (105) (cf. (105b)). What is going 1142 on here is that the subject of the verbal projection and the subject of the clause are dif-1143 ferent elements-the coordination is the subject of the latter but not the former. That the 1144 agreement here is with the coordination can then be taken to provide evidence that the 1145 source/locus of agreement is not its host, the verb. The same point can be made with SC, 1146 where the verb raises out of vP (see Bošković 2001 andStjepanović 1999); in fact both the 1147 auxiliary and the participle show plural agreement in (106a), indicating that even when it 1148 comes to the agreement that shows up on the participle, the verb is not the source/locus 1149 of the agreement. The system developed above may also enable us to account for some rather puzzling 1160 constructions noted by Goodall (1987). Consider (107)-(108), focusing on the former ex-1161 ample ( (108) is a result of a familiar ordering effect with late coordination constructions 1162 discussed briefly above, and in more detail in the appendix). ConjP1 is then inserted into SpecXP from (92) and ConjP2 undergoes right node raising 1174 (more on right node raising below). Agreement in this kind of double late coordination 1175 constructions works as in the constructions discussed above, as illustrated by (111), and 1176 can be accounted for in the same way. There is, however, one wrinkle raised by such constructions, which is that only one verb 1181 is pronounced. I suggest that there actually is coordination of the vPs here. The deriva-1182 tion discussed above (cf. (109)-(110)) would lead to John and Mary saw and saw himself and 1183 herself. It is possible that a haplology motivated PF deletion takes place here, deleting and 1184 saw. There is a potential alternative. A number of authors have argued that verbs in Eng-1185 lish undergo short V-movement (see e.g. Johnson 1999, Lasnik 1995b, Bošković 1997; such 1186 analyses often assume overt object shift in English, which would not affect anything 1187 given the discussion below). Under this analysis, and still assuming that there is vP co-1188 ordination here, the verb would undergo across-the-board head movement out of the vPs 1189 in (109). The only thing that would remain in the coordinated vPs would then be the co-1190 ordinator itself. I suggest that in such a case, where independent movement operations 1191 move everything out of a coordinated phrase but the coordinator, the coordinator itself is 1192 deleted. 1193 1194 7.5. Right node raising and tough-constructions 1195 In the examples discussed above, the late-formed coordination undergoes 1196 wh-movement (cf. (6)) or A-movement to SpecIP (cf. (90), whose derivation is given in 1197 (92)). There are other movement operations that the late-formed coordination can un-1198 dergo, like right node raising (112) or tough-movement ((113); additionally, SC and Jap-1199 anese examples discussed above involve scrambling). 1200 1201 (112) John sold, and Mary bought, gold rings and raw diamonds from South Africa res-1202 pectively. (Abels 2004) (113) George and Martha are respectively easy for me to fool and hard for anyone to take 1204 advantage of. (McCawley 1998:294) 1205 While exploring the issue in detail would take us beyond the scope of this paper, I will 1206 briefly note here that distributed coordination constructions may have some implications 1207 for the proper analyses of right node raising and tough-constructions. Thus, (113) seems 1208 to be difficult to handle under the null Op-movement analysis of tough-constructions, 1209 where what undergoes movement is a null Op that is licensed by a co-indexed 1210 c-commanding element, as in Johni is Opi tough to please ti. In (113), it is not clear that there 1211 would ever be a c-command relationship between George and Martha and the null oper-1212 ators that each of these DPs would need to license (for discussion and comparison of 1213 different analyses of tough-constructions, see Bošković 2020b).

1214
Regarding right node raising, constructions like (112) raise a serious problem for 1215 most accounts of right node raising (e.g. the ATB rightward movement or the PF deletion 1216 one; see Abels 1994), since gold rings and raw diamonds from South Africa would be gene-1217 rated in the object position of each verb, which would give us wrong interpretation (the 1218 whole thing should then be interpreted as the object of each verb). The late coordination 1219 formation analysis straightforwardly resolves the issue since under this analysis the 1220 coordination is not present in the base: what is present in the base (i.e. the relevant the-1221 ta-positions) is simply gold rings in the first conjunct and raw diamonds from South Africa in 1222 the second conjunct. The coordination in question is formed derivationally (the relevant 1223 elements are not interpeted as coordinated, as discussed above). Under the rightward 1224 movement analysis, the derivationally-formed coordination can then be inserted in the 1225 right-node raised position. 39 1226

1227
Returning to the derivation of non-ATB ATB constructions, we have seen above that 1228 the coordination that participates in agreement is formed during the derivation. Looking 1229 at the structures in (92) and (98), we can see that the late-formed ConjP is inserted into the 1230 Spec of the first projection (not counting the lower ConjP in determining the first projec-1231 tion) above the position where the relevant elements are interprete. 40 Such examples may 1232 then help us determine the timing of derivationally-formed coordination insertion into 1233 the structure (assuming that they are illustrations of a broader pattern): based on such 1234 examples I then tentatively conclude that the derivationally-formed ConjP is inserted 1235 into the first projection above the position where the relevant elements are interpret-1236 ed-(this can even be within the original ConjP, in which case the derivationally formed 1237 ConjP is inserted in both conjuncts, undergoing regular ATB out of the lower ConjP). The 1238 precise locality condition-"the first projection"-in the above statement may end up 1239 being revised upon future scrutiny of distributed coordinations 41 -what is important 1240 here is that the derivationally-formed ConjP is inserted very close to the positions where 1241 the relevant elements (i.e. its conjuncts) are interpreted (for the reason why this is the 1242 case, which ties the issue to the ordering of elements participating in late-coordination 1243 formation, see Appendix). 1244 The SC construction discussed in section 3 can help us become more precise re-1245 garding the derivation of non-ATB ATB constructions, the reason for that being that with 1246 left-branch extraction (LBE), which is employed in the SC construction in question, it is 1247 possible to introduce a locality/islandhood effect very close to the base-generation posi-1248 tion of the relevant elements. But before we discuss that, one point needs to be empha-1249 sized. 