How does Juicy Game Feedback Motivate? Testing Curiosity, Competence, and Effectance

‘Juicy’ or immediate abundant action feedback is widely held to make video games enjoyable and intrinsically motivating. Yet we do not know why it works: Which motives are mediating it? Which features afford it? In a pre-registered (n=1,699) online experiment, we tested three motives mapping prior practitioner discourse—effectance, competence, and curiosity—and connected design features. Using a dedicated action RPG and a 2x2+control design, we varied feedback amplification, success-dependence, and variability and recorded self-reported effectance, competence, curiosity, and enjoyment as well as free-choice playtime. Structural equation models show curiosity as the strongest enjoyment and only playtime predictor and support theorised competence pathways. Success dependence enhanced all motives, while amplification unexpectedly reduced them, possibly because the tested condition unintentionally impeded players’ sense of agency. Our study evidences uncertain success affording curiosity as an underappreciated moment-to-moment engagement driver, directly supports competence-related theory, and suggests that prior juicy game feel guidance ties to legible action-outcome bindings and graded success as preconditions of positive ‘low-level’ user experience.


INTRODUCTION
Juiciness or immediate "excessive amounts of feedback in relation to user input" [53, p. 139] is widely present in video games.It afects play time, player experience, and motivation [28,35] and is considered a key part of good game feel [26,53].Gamifcation researchers have therefore called it out as a motivational afordance that could make interactive systems in general more enjoyable and intrinsically motivating [17].But despite its ubiquity and recognised importance, views on what counts as "juicy" feedback diverge, and we do not know when and why immediate excessive feedback is motivating and enjoyable.Answering this question can both help designers better fne-tune feedback and advance HCI understanding of its interaction aesthetics.
Prior work has generated bottom-up characteristics of what developers consider good or juicy game feel [26,53] and exploratory empirical tests of various forms of amplifed feedback and player experience.To advance the feld, we need theory-guided, experimental work to systematically identify, isolate, and test precise design features and psychological mechanisms [3].To this end, we identifed three theories and candidate motives that mapped prior practitioner concepts: efectance-the basic positive experience of causing efects [71]; competence-exercising, expanding, and expressing one's abilities [57]; and curiosity-interest in generating and reducing uncertainty [36].We conducted a pre-registered, largescale online experiment (n=1,699) with a purpose-designed action role-playing game that allowed us to manipulate design features linked to these theories.In a 2x2+control study design, we systematically varied amplifcation (feedback volume), success dependency (whether feedback was triggered by actions or actions succeeding at a challenge), and variability (whether feedback was diverse).We measured efectance, competence, curiosity, and enjoyment with self-report, and tracked voluntary playtime.We used structural equation modelling to analyse relations between design features, experiences, and behaviour.
To our surprise, curiosity emerged as the strongest enjoyment and only playtime predictor, while success dependency drove curiosity, efectance, and competence alike.Contrary to theoretical predictions, enjoyment did not predict playtime, and amplifed feedback negatively impacted efectance and competence.We interpret these fndings to mean that even at low-level action-feedback loops, motivation and enjoyment arise from reducing uncertainty over action success, which depends on legible and diferentiated amplifed feedback.There can therefore be 'too much of a good thing' in amplifed feedback where it occludes causal action-feedback links or fattens success gradations.
Together, our fndings make several contributions to HCI: they evidence curiosity as a major, under-appreciated predictor of enjoyment and engagement in moment-to-moment interaction.They support that the efect of amplifed feedback on enjoyment and engagement is almost fully mediated by curiosity and competence, and to a lesser extent, efectance.They directly test and support self-determination theory claims that granular competence feedback drives competence experience and enjoyment.They surface sense of agency as a potential moderator for 'low'-level positive user experience.And they empirically support and psychologically specify prior design guidance on coherent juicy game feedback [26]: 'overloading' amplifed feedback likely interferes with competence and efectance by occluding action-feedback links and making diferences in task difculty and success less salient.

BACKGROUND
HCI traditionally conceives system feedback as the output part of the input-output interaction loop crossing the "chasm of evaluation" [51].That is, feedback helps a user form an accurate mental model of the current system state, e.g., whether it has correctly registered user input and shifted into the user-desired state in response.System feedback is thus framed as an informational usability factor, as can be seen in related design heuristics like "visibility of system status" or "ofer informative feedback" [32, p. 176].
In contrast, game feedback-particularly juicy feedback-is framed as an active contributor to a positive user experience [17,26].This makes juiciness an "interaction aesthetic" [44] or positively beautiful quality of interaction at the "motor level" of bodily, momentto-moment input-output loops [42].While juiciness is absent from relevant systematic reviews of interaction aesthetics [42], it is arguably related to Löwgren's "pliability", an "almost exaggerated quasi-physicality of the objects that are manipulated" [44].

Juiciness
The term "juice" or "juiciness" was likely coined by independent game designers Kyle Gray and colleagues [25]: "'Juice' was our wet little term for constant and bountiful user feedback.A juicy game element will bounce and wiggle and squirt and make a little noise when you touch it.A juicy game feels alive and responds to everything you do -tons of cascading action and response for minimal user input.It makes the player feel powerful and in control of the world, and it coaches them through the rules of the game by constantly letting them know on a per-interaction basis how they are doing." The term has since been in wide and varied use without a consensus defnition [8,17,26,33,53,61].Juiciness is commonly framed as a part of "game feel", "the tactile, kinesthetic sense of manipulating a virtual object" aforded by "real-time control of virtual objects in a simulated space, with interactions emphasized by polish" [64, p. 6].Thus, Hicks and colleagues' [26] framework of 13 heuristic-style characteristics of game juiciness equates game feel and juiciness.A more recent literature survey by Pichlmair and Johansen [53] analytically separates juiciness as one of three components of game feel: (1) tuning physicality, (2) streamlining support, and (3) juicing amplifcation, which describes an "intensifcation of experience" by "adding feedback to emphasise, clarify, and amplify [...] the intended game feel" [53, p. 147].
To summarise, "juiciness" is presently an ambiguous social category.This necessitates any research to explicate which defnition it operates with.Following the common denominators above, we here defne and focus on juiciness as the design feature of immediate "excessive amounts of feedback in relation to user input" [53, p. 139] at the level of moment-to-moment interaction.Thus, our present study does not speak to the full range of features or qualities associated with juicy feedback (e.g., tactility or amplifying any designer-intended afect).To avoid confusion, we will use amplifcation or amplifed feedback in the following whenever we refer specifcally to immediate excessive feedback as our chosen defnition.