1250 In any language I am aware of, only mobile elements can participate in ATB 1251 non-ATB constructions. The SC construction discussed in section 3 involves LBE, which 1252 is not possible in English, but is possible in SC. While non-ATB ATB involving LBE is 1253 possible in SC, as we have seen above, it is not possible in English (114a) The effect in question is actually also illustrated with English (11) As noted above, in contrast to the prepositional double object constructions (115c), the 1269 indirect object in DP DP double object constructions cannot undergo wh-movement 1270 (115d). 43 It also cannot participate in distributive extraction coordinations, as shown by 1271 (115b), which contrasts with (115a). e. *Whose sake did Ernie do that for? 1280 f. For whose sake did Ernie do that? (Postal 1998:127) 1281 1282 Informally, we can consider the PPs in (116) to be barriers, which renders the 1283 P-complements in (116) immobile (note that the whole PP can move). Importantly, the 1284 same effect is found with the distributive extraction coordination in (117) The mobility requirement (which is essentially a locality-of-movement requirement) is 1293 surprising from the point of view of Zhang's sideward merger analysis. Sideward 1294 movement/merger was originally employed by Nunes to get around islandhood/locality 1295 effects (provided it takes place while the island is an independent root phrase, which is 1296 the derivation we are focusing on): sideward movement/merger out of a context that 1297 would induce a locality/islandhood effect voids that effect. Whatever locality/islandhood 1298 effect is involved in (114) The parasitic gap constructions discussed above also require this movement. Con-1305 sider again (119). 1306 1307 (119) [Which secretary]1 and [which programmer]2 did Jerome respectively fire t1 after 1308 finding t1 drunk and hire t2 after finding t2 sober? (Postal 1998: 136) 1309 1310 As noted above, a parasitic gap is licensed by a moved wh-phrase that c-commands the 1311 parasitic gap. In accounts like Nissenbaum (2000) and Nunes (2004), it is not necessary 1312 for the wh-phrase to move to SpecCP to license a parasitic gap; movement to a lower 1313 position can do it. In fact, under Nissenbaum's account the wh-adjuncts in (119) are ad-1314 joined to their vPs, and the wh-phrases crucially need to move to adjoin to these vPs to 1315 license the parasitic gaps. This is the crucial step in parasitic gap licensing in this account. 1316 What is important for our purposes is that the wh-phrases need to undergo regular 1317 movement to license parasitic gaps within their conjuncts before undergoing sideward 1318 merger into ConjP in (119) The SC construction from section 3 can help us pinpoint the timing of regular and 1323 sideward movement. What is relevant here is that, in contrast to regular LBE as in ex-1324 amples like (22), what is in the literature referred to as deep LBE, illustrated by (121) For an account of the ungrammaticality of (121), the reader is referred to Bošković (2013a) 1337 and Corver (1992). The precise reasons for the grammaticality of (121) need not concern 1338 us here. What matters is that, as Corver (1992) and Bošković (2013a) show, the problem 1339 here arises with movement from NP1 to NP2 (there is no issue with movement out of 1340 NP1 per se (i.e. when the relevant NP is not dominated by another NP), otherwise even 1341 (120) would be unacceptable). 1342 Bearing this in mind, the following data shed a crucial light on the timing of regular 1343 and sideward movement involved in ATB non-ATB. There is a contrast between (122) and (123), which is particularly telling in light of the 1360 deep LBE effect from (121). The deep LBE effect is apparently still present in (123), but is 1361 voided in (122). What this means is that there is regular movement into NP1 in (123) (122)) given 1376 that, as Bošković (2013a) discusses, regular movement from the edge of the complement 1377 of N into the NP itself causes a locality violation in SC. Recall that, as discussed above, 1378 late-formed ConjP can be inserted into a phrase right above the original ConjP, which in 1379 this case is NP1. Since there is no regular movement from one NP domain into another 1380 there is then no locality violation. 1381 The relevant derivations are mapped out below: simplifying what exactly happens 1382 here, I will simply assume that the complement of a noun in this context is a barrier (the 1383 exact situation is more complicated (see Bošković 2013a, Corver 1992 and discussion be-1384 low), but this suffices for our purposes-the relevant phrase (i.e. the complement of the 1385 noun) is given in shadow bold red below). In (125), which corresponds to (123), regular 1386 movement (shown as movement of α) crosses a barrier, which induces a locality effect (in 1387 fact, the same effect as in (121)). In (124), which corresponds to (122), regular movement 1388 does not cross a barrier-only sideward movement crosses a barrier, but sideward 1389 movement voids locality effects (crossing here is metaphorical, sideward movement 1390 voids islandhood because it actually does not involve crossing of the island boundary: α 1391 is merged with β (forming a ConjP) in a separate derivational space, and then inserted 1392 into the position shown in (124) The contrast between (122) and (123)  The contrast between (123) and (127) quite strongly confirms the conclusions reached 1420 above regarding the contrast between (122) and (123) All of this confirms the movability requirement on elements that participate in distri-1447 buted extraction coordination: the relevant elements must be mobile, which indicates that 1448 they undergo regular movement before sideward merger. 1449 Putting everything that we have seen above together, we can map out the derivation 1450 of non-ATB ATB constructions more generally. It is apparently not possible for the rele-1451 vant elements to undergo sideward movement into late formed ConjP directly from the 1452 positions where they are interpreted. They have to undergo regular movement from that 1453 position, after which they can undergo sideward movement into the newly-formed 1454 ConjP. If a locality effect can be created right at the base-generated position, sideward 1455 movement will then not be able to obviate it (it would be taking place too late); but if the 1456 locality effect is created slightly higher than the base-generated position so that there is 1457 room for regular movement to take place before the locality effect kicks in, the locality 1458 effect gets obviated through sideward movement. Any locality effect higher up, i.e. 1459 higher than the point of insertion of the late-formed ConjP, which we have seen is still 1460 inserted close to the positions in which the relevant elements are base-generated, will still 1461 be in effect, due to the movement of the late-formed ConjP itself. All of this is mapped out 1462 in (132). (The brackets where a locality effect could in principle pop up due to regular, not 1463 sideward, movement crossing it are given in shadow bold red. For ease of exposition, I 1464 use a trace for the movement that precedes formation of the derivationally-formed ConjP, 1465 and a struck-out copy for the movement of the derivationally-formed ConjP itself. Two 1466 phrases are given between the final landing site and the original position of the 1467 movement of the derivationally-formed ConjP merely to indicate that this movement is 1468 generally longer than the movement that α alone undergoes, which is generally very 1469 short.) Regarding locality effects seen above, the locality effect in (123) arises due to the crossing 1474 of the redded XP between αi and ti in (132), which means with movement of the element 1475 that will later participate in late coordination ((122) crucially differs from (123) in that that 1476 step of movement in (122) does not cross a barrier; it essentially takes place below XP in 1477 (132) due to the presence of additional structure in (122)). The locality effect in (82) (90) and (94). The same holds for binding. If it takes place below YP, the individual 1486 conjuncts will function as binders. This is the case in (100), (102), (103), and (107). If it 1487 takes place above YP, the whole coordination will function as the binder, which is the 1488 case in (99). 48 1489 Importantly, the facts discussed above indicate that islandhood/locality effects are 1490 selectively present with non-ATB ATB constructions. In most cases they are present, but 1491 in some cases they are voided. This could not be captured if we were to simply adopt 1492 122) is thus problematic for this analysis). 50 On the other hand, the 1505 selective presence of locality effects can be captured on an analysis which essentially 1506 combines the two accounts just noted, on which there is both regular movement and 1507 sideward movement involved in the derivation of non-ATB ATB constructions. We have 1508 seen that such an analysis can also capture agreement and binding effects found with non 1509 ATB ATB constructions as well as the mobility requirement on the elements involved in 1510 non-ATB ATB and the ability of these elements to license parasitic gaps on their own. 1511

1512
This paper has provided additional evidence that it is possible to move different 1513 elements from conjuncts involved in the same coordination and that such constructions 1514 involve coordination formation in a non-base generated position, i.e. after movement (cf. 1515Zhang 2010. It was shown that such constructions are also subject to the ATB require-1516 ment: although different elements are moving out of conjuncts movement still must take 1517 place out of each conjunct. This means that the traditional ATB requirement needs to be 1518 reformulated: it is not the case that the moving element must move out of each conjunct 1519 but simply that movement must take place out of each conjunct. It can be the same ele-1520 ment that is moving out of each conjunct or different elements; the ATB requirement is 1521 satisfied as long as there is movement out of each conjunct (furthermore, the ATB re-1522 quirement does not hold across ConjPs). Traditional ATB, where the same element moves 1523 out of more than one conjunct, and what I have referred to as non-ATB ATB, where dif-1524 ferent elements are moving out of the conjuncts, can in fact be mixed under extraction out 1525 of the same coordination, as expected if all that is needed is that there is movement out of 1526 each conjunct. Furthermore, mixed non-ATB ATB cases have the same ordering re-1527 strictions (regarding the order of the conjuncts) as pure non-ATB ATB cases. 1528 We have also seen that there is a restriction on non-ATB ATB, where different ele-1529 ments are moving from different conjuncts, in particular, non-ATB ATB requires coor-1530 dination formation in the moved position. Additionally, head-movement cannot be in-1531 volved in non-ATB ATB, which in fact follows from the coordina-1532 tion-in-the-moved-position requirement if head coordination is disallowed, as Kayne 1533 (1994) argued. 1534 I have also discussed the precise timing of derivational coordination formation, 1535 concluding that the late-formed coordination is inserted into the structure very close to 1536 the phrase where the relevant elements are interpreted (under sideward merger analysis 1537 of distributed coordination; see Appendix for the reason for this), not in the final position 1538 of the relevant elements (as in Zhang 2010). The relevant elements, however, first need to 1539 undergo regular movement from the positions where they are interpreted: they cannot 1540 undergo sideward merger into the derivationally-formed ConjP straight from the posi-1541 tions where they are interpreted: The derivation of non-ATB ATB constructions then in-1542 volves both regular and sideward movement. 1543 The derivationally-formed coordination is semantically expletive in that elements 1544 participating in such coordination are not interpreted as coordinated; as a result only the 1545 most neutral coordinator is used in such coordination (even when a different element is 1546 used in the lower position). The coordination in question has also been shown to cause 1547 agreement/semantics mismatches, which arise because a coordination that is not seman-1548 tically interpreted participates in determining agreement. The presence of the coordina-1549 tor in question was shown to be motivated by formal reasons (this is in fact what enables 1550 extraction of different elements from a single coordination). 1551 Finally, the analyses and the coordination data discussed in this paper have been 1552 shown to have consequences for determining the proper analysis of a number of mecha-1553 nisms and constructions, in particular subject-oriented anaphors, right node raising, 1554 tough-constructions, the source of agreement, and the more general issue of clausal 1555 structure. Regarding subject-oriented anaphors, non-ATB ATB constructions were shown 1556 to provide evidence against approaches where the element located in SpecTP and/or the 1557 element that undergoes Agree with T function as the binder for subject-oriented 1558 anaphors based on constructions where different elements fill SpecvP and SpecTP (in 1559 such cases, the latter undergoes agreement with T but the former binds subject-oriented 1560 anaphors).  Above we have seen a number of cases involving coordination formation in the 1568 moved position, i.e. after movement. All these cases also involve coordination in the 1569 lower position, i.e. they involve extraction out of a coordination. A question arises if late 1570 coordination formation is possible without coordination in the lower position, i.e. if the 1571 relevant movements do not take place out of a ConjP. Citko and Gračanin-Yuksek (2013) 1572 argue that it is. They consider constructions like (133) and argue that crosslinguistically 1573 they can involve either larger (i.e. clausal) coordination where only the wh-phrase is re-1574 alized in the first conjunct, which they argue is the case with English (133), or coordina-1575 tion of wh-phrases, which they argue is the case with Bulgarian (134). The acceptability of (134) then indicates that we are dealing here with wh&wh, rather 1592 than clausal coordination. 52 1593 Citko and Gračanin-Yuksek also observe that coordinations like (134) are not pos-1594 sible with wh-phrases in situ. In fact, it is quite generally not possible to coordinate a 1595 subject and an object of the same clause, which means that (134) cannot involve base co-1596 ordination that would then undergo movement. In other words, the coordination here 1597 can only be formed after movement.