Prior empirical work.
Moving to the empirical literature, Juul and Begy's frst between-subjects study (n=46) of an experimental match-3 game found that juicy and non-juicy versions did not signifcantly difer in ease of use, rated quality, objective performance, or "feeling clever" [34].In two within-person experiments (n=40 and n=32) across three games, Hicks et al. [27] found a positive efect of visual juiciness on aesthetic appeal across studies, no efect on usability or performance, and an efect on curiosity, competence, and presence/immersion in only one game.Another within-person experiment (n=36) with a VR game comparing a baseline, juicy, gamifed and juicy+gamifed condition found signifcantly higher competence, autonomy, relatedness, presence, interest/enjoyment, and preference ranking for the juicy over the baseline condition [28].Two within-person studies (n=26, 38) that combined low and high visual juiciness with no, low, and high haptic juiciness found that haptic juice signifcantly induced player preference as well as enjoyment, appeal, immersion, and meaning, but not mastery [62].In a within-person study (n=113) of the automatic game design tool Squeezer [31], Johansen and Cook found that players signifcantly preferred game versions with expert-crafted juicy feedback over a non-juicy baseline, and signifcantly preferred 2 of 6 juicy feedback versions produced with their automated tool [30].Finally, in a large between-subjects online study (n=3,018), Kao found evidence for a goldilocks efect: juiciness signifcantly afected competence, autonomy, presence/immersion, as well as interest/enjoyment, playtime, and performance; however, medium and high juiciness conditions outperformed low and extreme juiciness across all these measures [35].
In summary, the evidence base is characterised by mixed results, high measurement and operationalisation variance, low statistical power, and exploratory efect searching: out of 6 papers, 4 had samples between 26 and 46 per study, 2 used unvalidated single items or preference rankings, and of the 4 papers using validated scales, 3 ran full batteries of 7 up to 20 constructs, while only 2 pre-specifed clear hypotheses for at least one individual construct.In tendency, players fnd games with juicy feedback more appealing, preferable, and enjoyable; for other constructs like mastery, immersion, usability, or objective performance, the evidence is mixed.This could be due to the above methodological variance and issues; the fact that studies used diferent understandings and operationalisations of juiciness; or because juicy feedback requires subtle and holistic design [26,53], which diferent studies were variously (un)successful to ensure.

Candidate Psychological Mechanisms
So what intrinsically motivating experiences might amplifed feedback aford?To answer this, we surveyed practitioner views and prior mediators proposed or tested in the literature, to then map these to existing theories of intrinsic motivation.Gray et al. identify being "powerful and in control of the world" [25] as an engaging feeling arising from juice; Swink [64, p. 10] proposes "the aesthetic sensation of control" and "the pleasure of learning, practicing and mastering a skill".Schell suggests an interaction feeling "powerful and interesting" [61].Hicks et al. 's developer survey lists visceral emotion, fantasy fulflment, mastery, and meaningful actions, but did not link these to existing theories [26].Juul suggests that juiciness could reduce enjoyment and ease of use (by creating extraneous cognitive load) and improve them [33], specifcally by helping players "feeling clever" [34].Deterding hypothesises that juiciness afords competence via "positive feedback on [...] actions and achievements" and curiosity via "unexpected variety" [17].Pichlmair and Johansen [53] suggest that juicing "empowers the player", but do not specify what psychological construct (if any) they refer to.
To map these accounts to existing theory, we assessed (1) which theories of intrinsic motivation have already been suggested and/or empirically supported in the literature, (2) whether these ft the experience labels proposed by practitioners, and (3) whether they provide a convincing rationale why amplifed feedback afords enjoyment and motivation.This resulted in three selected constructs: efectance, competence, and curiosity, detailed below.This selection excludes meaning, autonomy, relatedness, and presence/immersion which had been tested in prior work [27,28,35], since we saw no strong theoretical rationale why amplifed feedback should evoke these.We also excluded aesthetic appeal, which is conceptualised as a functional positive player experience, not as an intrinsic motive [69].
2.2.1 Efectance.Robert W. White introduced efectance to psychology as the proposed single intrinsic motive: in play and other exploratory behavior, we seek the positive satisfaction of a "feeling of efcacy" [71, p. 322].Building on this, Christoph Klimmt's Synergistic Multiprocess Model of Video Game Entertainment [37,39] posits efectance as one of three sources of gaming entertainment underlying interactivity, namely the basic, positively valenced sensation of our actions causing change in the world.Klimmt expressly distinguishes efectance from competence [57]: whereas efectance relates to 'raw', moment-to-moment causing efects, competence arises from the successful pursuit of difcult goals in longer episodic stretches of gameplay [37, pp. 81-85].Klimmt suggests that games aford efectance through the temporal immediacy and disproportionality of game output on player input [37, pp. 76-81].This exactly matches our common denominators of juicy feedback: a positive experience evoked by immediate amplifed feedback on momentto-moment inputs.It maps onto Gray and colleagues' original idea that juicy feedback "responds to everything you do", as well as a sense of "control" [25,64] and being "powerful" [25,61]; it arguably approximates "meaningful action" [26] and "empowerment" [53].
Manipulating the reliability with which a game registered player inputs as an operational proxy of efectance, Klimmt and colleagues found that players reported signifcantly lower enjoyment in the low-reliability/efectance condition [38].Another study found selfreported efectance to be higher in interactive narrative games when players could afect the game state versus watching replays [56].However, there have been no correlational or mediation analyses of the impact of efectance on enjoyment or behaviour, nor studies testing whether amplifed feedback afords more efectance than non-amplifed one.
2.2.2 Competence.Self-determination theory (SDT) has become a prevalent theory in games HCI [67], positing that we are intrinsically motivated to play games because it satisfes basic psychological needs for competence, autonomy, and relatedness, while 'fun' or interest and enjoyment are experiential signatures of this need satisfaction [59].
Moving to amplifed feedback, SDT posits that positive competence feedback-environmental events suggesting that we are efective in pursuing challenging activities and goals-is one main causal antecedent of competence satisfaction [57, pp. 121-131], that is, feeling capable in engaging the world as well as growth in our capabilities [57, p. 86].There is good evidence that positive (verbal) feedback enhances competence satisfaction, which leads to intrinsic motivation and engagement [57, pp. 153-157].Video games aford immediate, consistent, and dense competence satisfaction via "rich, multi-level, efectance-relevant, positive competence feedback." [57, p. 514] Restating amplifed feedback in their own terms, Rigby and Ryan highlight granular competence feedback (see also [57, p. 514]: In Guitar Hero, "there is the immediate feedback that players get on each and every strumming of their virtual guitar; through colourful fashes and sounds, the player sees and hears immediately whether or not they hit the note correctly.[...] these create mastery feedback loops [...] that instantaneously and consistently provide competence satisfactions [...] We call these [...] granular competence feedback because they have a one-to-one relationship to each of the player's individual actions." [55, p. 23].Thus, while efectance theory posits that amplifed feedback on all player action is already enjoyable and engaging, SDT makes competence success-dependent: Granular feedback informs players about "whether or not they hit the note correctly"; if they do so-if the feedback is positive about their abilities to achieve a challenging outcome-, then players feel competent [57, p. 152-3]. 1 This matches practitioner phrases like "the pleasure of [...] mastering a skill" [64], "mastery" [25], or "feelings of competence and mastery" [26]. Only few SDT studies have directly explored game feedback, with mixed results [9,48,52].One study found that competence satisfaction was higher when the textual result screen was positively rather than negatively phrased.Puzzlingly, positive feedback and competence correlated negatively with the likelihood to play the game again [9].Another study compared competence-supportive and non-supportive game conditions (which included feedback manipulation among others), where the competence-supportive version showed higher competence satisfaction, enjoyment, and motivation for future play [52].A study examining points as a form of granular competence feedback found no signifcant efect on competence satisfaction or intrinsic motivation [48].The abovereviewed studies variously fnd signifcantly higher competence under juicy conditions or not.Notably, while Gray and colleagues' original coinage [25] and efectance theory posit that amplifed feedback on just performing actions is enough to make players 'feel good', Swink [64] and SDT posit that amplifed feedback only leads to a competence experience when provided on an action that displays some player skill.To our knowledge, no prior work has tested these competing claims [27,28,35,62]; juicy conditions either comprise both raw action feedback (e.g., particle efects and animations on player movement) and success-dependent feedback (animation and particle efect on enemy death) (e.g., [27]), or are not documented sufciently to be able to tell whether both were manipulated (e.g., [35]).

Curiosity.
Curiosity is widely studied as the main motivator of play and exploratory behaviour in cognitive science, neuroscience, ethology, and developmental psychology [1,10,12,23,36], while games research on curiosity is more nascent [22,65,66,72].Curiosity is commonly theorised as an emotion or intrinsic motive driving information-seeking [36], leading organisms to seek out or generate stimuli with "collative variables" like novelty, surprise, uncertainty, complexity, diversity, or ambiguity [46] that promise information gain.White [71, p. 322] himself stated that environmental responses to our actions need novelty and variance to evoke efectance: "efectance is aroused by [...] diference-in-sameness.This leads to variability and novelty of response, and interest is best sustained when the resulting action afects the stimulus so as to produce further diferencein-sameness. [...] efectance motivation subsides when a situation has been explored to the point that it no longer presents new possibilities." This suggests that what makes juicy feedback engaging is (expected) information gain about causal action-feedback links and the range of possible game behaviour.This matches practitioner labels of juicy feedback like "interesting" [61], "feeling clever" [34] and characteristics like "uncertain outcomes" [26].As proposed by Deterding in his "lens of juicy feedback", "unexpected variety stokes curiosity" [17, p. 313]: the novel, surprising, uncertain, and complex ways that diferent items fall and explode and cascade into chain efects on every swipe keeps match-3 games like Candy Crush Saga interesting and pleasantly surprising, eliciting and satisfying curiosity as long as we haven't learned the possibility space of feedback patterns such that feedback has become fully predictable.
Supporting this logic, game designer Greg Costikyan identifes uncertainty as a core motivational afordance of games, including performance uncertainty and randomness [14].Developers in Hicks and colleagues' survey similarly mention "uncertain outcomes" as an important game feel characteristic [26].A recent qualitative study on moment-to-moment engagement in casual games found that uncertainty over the game's response to each player's action induced positively motivating curiosity [40].In a lottery task study, behavioural and neurological signatures of curiosity tracked the buildup and release of outcome uncertainty [68].Using an Asteroidavoidance game, another study found that people waited to see outcome uncertainty-resolving feedback even if doing so did not improve their performance, suggesting that it is disjunct from competence [29].One above-mentioned study found that juicy feedback drove curiosity [27].That said, we are not aware of studies testing whether collative variables like variety in amplifed feedback drive curiosity, nor severe tests of curiosity mediating enjoyment and voluntary engagement from amplifed feedback.

The present study
To summarise, prior work has generated diferent conceptualisations of juiciness, descriptive synthesis of practitioner views, and initial, if mixed and mostly exploratory empirical data supporting that amplifed feedback can be enjoyable and motivating, due to various possible (but largely unspecifed) mechanisms and design features.Mapping this work to prior theory, we identifed three candidate theories specifying diferent antecedents and mediators: efectance aforded by all amplifed feedback; competence aforded by success-dependent amplifed feedback; and curiosity aforded by variable amplifed feedback.None of these theoretical predictions have been directly tested to date.Thus, our study aims to directly test competing theory-derived predictions about how amplifed feedback afords enjoyment and intrinsic motivation, thereby also clarifying conficting claims in prior practitioner views Just as multiple motives and emotions can direct our behaviour in parallel [54], amplifed feedback could operate via one, two, or all three predicted pathways in parallel.We therefore adopted and pre-registered an experimental and analysis strategy that would allow inference to the best explanation under the possibility of multiple parallel efects. 2  Because success dependence and variability are qualifcations of amplifed feedback, we can only manipulate them if feedback is already amplifed.This led us to create 2x2 experimental conditions that varied success dependence (yes/no) and variability (yes/no) of amplifed feedback, with a dangling control condition with 'standard', non-amplifed feedback.We specifed two matching confrmatory structural equation models, one for amplifcation, one for variability and success dependence (see Figure 1).
For each construct, our pre-registration specifed two hypotheses about experimental conditions and matching path predictions in our structural equation models, indicated as thick arrows in Figure 1, where the direct paths from amplifcation, variability, and success dependence to enjoyment tests for the absence of non-mediated efects.
H1 Efectance Based on theory, we expect that amplifed feedback will generate enjoyment and subsequent voluntary engagement, fully mediated by efectance.This means: H1a: Enjoyment will be higher under non-success dependent amplifed feedback than standard non-success dependent feedback.H1b: Voluntary engagement will be higher under non-success dependent amplifed feedback than standard non-success dependent feedback.As causal paths, our model tested: • There will be a signifcant positive correlation between amplifcation and efectance.• There will be a signifcant positive correlation between effectance and enjoyment.
H2 Competence Based on theory, we expect that amplifed feedback will cause competence experience if it is success-dependent, which will lead to enjoyment and subsequent voluntary engagement.This means that: H2a: Enjoyment will be higher under success-dependent than non-success-dependent amplifed feedback.H2b: Voluntary engagement will be higher under successdependent than non-success-dependent juicy feedback.As causal paths, our model tested: • There will be a signifcant positive correlation between success dependence and competence.• There will be a signifcant positive correlation between competence and enjoyment.
H3 Curiosity Based on theory, we expect that varied amplifed feedback will lead to enjoyment and subsequent voluntary engagement, fully mediated by curiosity.This means that: H3a: Enjoyment will be higher under variable than non-variable amplifed feedback.H3b: Voluntary engagement will be higher under variable than non-variable amplifed feedback.As causal paths, our model tested: • There will be a signifcant positive correlation between feedback variability and curiosity.• There will be a signifcant positive correlation between curiosity and enjoyment.
2 Pre-registration: https://osf.io/yvu3c/Voluntary engagement As all theories suggest that enjoyment leads to voluntary engagement, our model tested: There will be a signifcant correlation between enjoyment and voluntary engagement.