Note also that, like distributed coordinations, wh&wh coordinations are sensitive to 1599 islandhood, as shown by Bulgarian (136), involving an adjunct island (note that Bulgar-1600 ian does not show Comp-trace effects).  (134) is correct, such 1607 examples provide evidence that late coordination formation is not limited to construc-1608 tions involving movement out of a coordination. Notice also that, like the derivational-1609 ly-formed coordination discussed in the main text, the coordination discussed in the 1610 appendix is also not semantically interpreted. Thus, the interpretation of (134) is simply 1611 'who bought what', there is no coordination of the wh-phrases in the interpretation of 1612 this construction. 53 Not being interpreted, i.e. being semantically expletive in the relevant 1613 sense, can then be taken to be the hallmark of derivationally formed coordination (re-1614 gardless of whether late coordination formation takes place out of a coordination or not). 1615 In other words, coordination can be formed during the derivation but if that happens it 1616 has no semantic import-only base-coordination is interpreted. 1617 Notice furthermore that given that the structure instantiated by Bulgarian (134) is 1618 apparently not allowed in English, the availability of non-distributed wh&wh coordina-1619 tions, which, if Citko and Gračanin-Yuksek (2013) are right in their treatment of such 1620 coordinations involve late coordination formation without movement out of a coordina-1621 tion, should not be tied to the availability of constructions like (6) (which involve coor-1622 dination in the lower position, i.e. late coordination formation out of a coordination), in a 1623 sense that the availability of the latter would imply the availability of the former (though 1624 SC happens to allow both, see (44) and the discussion below). 1625 It should, however, be noted that the former (i.e. non-distributed late coordination) 1626 is not possible with the non-wh arguments in the SC example in (137) (the fronting in 1627 (137a) can in principle involve topicalization, focalization, or scrambling (see Bošković 1628 2004a), the construction is apparently ruled out regardless of which of these options is 1629 taken, in contrast to (44) Given that there clearly must be rather strong additional restrictions on non-distributed 1646 wh&wh coordinations, which are not operative with Postal-style distributed coordina-1647 tions, it is not out of question that the unavailability of the former in English (in contrast 1648 to the availability of the latter) is due to those additional restrictions, i.e. that we are not 1649 dealing with a deeper point of variation in this case, where English would allow late co-1650 ordination formation only out of another coordination. (Recall that SC allows it regard-1651 less of whether late coordination formation takes place out of a coordination or not.) 1652 Rather, more construction-specific issues could be involved. 1653 Citko and Gračanin-Yuksek (2013) in fact tie the availability of wh&wh coordination 1654 to the availability of multiple wh-fronting: since English does not have multiple 1655 wh-fronting it cannot then have the structure in question. However, it is not clear why 1656 multiple wh-fronting should be relevant here. Under Citko and Gračanin-Yuksek's 1657 analysis, the interrogative C in (134) has only one Spec, which is filled by ConjP. 1658 There are additional reasons why the availability of (134) should not be tied to 1659 multiple wh-fronting. Thus, as another argument for the wh&wh (as opposed to clausal) 1660 coordination account of Bulgarian (134), Citko and Gračanin-Yuksek (2013) observe that 1661 such constructions show Superiority effects, i.e. strict ordering of coordinated 1662 wh-phrases. They argue that this would not be expected if (139) involves coordination of 1663 two clauses, where each clause has only one wh-phrase, which undergoes movement.  Citko and Gračanin-Yuksek (2013) argue that what is relevant here is that Bulgarian is a 1670 multiple wh-fronting language. Multiple wh-fronting languages differ regarding 1671 whether they show superiority effects under multiple wh-fronting (see for example 1672Rudin 1988, Bošković 2002. Bulgarian does show such effects (see (140)), just as it does 1673 with wh&wh coordinations (see (139)). In light of this, Citko and Gračanin-Yuksek (2013) tie the possibility of wh&wh coordi-1681 nations to multiple wh-fronting. The correlation is, however, rather difficult to maintain. 1682 Under the standard account the superiority effect in (140) arises as a result of the inter-1683 rogative C attracting two wh-phrases, where these wh-phrases undergo separate 1684 wh-movements, occupying separate CP Specs. This is, however, not the case with (139) 1685 under Citko and Gračanin-Yuksek's (2013) analysis, where the coordination of 1686 wh-phrases (i.e. ConjP dominating the wh-phrases) is merged into SpecCP-there are no 1687 two separate wh-movements or two CP Specs in (140). Furthermore, Citko and 1688 Gračanin-Yuksek (2013) also observe that some speakers of Bulgarian do allow free or-1689 dering of the coordinated wh-phrases in (139). On the other hand, there is no speaker 1690 variation regarding superiority effects with multiple wh-fronting. 1691 Also relevant here is SC, which does not show matching in the ordering of 1692 wh-phrases in simple multiple wh-fronting constructions and wh&wh coordinations. As 1693 noted above, multiple wh-fronting languages differ regarding whether or not they show 1694 superiority effects in examples like (140). Thus, as discussed in Rudin (1988) and 1695 Bošković (2002) When there is nothing following the first wh-phrase there is a superiority effect, as shown 1724 by (142). 58 On the other hand, when the clitic follows the first wh-phrase, which clearly 1725 shows that in such cases the first conjunct is larger than the wh-phrase itself, there is no 1726 superiority effect (see (143)- (144)). These data indicate that when there is no additional 1727 material following the first wh-phrase we are indeed dealing with a wh&wh coordina-1728 tion. 59 These facts also indicate that there is no parallelism between Superiority effects in 1729 simple multiple wh-fronting constructions and wh&wh constructions, given the contrast 1730 between (141) and (142) (more precisely, (141b) and (142b)), i.e. the parallelism shown by 1731 Bulgarian (140) and (139) is accidental.