Modifcations to preregistered analysis
We added (1) potential covariances between curiosity, efectance, and competence and (2) potential direct efects from design features (juiciness, success dependency, and variability) to voluntary engagement as exploratory analyses to our pre-registered models.These explore potential pathways to voluntary engagement not mediated by enjoyment since our enjoyment measure, the Intrinsic Motivation Inventory (IMI; 58), is somewhat SDT-specifc, and not all intrinsic motivation theories predict that enjoyment mediates [50].Figure 4 shows the fnal models.See SEM_Analysis.R for the pre-registered and main_analysis.R for the fnal analysis script on the OSF repository. 3The pre-registration and connected materials use the term "juicy" throughout to refer to immediate excessive feedback.The present manuscript replaces "juicy" with "amplifed" throughout to signal the specifc defnition of juiciness we refer to.

METHODS
Our preregistration 2 as well as data and analysis scripts 3 are accessible on the Open Science Framework (OSF).Detailed documentation of study conditions (including gameplay video) can be found on our online documentation. 4.1 Study Game Platform The game is a PC fantasy action roleplaying game (RPG) controlled with keyboard and mouse, similar to, e.g., Diablo, where the player controls a knight.The game begins with a skippable ~1 minute tutorial teaching movement and combat controls.Players can move (using W, A, S, D), jump (spacebar), and sword attack (left-click).The game features fve areas with diferent monsters (e.g., goblins, slimes, spiders) and quests (e.g., defeat the hobgoblin).Earning experience points from defeating monsters and completing quests levels up the player's character, increasing their health and damage dealt.If a player dies, they respawn at the beginning of their current area.Monsters move and attack diferently, and some have special abilities, e.g., giant spider attacks immobilise the player.The environment further features grass and butterfies that respond if the player character moves through them.

Development and Validation.
Over the course of 20 months, we developed and tested the game using Rapid Iterative Testing and Evaluation (RITE) with university students recruited through a university's newsletter as test users.We conducted four rounds of testing with 3-4 participants each.Participants played all study conditions in randomised order, then were interviewed using semistructured interviews (see OSF repository for interview script 3 ), and were compensated with a $15 gift card.Feedback from each round was synthesised and translated into changes to the game for the next round until no additional issues were raised.This was interwoven  with three additional rounds of feedback and iteration with the authors and external experts to ensure the game's suitability for the study.Absent well-specifed consensus defnitions or operationalisations of juiciness, we considered but discarded expert member checking with game developers as a construct validity test: since game developers have diverging and diferently well-developed views on what constitutes juiciness [8,24,26], they may not have aligned with each other nor our tested conceptualisation of juiciness.Instead, we directly involved creators of the empirically grounded game juiciness framework [26] as an external validity check that our amplifed conditions ft their characteristics of game juiciness, while our standard, non-amplifed condition did not.This led, e.g., to the addition of rising butterfies and moving grass animations to match their characteristic B4, "ambient feedback" [26].Similarly, we ended up using commercial Unity animation assets (e.g., sword swing animations from the "Magic Slashes FX" package) that ft the fantasy setting of the game, in line with characteristic A2, "thematic coherence." The involved creators were satisfed that our fnal version ft their characteristics of game juiciness.
For construct validity, we mainly aimed to ensure that our conditions operationalised our defnition of amplifed feedback and theoretically relevant design features (success dependence, variability), as our main aim was theory-testing.Matching our defnition, all manipulated feedback is directly caused by and immediately follows moment-to-moment player actions: moving and sword attack.Feedback comprised direct movement and sword swinging sound and animation, and indirect grass movement, butterfy movement, enemy hit, enemy kill, potion pickup, and experience point gain/levelling up depending on when and where the player moved or attacked.During RITE, playtesters played all conditions and were then openly interviewed to describe diferences they noticed between conditions, and we checked that their descriptions matched our desired diferences and no condition was unclear.
For amplifcation, we further checked that playtesters described amplifed conditions as having stronger or more feedback than the standard condition, by asking them to identify diferences between conditions.At the same time, we expressly prompted playtesters to report whether the standard condition provided sufcient feedback such that they still understood gameplay and the consequences of their actions.For success dependence, our manipulation needed to operationalise SDT predictions that competence arises from at least marginally difcult and skill-requiring tasks.We chose to manipulate sword attacking as a basic, thematically ftting game action that (a) produces plausible immediate feedback on actuation (swooshing sound and animation) for amplifed non-success-dependent feedback, and (b) also involves a non-trivial challenge with plausible amplifed success-dependent feedback, here: sound and animations on actually hitting a moving enemy, as well as enemy death.To ensure the sword attack involved challenge, we balanced the game during RITE such that players would sometimes miss enemies on attack.To operationalise variability, varied conditions used randomised diferent sounds and animations from the commercial Unity asset packages, while the non-varied conditions used a single fxed sound and animation from the same package, again aiming for thematic coherence.We asked particular interview questions to check that the audio-visual efects for varied feedback were all similar in perceived strength for each action as well as comparable to the non-varied efect for each action.
To avoid game difculty and learnability confounds, we checked whether the game was easy to learn and playtesters performed roughly equally well across conditions.To control for variability in performance, we frst ran a benchmarking procedure with 80 online participants recruited from Prolifc.The benchmark tested the most performance-intensive actions in the game for each condition and recorded the frames per second (FPS).Based on these results, we (a) optimised code to improve performance (focused on collision detection computations), (b) created a set of minimum machine requirements for participants, and (c) devised a short benchmarking test that ran at startup in the main study.This test, performed invisibly to the user behind a loading screen, simulated the most performance-intensive situation: attacking an enemy consecutively with background juicy environmental features present.We performed this test for all fve conditions and then took the lowest average FPS across conditions and set this as the maximum frame rate for that participant.This controlled for variance in frame rate across conditions and participants-a participant's FPS will always be limited by the most performance-intensive condition, regardless of their actual assigned study condition.If that lowest FPS was less than 30, we informed the user that their machine doesn't meet requirements and excluded them from completing the study.

Conditions
The study used a between-subjects, 2 x 2 design with amplifed (A) and variously success-dependent (+/-SD) and varied (+/-V) conditions plus an additional dangling non-amplifed control group (STND), leading to 5 total groups that participants were randomly assigned to (Figure 2 and Table 1; see the online documentation 4  for condition details including video, audio, and source code).

Measures
3.3.1 Dependent variables.We measured efectance using Klimmt et al. 's efectance in games scale [38], specifcally, the adapted version that is recommended in Ballou et al. 's validation study [4] (i.e., items 5, 7, 9, and 11, measured on a 7-pt Likert scale from 1 ("not at all true") to 7 ("very true").We use the adapted version because the original scale was found not to be unidimensional, suggesting it is not construct-valid [4].We verifed using confrmatory factor analysis that the adapted 4-item measure functioned well in our sample (strong model ft and all item loadings > .69;see OSF repository 3 ).We measured curiosity and competence using the curiosity and mastery subscales of the Player Experience Inventory/PXI [69], measured on 7-pt Likert scales from -3 ("strongly disagree") to 3 ("strongly agree").We measured enjoyment and self-reported intrinsic motivation using the interest/enjoyment subscale of the Intrinsic Motivation Inventory/IMI [58], measured on a 7-pt Likert scale from 1 ("not at all true") to 7 ("very true").We omitted all other subscales from the PXI and IMI since they were not of theoretical relevance to our study and the PXI and IMI subscales are all fully disjunct.Both PXI and IMI are well-validated scales widely used in games HCI and prior juiciness research [27,28,35], and showed very good model ft with high loadings in our sample (all items loaded > .74 with the exception of one IMI item at .59; see OSF repository 3 ).Finally, we operationalised voluntary engagement or the behavioural expression of intrinsic motivation as voluntary time on task, in line with common practice [58].After 10 minutes of mandatory playtime, players flled in a survey on the measures above and were informed that they could now end the experiment or continue playing for as long as they want.We measured voluntary engagement as minutes of play after this point.

Other measures.
To describe our sample and check for potential confounds, we additionally measured prior play experience for video games and action role-playing games using the questions "How would you rate your prior experience playing video games?" and "How would you rate your prior experience playing action roleplaying video games?" both on a 7-pt Likert scale from 1 ("minimal") to 7 ("extensive").We also asked "How many hours of video games do you play approximately per week on average?"To test that game difculty did not accidentally co-vary across conditions,we measured challenge using the PXI challenge subscale [69] on a 7-pt Likert scale from -3 ("strongly disagree") to 3 ("strongly agree").We also included one attention check item.

Sample Size Determination
Although we use validated questionnaires, these have not yet been used in studies to determine efect size benchmarks (e.g., using anchor-based methods [2]), and prior literature was insufcient to estimate expected efect sizes or specify a smallest efect size of interest, precluding an a priori power analysis.We followed recommended practice in these circumstances [41] to instead base our sample size on the largest number of people that could be collected with the monetary resources available to the project.This maximally afordable sample recruited was N=1,700.Based on past experience, we expected to remove up to 200 participants who failed our attention check, meaning we would have a sample of N=1,500 valid participants.

STND
Standard Non-Varied Feedback without amplifcation.Sound efects on swinging, hitting, and enemy death.No enemy death animation (the enemy disappears).A-SD-V Amplifed Non-Success-Dependent Non-Varied Feedback.Exaggerated audiovisual impact efects occur even without hitting an enemy, but there is only one efect for each feature, e.g., one swing efect, one impact efect, one enemy death sound efect, etc. A-SD+V Amplifed Non-Success-Dependent Varied Feedback.Exaggerated audiovisual impact efects occur even without hitting an enemy, and there are many possible efects, e.g., many swing efects, many impact efects, many enemy death sound efects, etc. A+SD-V Amplifed Success-Dependent Non-Varied Feedback.Impact efects will occur only when the player successfully hits an enemy, but there exists only one efect, e.g., one swing efect, one impact efect, etc. A+SD+V Amplifed Success-Dependent Varied Feedback.Impact efects will occur only when the player successfully hits an enemy and are varied, so there exist many possible efects, e.g., many swing efects, many impact efects, etc.
To determine the efect size sensitivity of this sample size, prior to collecting data, we ran a Monte Carlo simulation in R using lavaan version 0.6-11.Simulation results indicated that with 1500 participants, we have 89-91% power to detect standardised efects of = .1 on the pathways from curiosity, efectance, and competence to enjoyment; and 77-80% power to detect standardised efects of = .1 on the pathways from juiciness, variability, and success dependence to curiosity, efectance, and competence.By common benchmarks, our study thus had adequate power to detect small ( < .2) efects [19].Further details and results of our simulations are available in the script SEM_Analysis.R found in the preregistration.

Participants
We recruited 1,706 participants via the online recruitment platform Prolifc.Out of these, only 6 failed the benchmarking test and only 1 was removed for failing the attention check question.This left us with n=1,699 valid participants for an average 339.8 participants per condition (SD=22.1)(see Table 1 for n per condition).We used simple randomisation, which can naturally lead to some variation in group sizes.Our group sizes deviated at most 2% from the average 20%, which is considered acceptable especially at large sample sizes, and does not afect the robustness of our results.60.8% of participants identifed as men, 36.4% as women, 2.1% as gender variant/nonconforming, and 0.7% as transgender.Participants had an average age of 27.1 (SD=8.0).Using Prolifc's pre-screening criteria, the task was available only to participants who were at least 18 years old, had English fuency, and had a desktop computer with working audio.Participants were told before accepting the task that they must own a Windows or Mac machine that meets our minimum requirements which are specifed in the preregistration.There was no limitation on geographic location.Participants were from South Africa (18.1%),Poland (17.5%),Portugal (15.8%),Italy (8.3%), Mexico (8.1%),United Kingdom (6.6%), Greece (4.7%), Spain (4.4%), Hungary (2.6%), Chile (1.8%), Netherlands (1.6%), Czech Republic (1.5%), France (1.4%), Germany (1.1%), Canada (0.8%), and a combined 5.9% from 15 other countries.The Institutional Review Board (IRB) at the frst author's institution approved the study.All participants provided informed consent, and were paid an average US$11.29 per hour, including excluded participants.

Procedure
Participants are tasked to fll out a consent form and then randomly assigned to a condition.They then undergo an audio check, during which they are required to type a spoken English word.Next, they download and play their condition version of the game that is compatible with their operating system (Windows or MacOS).Upon launching the game for the frst time, the benchmarking test is automatically conducted, as described in Section 3.1.2.If a participant fails the benchmarking, they are instructed to return the task on Prolifc.
Participants passing the benchmark are instructed to play the game for a minimum of 10 minutes.At the 10-minute mark, participants are automatically prompted with an in-game survey containing measures for efectance, curiosity, competence, and enjoyment.The order of survey measures and items within each measure are randomised for each participant.Upon completion of the survey, participants are informed that they may exit the game whenever they wish.Voluntary engagement time is automatically logged thereafter.Upon exiting the game, participants complete a postsurvey that includes measures for prior play experience and challenge, followed by a demographics questionnaire.