1732
In fact, there is reason to believe that whatever is going on with the ordering of 1733 wh-phrases in wh&wh coordinations is different from superiority effects with multiple 1734 wh-fronting. As discussed in Bošković (2002), quite generally when the superiority effect 1735 is found with multiple wh-fronting it holds only for the first and the second wh-phrase; 1736 beyond that the ordering of the wh-phrases is free. This is shown by Bulgarian (145): 1737 when only two objects undergo wh-movement, the indirect object must precede the di-1738 rect object, a superiority effect given that the former is higher than the latter prior to 1739 wh-movement. However, when a higher wh-phrase is present, the ordering of the indi-1740 rect and direct object is free (the nominative must be first in (145c-d) as well as (146c) and 1741 (147) below). The same point is illustrated by SC (146), where the superiority effect also 1742 holds only for the first and the second wh-phrase (see Bošković 2002 for discussion of the 1743 superiority effect in (146a-b) As noted above, Citko and Gračanin-Yuksek (2013) take the superiority parallelism be-1804 tween Bulgarian (140) and (139) to indicate that the availability of multiple wh-fronting 1805 underlines the availability of wh&wh coordinations (which, recall, involve late coordi-1806 nation formation). The fact that, as shown above, wh&wh coordinations do not track 1807 multiple wh-fronting with respect to Superiority suggests that the two should be di-1808 vorced. There should then be no connection between multiple wh-fronting and the pos-1809 sibility of late coordination formation (which underlines the possibility of Postal-style 1810 distributed coordination-this is desirable given the possibility of the latter in English).

1811
At any rate, there is crosslinguistic variation regarding non-distributed wh&wh 1812 coordinations, whose availability should not be tied to either the availability of multiple 1813 wh-fronting or Postal-style distributed coordination in the language. 1814 While the issues discussed in this appendix merit a much more extensive scrutiny 1815 than they could be given in this appendix, whose scope is rather limited, what we are 1816 seeing here is that languages differ with respect to how they behave regarding the rele-1817 vant properties of coordinate constructions. The point of the above discussion was 1818 merely to outline some of the possible crosslinguistic variation in the relevant domain, as 1819 well as to highlight the need for more extensive crosslinguistic investigations of the rel-1820 evant properties of coordinations (recall that languages also differ regarding whether 1821 they allow conjunct extraction, see for example SC (77), which is unacceptable in Eng-1822 lish). Hopefully, such investigations will reveal correlations between the properties of 1823 coordination investigated in this paper and other properties, which should help deter-1824 mine in a more principled way the factors that are behind the phenomena (and the vari-1825 ation with respect to these phenomena) discussed in this paper. 1826 I will close this appendix with a note on the ordering effect. Postal-style distributed 1827 extractions pattern with wh&wh coordinations in the relevant respect. As discussed in 1828 section 3, there is an ordering effect with Postal-style distributed extractions-the order 1829 of the conjuncts within the newly formed ConjP must match the order of the conjuncts 1830 from which extraction takes place in the original ConjP-which in the cases involving 1831 three conjuncts holds for all conjuncts, as shown by SC (152) which gives the only possi-1832 bility for the distributed interpretation of the extracted adjectives. The same holds for 1833 English distributed coordinations, as shown by (153) If the ordering effect in the two constructions is to be captured in a unified manner, 1845 the ordering effect with Postal-style distributed coordinations cannot be due to a 1846 matching effect between two coordinations (as suggested briefly in section 3) since with 1847 wh&wh coordinations there is no lower ConjP. I would therefore like to suggest an al-1848 ternative. In particular, I suggest that derivational coordination formation needs to occur 1849 as early as possible. More precisely, the relevant element needs to merge with the 1850 non-base coordinator as soon as it is eligible for such merger. (142), repeated in (154), 1851 would then be derived as in (155)  The element that enters the structure first then has to undergo sideward merger wih the 1861 coordinator first, as shown in (154b) (i.e. what merges with and before who does). The re-1862 sult of this is that the order of elements in the derivationally formed ConjP will corres-1863 pond to the order of these elements prior to derivational coordination formation, with 1864 ordering imposed on all conjuncts, not just on one relevant element (as it is with Supe-1865 riority). 61

1866
All this also works for Which booki and which journalj did Sue [buy ti] and [read tj] re-1867 spectively; given the cycle/bottom up structure building, which journal has to merge with 1868 and, the head of the derivationally formed ConjP, before which book (since the conjunct 1869 that dominates it is integrated into the structure by merging with the base coordinator 1870 and first). 62

1871
The above suggestion also enables us to deduce the contiguity requirement on 1872 mixed ATB and non-ATB ATB cases. To illustrate it again, in (155)-(156) regular ATB can 1873 hold between contiguous conjuncts, it cannot hold between the first and the third con-1874 junct, skipping the second conjunct. 1875 1876 (156) ?How many cakes and how many letters did Mary bake, read, and mail respec-1877 tively? 1878 (157) *How many lettersi and how many cakesj did Mary read ti, bake tj, and address ti 1879 respectively? 1880 1881 Given the earliness requirement on late coordination formation, how many letters must 1882 merge both before and after how many cakes into the late formed ConjP in (156) (more 1883 precisely, it has to be sideward merged into it after and address…. is formed; there can be 1884 no regular ATB after that step, i.e. out of the late-formed coordination, since late-formed 1885 coordinations are opaque for merger out of them, see section 6)). The early sideward 1886 merger requirement (i.e. the requirement to form the derivationally formed ConjP as 1887 soon as possible) thus deduces the strict ordering requirement as well as the contiguity 1888 requirement on derivationally formed ConjPs. 1889 Recall now that the derivationally formed ConjP has to be inserted into the structure 1890 very close to the step of regular movement that the relevant elements need to undergo. 1891 This can be interpreted as indicating that the derivationally formed ConjP has to be in-1892 serted into the structure as early as possible. From this perspective this earliness re-1893 quirement can be looked at as part of a more general earliness requirement on deriva-1894 tional ConjP formation: such ConjPs must be formed as soon as possible and inserted into 1895 the structure as soon as possible.
1896 1897 Notes 1 (4) also involves a ban on extraction of conjuncts, which will not be examined in this work (the ban on extraction out of conjuncts and the ban on extraction of conjuncts have anyway ban argued to be independent conditions, see e.g. Grosu 1973, Postal 1998, Stjepanović 2014a, Oda 2017).
2 There are some differences across speakers regarding the most natural prosody of such constructions. The judgments given below reflect the most natural prosody for the speakers in question (not all speakers accept such coordinations in the first place).
accept (20)), which confirms that non-ATB ATB and regular ATB indeed behave in the same way with respect to the parallelism requirement in question. 7 These authors argue that constructions like (22) involve extraction of the AP out of the NP. There are two alternative analyses: remnant movement of the NP which contains only the AP (Franks and Progovac 1994;Abels 2003) and full NP movement with scattered deletion, where the NP is deleted in the highest copy and the AP in the lower copy (Fanselow and Ćavar 2002). There are a number of arguments in the literature for the left-branch extraction analysis, which is adoped here; see e.g. Bošković (2005), Stjepanović (2010Stjepanović ( , 2012; Talić (2013Talić ( , 2017, and Despić (2015). The reader is also referred to Bošković (2019) for discussion of the CSC regarding SC, where it is shown that (1) is operative in SC. 8 All the judgments below are given only for the distributive reading, indicated in the translations of (23)-(24) (and with traces when they are given in the structures below). 9 A referee observes that, as expected, (i), where there is no adjective at the edge of the conjunct that is not involved in ATB, is also unacceptable: (i) 10 What may matter here is the following: Chomsky (2013) proposes that the first conjunct determines the category of the whole coordination (which essentially means that the coordination itself does not inherently have it; note that Chomsky's proposal is stated somewhat differently, in terms of labeling), and a number of authors (e.g. Sag et al 1985, Takahashi 1994, Bošković 2019 have argued that the ATB requirement is related to the coordination-of-likes requirement (see Chomsky 1957, Schachter 1977, Williams 1978, Sag et al 1985, Bowers 1993, Beavers and Sag 2004, among others, on this requirement). An intuitive idea here is that when the first conjunct, which is supposed to determine the category of a coordination, is itself a coordination, the category of the higher coordination is undetermined-this then voids the ATB requirement, which is tied to category specification (this is what is relevant to the coordination-of-likes requirement). This makes a prediction, which is borne out: if the order of the conjuncts in (29) is switched, the category of the coordination will be determined since the first conjunct is not a coordination; this then activates the ATB requirement, ruling out ( The contrast between (33) and (39) in fact parallels the contrast between (ib) and (ic).
Another case of the maximize ATB effect may be provided by the contrast between (33) (49). For some speakers, under the most natural prosody the fronted adjectives bear focal stress and are followed by a pause. This causes an issue regarding clitic placement. There is variation across speakers whether under certain conditions a clitic can follow a sentence internal pause (see Bennett 1987, Percus 1993, Browne 1975, Schütze 1994, Bošković 2001. I ignore here speakers for whom there needs to be a pause following the fronted adjectives and who disallow clitic placement after such a pause (pronominal clitics are quite generally disallowed in that case, hence they are avoided below in this context). 16 On the relevant reading, onu is not a separate nominal in (51)-(52) (demonstratives can be separate nominals, as in I like this) but modifies kuću, just like staru does (there is only one nominal on this reading, 'that old house'). Below, where possible different gender will be used for the demonstrative and the adjective to control for this. 17 It is worth noting here that NP ellipsis that strands demonstratives and adjectives is also possible in SC, see Bošković (2013b).
One might try to treat (24) this way. The ungrammaticality of (53), however, provides evidence not only against the clausal ellipsis analysis, but also against the NP ellipsis analysis. 25 Such crosslinguistic differences underscore the need for more in depth crosslinguistic investigations of the distributed extraction coordination construction, one of the main goals of this paper in fact being to spur such investigations. Note that examples like (74) indicate that we cannot simply have allowing or dissalowing distributed extraction coordination as the point of crosslinguistic variation since the same language can allow it with some but not other movements. In fact, the variation can also concern the coordination from which it takes place; thus, Japanese allows it out of -to coordinations (see section 4) but disallows it out of sosite coordinations: In contrast to (55) 27 Sideward merger might be a more appropriate term (if movement is taken to involve a c-command relation between the relevant positions); at any rate I will use the terms interchangeably below. 28 This does not mean that phases in general are islands; phases have the potential to induce locality violations, which can then capture islandhood. 29 One issue these works were concerned with is the well-known impossibility of reconstruction into the second conjunct with regular ATB (see Munn 1992and Nunes 2004 for relevant discussion under the null operator movement and the sideward movement analysis respectively). 30 Chomsky's approach to the PIC differs in this respect from Hiraiwa (2005) and Bošković (2015), where the edge of the edge of phase XP is not at the edge of XP. I am adopting Chomsky's approach to the PIC in this respect. (Notice also that Bošković 2018 argues that movement to the edge of individual conjuncts is quite generally allowed).
It should be noted that it is beyond the scope of this paper to examine the nature of the ATB requirement itself. As noted in fn. 10, several works have argued that the ATB requirement actually follows from the coordination-of-likes requirement, namely Sag et al (1985), Takahashi (1994), and Bošković (2018). Under the implementation of this approach in Bošković (2019), extraction out of the first conjunct is allowed in certain cases even where there are no gaps in other conjuncts (which is in part due to the accessibility of the edge of the first conjunct, see Bošković 2019 for other factors involved). Any gap in a non-initial conjunct, on the other hand, obligatorily "activates" the ATB requirement, forcing the presence of a gap in each conjunct. 31 We will see below some cases where the relevant elements move to the edge of ConjP. Given that only the outmost edge of a phase is actually available for movement out of a phase, as argued extensively in Bošković (2016) (93). 35 Such examples were noted in Dougherty (1970), McCawley (1998), Postal (1998), and Zhang (2010). agreement in (94) (112), is the object in the first conjunct, with the object in the second conjunct staying in situ, which seems to leave no room for insertion of the late formed ConjP in (112). Bošković (2004b), however, shows that the element in the second conjunct can undergo rightward heavy NP shift within that conjunct. The late formed ConjP can then be inserted in that position.