Analysis
We ft two structural equation models to analyse our results, specifed in Figure 1.Models are ft using the sem() function of lavaan version 0.6-11 using the robust MLR estimator.Precise model syntax and model ftting code are available in the main_analysis.R script on OSF. 3  In Model A, we assess the efect of amplifcation on efectance, competence, and curiosity, and ultimately enjoyment and freechoice playtime.We include only the standard feedback condition (STND) and the amplifed non-success-dependent, non-varied feedback condition (A-SD-V); this allows us to isolate the efect of amplifcation itself.In Model B, we assess the efects of successdependence and variability on efectance, competence, and curiosity, and ultimately enjoyment and free-choice playtime, using all amplifed conditions.
For both Model A and Model B, we conduct model comparisons with alternative models where the mediated pathways (i.e., the pathways from the design feature → efectance, competence, and curiosity; and from efectance, competence, and curiosity → enjoyment) are fxed to 0. We compare model ft between the restricted and unrestricted models using ΔCFI, ΔRMSEA, and Δ 2 statistics.
We evaluate the ft of each SEM model using standard model ft indices (CFI, RMSEA, SRMR, 2 ).We do not specify cut-of criteria in advance, given evidence that these are overly simplistic [11].Instead, we decide whether the models ft appropriately using all these indices in tandem.Where the models do not ft sufciently well, we use residual covariances, modifcation indices, and domain expertise to make adjustments that improve the model ft, and transparently report all decisions and alterations made.If we achieve a well-ftting model, we infer the presence of a signifcant efect when the p-value of a particular pathway is < .05.

Descriptive Measures
Descriptive results can be seen in Figure 3 and Table 2.All efect sizes are presented in Figure 4 as unstandardised coefcients, which we believe are more informative than standardised ones, including levels of signifcance.
As predicted, success-dependence led to signifcantly greater competence (.45, p<.001), and competence was again signifcantly positively associated with enjoyment (.26, p<.001).Competence and voluntary engagement were higher in success-dependent than non-dependent juicy conditions (Table 3); we found no signifcant direct efects of success-dependence on enjoyment or voluntary engagement.This supports H2a and H2b and our broader SDT-based hypothesis that success-dependent amplifed feedback drives enjoyment mediated by competence.Contrary to prediction, enjoyment again was not signifcantly associated with voluntary engagement, which we return to below.
Also counter to our predictions, variability showed no signifcant correlation with curiosity, while as predicted, curiosity strongly correlated with enjoyment (.76, p<.001).Varied conditions showed slightly higher enjoyment than their non-varied counterparts, mirrored in a small but signifcant correlation between variability and enjoyment (.11, p<.05).

DISCUSSION
We will discuss ramifcations of our results for each theory in turn, followed by general refection on enjoyment and voluntary engagement and games HCI, prior work on juiciness, and implications for design.

Amplifcation, Efectance and Klimmt's Multi-Process Model
Surprisingly, amplifcation had a signifcant negative efect on efectance and competence.Is this a case of juicy feedback frustrating users via extraneous cognitive load, as Juul [34] proposes?We think not, or rather, we think this explanation is a case of imprecise theory use by Juul.In cognitive load theory [63], extraneous load refers to information unnecessary for learning something that interferes with limited working memory; swapping gems (as studied by Juul [34]) or in our case, actuating movement or sword attacks, arguably involve no such information memorisation.We propose that our results can instead be explained by the amplifed condition (A-SD-V) tested in Model A unintentionally impeding so-called outcome binding-attributing an observed event to one's prior intentional action [43].Outcome binding is a constituent subprocess of sense of agency or "the feeling of control over actions and their consequences" [49].Both may sound the same as efectance, but describe a 'neutral', low-level cognitive process, while efectance describes a (resultant) positively valenced afective-motivational state.How did condition A-SD-V impede outcome binding?Inspecting gameplay video of both the control (STND) and A-SD-V (Figure 5) shows that in the latter, the large light glow ball of the Weapon Swing animation often visually engulfs and occludes subsequent triggers of the same animation as well as separate animation effects for hitting or killing an opponent.Thus, players may not have sensed that they hit repeatedly or caused enemy hits or deaths.Our logic here is simple: To feel positively efcacious or competent over causing an observed event, we must frst sense that our actions caused it -and that may have been impeded by sword attack feedback occluding success feedback.In line with this logic, prior HCI work on sense of agency fnds it impeded by delayed (i.e., laggy or nonresponsive) as well as incongruous or unreliable feedback [43].This maps Klimmt's proposal that efectance depends on the temporal contiguity [39], and his fnding that unreliable feedback is less enjoyable [38].Next, we turn to the major disagreement between Klimmt's efectance-based Multi-Process Model [39] and SDT [57].SDT (like Swink [64]) argues that enjoyable juicy feedback must be successdependent, because positive (competence) experience requires observing attaining intended and at least marginally difcult tasks, while Gray and colleagues [25] and Klimmt argue that merely observing change caused by actions is already a positive (efectance) experience.If the latter were the case, success-dependence should not afect efectance.Yet we found a small-to-medium positive association between the two (.26, p<.001).A plausible counter-argument would be that our success-dependent conditions gave players more diferent action-consequence links to observe: swinging and hitting and killing.Also, we found signifcant (if small) independent associations between efectance and enjoyment.Thus, we do not see sufcient warrant to accept or reject Klimmt's proposal that efectance is a fully separate mediator.Further research with more careful manipulations is warranted here.

Success Dependence, Competence, and SDT
Our study directly tested and found support for SDT claims on granular positive competence feedback [55,57]: success-dependent amplifed feedback caused competence satisfaction with a moderately large efect size, which was signifcantly associated with enjoyment, while success-dependent amplifed feedback had no signifcant direct efect on enjoyment, supporting mediation.
The observed negative efect of amplifcation on competence can be explained in at least two SDT-congruent ways: (1) if amplifed feedback efects are trivial to accomplish (just by actuating a sword swing), they do not engender competence over doing well at something challenging, but may also comparatively 'cheapen' the attainment of more difcult tasks (like hitting or killing enemies) triggering the same intensity of competence feedback.Hicks and colleagues [27] ofer a similar explanation for why juiciness increased competence in only one of three games in their study.(2) As suggested above, our amplifed feedback condition may have impeded action-feedback outcome binding, which is a logical precondition for feeling competent over observed feedback.Interestingly, SDT has long proposed that competence satisfaction from a challenging task only occurs under perceived self-determination or having intentionally engaged in the task [15,16].This has been later reframed into competence satisfaction only occurring under parallel autonomy satisfaction [57].Our fndings suggest that at a moment-to-moment level, outcome binding ("my action caused this efect") as an aspect of agency could be the actual precondition of competence, or a further one in addition to intentional binding ("I caused this action").A recent review has highlighted a similar ambiguity of autonomy and agency concepts in HCI, variously referring to diferent aspects of causality and identity [6].Thus, our fndings point to plausible 'microscopic' feedback design features (outcome binding and success gradation) as preconditions not discussed or tested in prior work.

Variability and Curiosity
Curiosity was the strongest predictor of enjoyment by far and the only construct signifcantly correlated with voluntary engagement.Every point increase in curiosity was associated with a .75(Model A)/.76 (Model B) increase in enjoyment and .88/.86 additional minutes of gameplay, or a 10.8% gain of the average playtime of 7.9 minutes.
However, contrary to our hypothesis, varied amplifed feedback showed no signifcant association with curiosity, while success dependence did.One possible explanation comes from recent accounts of curiosity as expected uncertainty or prediction error reduction [13,18,21]: In varied state spaces with irreducible aleatoric uncertainty (that is, true stochastic randomness like dice rolls or our varied feedback), once possible states are known, there is no further satisfying uncertainty reduction to be had: we know which sides of the die can show or which animation loops can be triggered.In contrast, in state spaces where variance follows a reliable generating rule, even after the possible states are known, we can continue to satisfyingly reduce epistemic uncertainty over learning the rule when and why each state is likely to occur [60].This would also explain why success dependence increased curiosity: as Kumari and colleagues [40] found, even at moment-to-moment levels of input-output loops, people are curious to reduce "outcome uncertainty" over whether they succeed at each action.Our successdependent non-varied conditions could have made action-outcome links with uncertain success (i.e., enemy hits and kills) more salient than their counterparts: something players perceived that they could learn to reliably predict and cause better and better.
In summary, our fndings support that curiosity mediates the efect of amplifed feedback on enjoyment and engagement, but is aforded by reducible outcome uncertainty in success-dependent feedback, not by uniform amplifed nor (truly) randomly varied amplifed feedback.While nascent work has suggested curiosity as a potential driver of gaming enjoyment and motivation [45,65,66], our study provides (to our knowledge) the frst quantitative empirical support in games HCI that curiosity strongly drives enjoyment and voluntary engagement.Prior games HCI work has suggested gameplay forms like exploration aforded by open world games or social simulations aford curiosity [22].Our data supports that curiosity already operates at the level of game feel.Our fndings also suggest that aleatoric uncertainty or stochastic randomness may not uniformly add to curiosity (as Costikyan [14] seems to imply).

Enjoyment and Voluntary Engagement
SDT as the main theory of intrinsic motivation proposes that proximate experiences like competence need satisfaction lead to felt intrinsic motivation (here: enjoyment/interest, measured with the IMI), which leads to intrinsically motivated behaviour, here measured as voluntary engagement [57].In line with this, our data showed expected motive-enjoyment correlations for efectance, curiosity, and competence.Although voluntary playtime substantially varied between conditions (lowest 6.17 min in STND, highest 10.18 min in A+SD-V), we found no signifcant enjoymentplaytime association, as predicted by SDT.Meanwhile, curiosity was strongly directly associated with voluntary playtime in both models.Put diferently, players played more when they were curious in the game, but not when they enjoyed it.
We see one potential explanation in well-evidenced neurological diferential mechanisms for 'liking' and 'wanting' [7].Liking refers to positive hedonic experience upon consummatory behaviour (e.g., enjoying licking ice-cream), while 'wanting' or incentive salience refers to neural systems driving both anticipatory approach and consummatory behaviour even in the absence of liking (e.g., going to the ice cream store in order to buy and eat an ice cream).Enjoyment measures like the IMI or player experience measures like 'mastery' in the PXI chiefy capture post-hoc hedonic 'liking' experiences, while curiosity-motivated behaviour has recently been linked to 'wanting' [20].More specifcally, Murayama's reinforcement learning model of curiosity suggests that a 'wanting' state (manifesting as a sense of interest or curiosity) energises behaviour, while satisfying this want generates enjoyment or 'liking' experiences of surprise, insight, learning, or uncertainty reduction, reinforcing incentive salience [50].Put diferently, our players played voluntarily if they wanted to satisfy their curiosity about a novel game or whether they can beat it.Satisfying this curiosity then generated positive liking experiences of enjoyment, reinforcing their wanting to play.This would ft that curiosity correlated with enjoyment, but also, and independently, with voluntary engagement, and the lack of enjoyment-engagement links.We think this points games HCI toward curiosity and expectations as potential under-appreciated mediators of actual play behaviour.
Summarising across theories, we can say that the impact of amplifed feedback on enjoyment is mediated by curiosity and competence (and potentially, efectance), but not fully shaped by the design features predicted in each theory: success-dependent feedback afords competence as predicted, but also curiosity and efectance, while random feedback variety has no impact on curiosity.

Implications for Juiciness and Games HCI
We highlighted throughout that the current ambiguity of the term "juice" urges clarity and care in defnitions, operationalisations, and delimitations.We here specifcally tested and speak to variants of juiciness understood as amplifed feedback in immediate response to moment-to-moment player input as a common denominator of much prior work [17,26,28,33,35,61,62].Here, our fndings support prior observations [27,34,35] that amplifed feedback on its own does not necessarily lead to enjoyment or intrinsic motivation, and can even detract from them.Further, our fndings add psychological specifcation to prior design guidance.First, by sword swinging feedback engulfng enemy hit and death feedback, our A-SD-V condition arguably violated characteristic C3 of Hicks and colleagues' game juiciness framework [26]: "Unambiguous: Can information be connected to actions and only interpreted in one way?"As unpacked in detail above, our fndings suggest an explanatory mechanism for this characteristic not previously discussed in the literature: when amplifed feedback occludes or becomes undiferentiable from other action feedback, it may hamper action binding [49] as a sub-component of sense of agency, which is a logical precondition for both efectance and competence.
Second, to test competing claims of efectance theory and SDT, our non-success-dependent conditions ensured equally strong amplifed feedback on both acting (sword swinging) and succeeding (hitting and killing an enemy).This may have violated characteristics A4 ("Feedback Coherence: Does feedback refect the importance of the event?"),B3 ("Highlighting: Are feedback elements that highlight information in harmony with other systems?"), and C4.A ("Relevant: Is feedback giving in response to game critical events or is feedback received on minor player actions that require no further action.")[26].All three principles (A4, B3, C4.A) imply a kind of coherent information hierarchy, which Hicks and colleagues have tied to an informational function [26, p. 6].By inadvertently manipulating this hierarchy, our conditions may have violated good "game juiciness" as specifed in [26].Yet in doing so, they also provide empirical evidence that coherent diferentiated feedback has a direct motivational impact, not just an informational one, and that this impact is mediated by curiosity, efectance, and competence.We derive three new testable mechanisms for why this may be for future research: For competence, we suggested that the more difcult to obtain an outcome (killing enemy > hitting enemy > swinging a sword), the more amplifed feedback should be to signal the extent of the player's displayed competence.This seems intuitive but to our knowledge, has not been suggested nor tested in the literature [26,53,55].For efectance, we propose that diferentiated feedback allows the observation and learning of multiple, diferentiated action-outcome links.For curiosity, we suggest (in line with [40] that diferentiated success feedback makes reducible outcome uncertainty more salient. For games user research and games HCI more widely, we see three upshots from our study.First, it provides robust evidence that curiosity is an important enjoyment driver for moment-to-moment game control, also outside 'typical' gameplay forms and genres like open-world exploration.Second, curiosity not enjoyment predicts moment-to-moment play continuation, counter to SDT-informed games HCI [67] and wider received wisdom [47,70].This suggests that expectational wanting may matter more and more directly for play behaviour than consummatory liking [7], in line with emergent work on the role of expectations in player (dis)engagement [5].This invites future games HCI to explore curiosity as a potential alternative mediator to well-established concepts (like fow or need satisfaction), and separately study expected and realised positive player experiences (where common practice is to rely on self-report measures of the latter, such as the PXI).A third upshot is that outcome binding as part of sense of agency could be an important moderator of positive user experiences.It also provides an alternative, more low-level mechanism for SDT-informed games HCI why competence and autonomy satisfaction interact [57] with concrete and diferent interaction design implications.And it could help theoretically clarify ambiguity around the concept of autonomy in HCI itself [6].