Another case of distributed extraction coordination interacting with right node raising is provided by (i The partial structure in (ii) raises a question: how is the 'governor' interpreted in the object position of the first conjunct? The answer is provided by right node raising, which is possible in this context: (ii) John denounced and Martha was denounced by the governor. 40 Above I have briefly aluded to an alternative analysis on which agreement is established in a Spec-Head relation. While for reasons of space I did not discuss this analysis, this conclusion also extends to the Spec-Head agreement analysis (although this analysis comes with different structural assumptions, see fn. 33). 41 A small adjustment will in fact be made below to accommodate an additional step in the derivation of distributed coordinations discussed below. (I have assumed above that there is no vP in passives although there is some controversy regarding this issue. At any rate, what is important here is simply that the late-formed ConjP is inserted very close to the base positions of the relevant elements.) 42 In this respect, note that Slovenian speakers generally disallow regular adjectival LBE and they also disallow it with distributed extraction coordinations of the kind discussed for SC here, which confirms that the mobility of the relevant elements matters. 43 There is actually some speaker variation in this respect in British English, see Holmberg et al (2019). 44 The first-projection-locality-requirement on late-formed ConjP insertion discussed above should now be adjusted to take into consideration this short regular movement (it would be the first projection above the phrase where this short regular movement lands (or even within that phrase in some cases-I return to this issue below). At any rate, what is important is that the late-formed ConjP is inserted very close to the landing site of this movement.
It should be noted that there is one exception to the mobility requirement, which concerns right node raising. Right node raising is known to be able to affect elements that are otherwise immobile, see e.g. Bošković (2004b). The same holds when right node raising involves distributed extraction. It then seems that we are dealing here with a right node raising-specific issue, which is independent of the mobility requirement on distributed extractions discussed in the text: (i) a. John likes cheap, and Mary likes expensive, swimming suites.
b. John is asking when, and Mary is asking why, Peter is leaving.
(ii) a. John likes cheap, and Mary likes expensive, trousers and dresses respectively.
b. John is asking when, and Mary is asking why, Peter is leaving and Amy is late respectively. 45 Slightly more complicated are cases like (34), which involve a mixture of non-ATB ATB and traditional ATB.  Nunes's (2004) sideward merger analysis of traditional ATB. Under this analysis and the current analysis of distributed coordinations, (i) is derived as follows: "white" is merged with "coat"; "red" is merged with "shirt" and then undergoes sideward merger with "skirt" (this is regular ATB). Both "red" and "white" then move to the edge of the lower ConjP, after which derivational coordination formation takes place, with "red" and "white" undergoing sideward merger that forms what I have called late-formed ConjP (note that only derivational ConjP formation is subject to the not-directly-from-the-interpreted-position restriction). 46 Bošković (2013a) argues that in inherent case configurations, there is actually a PP-like projection between the two NPs, which means that the higher noun does not directly take NP as its complement in (126) (and (127) below), in contrast to (121) and (123) (see Bošković 2013a for discussion why this matters). 47 As noted by Sandra Stjepanović (p.c.), examples like (127) can also help us control for an interfering factor regarding (30). In (i), the adjective can modify both conjoined nouns, which means (ib) doesn't necessarily involve ATB movement of the adjective; it can also involve regular ATB of a single adjective with (ia) as its input (note this is not possible in three conjuct examples (on the true coordination non-list reading), as with the SC counterpart of white coat, red skirt, and dress). Given this, an issue arises: how do we know that (30) doesn't involve adjectival ATB from the first conjunct, instead of one adjective being adjoined to the ConjP, taking scope over both Ns?
please me self red skirts and dresses 'I like red skirts and (red) dresses.' Deep left-branch extraction with inherently Case-marked NPs controls for this interferring factor: due to the extra layer of embedding, AP adjunction to ConjP that conjoins the relevant nouns is not possible in (ii) (kongresu and parlementu are not conjoined), which means that the possibility of one adjective taking scope over, and modifying, two conjoined nouns prior to extraction is not an option here. (iii) must then involve regular ATB. this-year's is supported resistance congressDAT help parliamentDAT and threats presidentDAT 'He supported resistance to this year's congress, helping this year's parliament, and threats to this year's President.' When regular ATB dependency is blocked, as in (iv) (in (iii) all relevant elements, the adjective and the nouns, are masculine; in (iv) 'government' is feminine, blocking ATB), we get an ungrammatical construction.  (90) and (94) given above can be easily adjusted to accommodate the movement from the interpreted position prior to sideward merger into the late-formed ConjP. Regarding (94) (cf. the derivation in (95)-(98)), one possibility is that there is a projection between XP and VP, with the relevant elements moving to that projection within their respective conjuncts prior to sideward merger (i.e. prior to step (96)). There is, however, another possibility which does not require an additional projection: the individual conjuncts, John and Mary, undergo separate movements to Specifiers of XP in (97) prior to the late-formed ConjP insertion, with this ConjP inserted into the higher Specifier of XP (note that the Spec-Head Agreement analysis, briefly discussed in fn. 33, would also require a projection between VP and TP to accommodate the regular movement that precedes sideward merger into the late-formed ConjP). As for (90) (cf. the derivation in (92)), given that it is not possible for different elements to undergo movement out of a ConjP (so sideward merger is needed to obviate the locality effect, as discussed above), a dog and a rooster would move separately to the edge of ConjP if movement to the edge of ConjP is also allowed in English, or, if this is not possible in English, (90) would involve coordination on a slightly higher level (than vP in (92a)), with a dog and a rooster moving to the edge of the projection that undergoes coordination prior to undergoing sideward merger into the late-formed ConjP. Further research is needed to tease apart the options in question. 49 If there are more than two conjuncts, the higher ConjP can have multiple Specs, or there can be additional ConjP shells. 50 There are other issues that would arise on this analysis. Under this analysis, it is not clear how the relevant elements could trigger agreement together (cf. (90), (94), (105), (106), (111)) or function as binders together (cf. (99)) (assuming the ConjP in question is on top of the structure; we might actually then expect the coordination in question to be a strictly matrix phenomenon, which it is not (cf. Peter inquired which book and which magazine John bought and Mary sold respectively). Notice also that the relevant elements can be clefted together (cf. it is this book and that song that Mary wrote and Tom sang respectively). This indicates that they are a consitutent, which is not the case under the alternative analysis outlined in the text. 51 I will use the term (non-distributed) wh&wh coordination to refer to constructions which involve coordination of wh-phrases (not a larger constituent) and where the wh-phrases are not extracted out of a coordination. For ease of exposition, to differentiate such cases from examples like (6), where there is a lower ConjP, I will refer to the latter as (Postal-style) distributed coordinations.
52 Furthermore, Gračanin-Yuksek (2007) note that wh-DP external material can occur within the relevant ConjP in English, as shown by (i). Nothing of that sort is possible in Bulgarian. (Note that SC allows both obligatory arguments and DP-external material in the coordinations in question, as shown by (44)-(45), the reason for this being that SC allows both wh&wh and larger clausal coordination, see Citko and Gračanin-Yuksek 2013 (I am simplifying their discussion here) and the discussion below.) (i) What did Peter and why did Peter eat? (Gračanin-Yuksek 2007) 53 It would be interesting to see whether such coordination is possible in rare languages that disallow multiple questions (e.g. Italian, which disallows constructions like who bought what); whether or not such languages would allow late coordination formation of wh-phrases might shed light on the reason for the unacceptability of multiple questions in such languages. 54 I discuss restrictions on non-distributed late coordination in SC in work in progress, where I show that this kind of coordination, without coordination in a lower position, is actually not completely restricted to wh-phrases/wh-movement. 55 In SC, Postal-style wh-movement distributed coordinations are actually more restricted than in English. One issue could be that SC does not have a real counterpart of respectively that is used in such cases in English. What is happening in SC is that (possibly due to the lack of 'respectively' or the possibility of wh&wh coordinations), the non-distributed reading on which the coordination of wh-phrases undergoes ATB-movement from each object position is the only reading in the counterpart of (6) in SC. (i) [Koju knjigu i koji magazin]i je Jovan kupio ti i Ivan prodao ti?
which book and which magazine is Jovan bought and Ivan sold 'Which book and which magazine did John buy and Ivan sell?' However, when such non-distributed ATB-movement parse is not possible, as in (69), Postal-style distributed coordination is available with wh-phrases in SC. It is also marginally available in constructions more similar to (i) where the pragmatics increases the saliency of the distributed reading (with a different coordinator though; a is a clausal coordinator, which does not coordinate nominals).
how-many dishes and how-many letters is Maria made and Ivan wrote 'How many dishes and how many letters did Maria make and Ivan write?' 56 (142b) improves if there is a pause following the first wh-phrase, which is not necessary in (141b). For the relevance of this prosodic pattern, which I put aside in the text, see fn. 59. 57 There may be a null subject in the first conjunct and a null object in the second conjunct here. SC is a pro-drop language so the former is not surprising. On null objects in similar constructions, see Zanon (2015) and references therein (for other perspectives on these issues, see Citko andGračanin-Yuksek 2013, Gračanin-Yuksek 2007). 58 The same holds if the clitic follows the coordinated wh-phrases. who and what is bought 59 This in itself is quite interesting. We may be dealing here with an economy of representation effect (see Bošković 2011 and references therein): when both a smaller and a larger structure are in principle available for X, if there is no evidence for the larger structure X is analyzed in terms of the smaller structure. (We would not necessarily expect to find this effect in all languages of this sort since the effect would hold only in an all-else-being-equal scenario, which is not always the case; e.g. lexical properties of elements that are elided on the larger structure option could block the effect-see Bošković 1997.) There may, however, be another factor at work here. Recall that, as noted in fn. 56, (142b) improves if there is a pause following the first wh-phrase. It may then be that wh&wh and clausal coordination are associated with two distinct prosodic patterns in the cases where nothing intervenes between the coordinated wh-phrases, the prosodic pattern with a pause following the first wh-phrase reflecting clausal coordination structure. This prosodic pattern is forced by a superiority violation, which is not allowed in the wh&wh derivation. It should be noted that according to Citko and Gračanin-Yuksek, there is no superiority effect in wh&wh coordinations in Croatian.
The speakers I have consulted, all of which come from Bosnia, do show a superiority effect here (in fact all the data discussed above come from the Bosnian variety of what I have referred to as SC). It is not out of question that there is no real variation here, and that Citko and Gračanin-Yuksek were checking the prosodic pattern associated with clausal coordination. (This may also be behind what they report as speaker variation in the ordering effect in Bulgarian; while Citko and Gračanin-Yuksek do not give relevant Croatian data they do give superiority-violating examples from Russian. However, it turns out superiority violations in Russian are possible only under the prosodic pattern associated with clausal coordination). At any rate, while the issue under consideration is quite interesting, it goes beyond the scope of this paper, which focuses on distributed coordinations. 60 As discussed in Bošković (2002), SC shows superiority effects in certain contexts; one such context is the one given in (146). 61 (i) may indicate that the wh-phrase that enters the structure first may not need to merge with the coordinator immediately, but simply before the other wh-phrase enters the structure. This is so if 'what car' in (i) moves in front of the verb before kakva is sideward-merged into the late formed ConjP. Alternatively, it is possible that kakva sideward merges into the late formed ConjP before the object moves in front of the verb in (i). (i)  what-kind-of and whose today car father is-buying 62 While I have been assuming that sideward merger into ConjP takes place after the short movement step, it could even take place before-in the latter case the original copy would both sideward-merge and undergo the short movement step; note the lack of a c-command relation, even the ConjP doesn't c-command anything before integration into the structure.