Implications for Design
The unexpectedly negative efects of our A-SD-V condition underlines practitioner wisdom that fne-tuning juicy feedback is a delicate holistic design task [24,26].Our fndings add some-in parts speculative-design guidance to existing work [26,53].First, we have good evidence that success-dependent amplifed feedback enhances competence, curiosity, efectance, and enjoyment.Including such granular success feedback is already a design recommendation derived from SDT [55], but hadn't been empirically validated before.Second, amplifed feedback on mere action should be structured and proportioned to not diminish the relative value of success on challenging tasks.This aligns with existing guideline C4.A by Hicks and colleagues [26], but specifes it.Third, random variety in feedback is unlikely to enhance player experience and thus not recommended -this specifes Deterding's [17] design guidance for unexpected variety in juicy feedback, which did not preclude randomised variety.Fourth, when assessing the efectiveness of juicy feedback, game designers and games user researchers should not exclusively rely on self-reported enjoyment (e.g, via the IMI), but also assess actual voluntary playtime (and/or potentially, curiosity).

LIMITATIONS AND FUTURE WORK
If our occluding non-success-dependent feedback impeded agency, this could have suppressed true efectance efects.Future work should repeat our design, but with expert game design practitioners helping to ensure unambiguous and easily diferentiable actionfeedback links.Positively, we see rich future work in testing the potential moderating role of outcome binding and sense of agency on positive low-level user experience, including competence and efectance.Similarly, we identifed further candidate detail design features worthy of further study: diferentiable gradations of success feedback and truly random versus learnably generated feedback variation.Finally, on a theoretical level, we proposed sense of agency/outcome binding as a potential explanatory mechanism for SDT's proposed autonomy-competence coupling, as well as wanting and learned expectations as potential engagement drivers besides enjoyment.Both are exciting areas of future work.
In terms of generalisability, our study does not speak to all understandings of juiciness (such as enhancing any designer-intended game feel, tactility, aliveness, ...), nor claim external validity toward 'industry-level' polish.It was conducted in an action RPG and over a relatively short play duration, suggesting future work to test longer time scales and other genres, particularly turn-based and/or puzzle games where juicy feedback is less likely to directly interfere with real-time action.Our between-subjects design is more ecologically valid (few people play design variants of the same game in direct comparison), but may be less sensitive to real efects than within-person studies prevalent in prior work.

CONCLUSION
Juicy feedback is a widely used yet poorly specifed game design concept for positive user experience aforded by immediate amplifed feedback on player action.To systematically unpack when and why amplifed feedback is engaging, we tested three candidate theories with linked mediators and design features: efectance and amplifed feedback; competence and success-dependent feedback; and curiosity and varied feedback.For games HCI, our fndings contribute strong direct support for SDT-based claims that success-dependent amplifed feedback drives enjoyment, mediated by competence.They also suggest that sense of agency, specifcally outcome binding ("My action caused this efect") may be an important moderator of competence, efectance, and positive user experience in moment-to-moment interaction.Outcome binding may explain why good game feel is associated with unambiguous, coherent, highlighting and relevant feedback, and when and why amplifed feedback becomes 'too much of a good thing' (as opposed to cognitive load, proposed in prior work).Players need to be able to learn to attribute screen events to their actions, which requires reliable (nonrandom) and non-mutually occluding action feedback.And they may feel more competence if feedback kind and volume communicates the difculty of an attained game state.Our fndings further show curiosity to be an under-appreciated factor of enjoyment and engagement even for low-level interaction, likely driven by reducible outcome uncertainty, and propose that not all forms of stochastic randomness produce such engaging uncertainty.They also suggest curious and/or expectation-driven wanting as a potential overlooked engagement driver in games next to enjoyment and proximate positive experiences like need satisfaction.Overall, we hope to have contributed to moving games HCI research on juicy feedback from exploratory efect searching to careful and systematic theory-testing.

Figure 1 :
Figure 1: Structure of the two confrmatory structural equation models.

Figure 2 :
Figure 2: Screenshots of conditions: mid-swing (top row), hitting an enemy (middle row), and hitting with no enemy present (bottom row).Success-Dependent (+SD) impact efects are triggered only when hitting an enemy.Non-Success-Dependent (-SD) impact efects appear with every swing, regardless of a hit.Varied (+V) efects are selected randomly each time they are triggered.Non-Varied (-V) efects have a single efect.

Figure 4 :
Figure 4: Results of the two confrmatory structural equation models.Solid black paths reference primary hypothesis tests, with dashed lines for all other relations.All coefcients and 95% CIs (in brackets) are unstandardised efects: efects on curiosity, efectance, competence, and enjoyment represent points on a 7-pt Likert scale, and efects on free-choice playtime refer to minutes.Red paths were not originally included in the preregistered analysis plan.* < .05;** < .01;*** < .001.

Figure 5 :
Figure 5: Still frames depicting three consecutive attacks in sequence from left to right.The top row (STND) and bottom row (A-SD-V) highlight the contrast in visual occlusion between the two conditions.

Table 2 :
Descriptive statistics for player experience constructs and playtime across each condition (numeric).