We do as we construe: extended behavior construed as one task is executed as one cognitive entity

Participants were asked to keep a covert count of trials in threes (1–2–3–1–2–3) or fives (1–2–3–4–5–1–2–3–4–5) depending on an onscreen (‘steps of 3’ or ‘steps of 5’) instruction at the beginning of each block consisting of 6–35 trials (Fig. 1). This instruction remained on until participants pressed the spacebar. On each trial two numbers differing in value (between 0 and 99) and font size (Arial font 60 or 20) were displayed on each side of the fixation cross. A margin box appeared around, and at the same time as, each number display. Margin color determined the rule relevant on that trial—blue: choose the smaller value, green: choose the smaller font. Participants’ decisions were conveyed via buttons presses that were spatially congruent with their choice (Numpad 1 for left and Numpad 2 for right on a standard QWERTY keyboard number pad). In this and all subsequent experiments the stimuli remained onscreen until a response was made. Following the response there was a fixed interval (iTi) of 500 ms. before the onset of the next stimulus/margin.

To encourage participants to view the 3- or 5-trial groups as task episodes and to check whether they were oriented to their stage within them, a probe (“?”) appeared at the end of the 6–35 trial blocks. Participants were to key in the episode step number that they had executed immediately before the probe appeared. For example, in a 3-trial episode block that ended after seven trials, the correct response would be 1 (1–2–3–1–2–3–1-probe). Feedback was given on the accuracy of probe responses (high pitch tone for correct, low pitch tone for incorrect).

This and subsequent experiments were created in Visual Basic.net and run on a Dell computer with an 85 Hz refresh rate monitor situated at a comfortable distance from the participant. The experiment was conducted on an individual basis in a testing room designed to minimize visual and noise distraction. Participants were first given seven trials of practice on each of the two trial types (value and font judgments). They then completed a 30-trial practice block in which the stimulus margin color, signaling the currently relevant rule, changed randomly. They were then told to execute trials while keeping a count in threes. They practiced on two such blocks before proceeding to the main experimental session. Participants completed 70–120 blocks each consisting of 6–35 trials with 3- and 5-trial blocks being randomly interlaced. Participants were to respond as quickly and as accurately as possible.

Figure 2 and Tables 1 and 2 summarize the main results. As is evident, these concur with the key predictions: (1) significantly elevated RTs to the first trial of each construed episode; (2) longer RTs for 5-trial than 3-trial episodes; and (3) lack of significant switch costs when switches occurred across the boundaries between episodes. The first trial of the conceived episode took longest to execute (Table 1, and main effect of serial position in Table 2). Cohen’s d (effect size: mean difference/standard deviation) was 1.68 and 1.99 for 3 and 5-trial episodes, respectively. This trial 1 RT was higher for 5-trial compared with 3-trial episodes [95% CI of difference = (46, 113), Table 1 and interaction between the effects of serial position and episode length in Table 2, Cohen’s d = 1.3]. While the performance on switch trials was poorer than on repeat trials (Table 2, main effect of rule switch), this effect of rule switch differed across the serial positions within the episode (Table 2: Rule Switch × Serial Position). Specifically, as predicted no statistically significant switch costs were observed at trial 1 [RT: paired t₁₄ = 1.05, p = 0.3, 95% CI of difference = (− 19, 54); accuracy: paired t₁₄ = 1.1, p = 0.3, 95% CI of difference = (− 0.02, 0.01)].

It’s noteworthy that while the elevated RTs on switch trials compared to repeat trials were accompanied by the expected reduced accuracy, the elevated RTs on trial 1 compared with within-episode trials was, if anything, associated with greater accuracy. This was the case even though the within-episode switch cost was 92 ms, while the episode onset ‘cost’ (relative to within-episode trials) was in the region of 150–200 ms. This is to be expected, because response on switch trials got delayed because of control demands that was intrinsic to executing those trials, e.g., choosing the correct response in light of the newly relevant rule and overcoming interference from associations and configurations related to the previous rule. In contrast, trial 1 response got delayed because the episode-related program had to be assembled prior to the actual execution of trial 1, control demands related to choosing the correct response of trial 1 remained the same.

Stimuli and trial rules were identical to experiment 1 (Fig. 3). Before 3- and 5-trial task episode blocks participants saw ‘3 Step Tasks’ and ‘5 Step Tasks’, respectively, this instruction screen remained till the spacebar was pressed. Before the independent trial block participants saw ‘Independent Trials’, and the individual trials of this block had the number ‘1’ in the stimulus background. The 3- and 5-trial task blocks had 70 and 120 trials, respectively, and the independent trial blocks had 70 trials. The order of the three block types was random. Eighteen participants (11 females) did a total of 85–100 blocks. They were not explicitly told about the temporal grouping of the trials in the main experiment, and were asked to ignore the faded number in the stimulus background.

Tables 3 and 4 summarize the key results. The first trial of the 3- and 5-trial episodes took longest to execute (Cohen’s d 1.3 and 1.6; Table 3, and main effect of serial position in Table 4). While there was a main effect of rule switch on performance (main effect of rule switch in Table 4), this varied across serial positions (Table 4: rule switch × serial position). Specifically, as in the previous experiment, it was not significant on trial 1 [RT: paired t₁₇ = 1.7, p = 0.1, 95% CI of difference = (− 8, 71); accuracy: paired t₁₇ = 0.8, p = 0.4, 95% CI of difference = (− 0.01, 0.02)]. Unlike the previous experiment, the trial 1 RT for 5-trial episodes was not higher than 3-trial episodes [95% CI of difference = (− 39, 59); Table 4 interaction between serial position and episode length, Cohen’s d = 0.1]. Finally, as in the previous experiment, the strong difference in RT between the first and subsequent steps was not accompanied by changes in accuracy (Table 3, and main effect of serial position on accuracy in Table 4).

Trials conceived as parts of a task episode were executed faster than independent trials. As is evident from Table 3 the switch and repeat independent trials were higher than RTs on trials 2 and beyond of task episodes (t₁₇ > 5.5, p < 0.001; Cohen’s d > 1.61). Accuracies, however, were not significantly different. Thus, organization of trials into larger task episodes created a slower trial 1 but resulted in faster execution of subsequent trials. Was the price paid at trial 1 offset by gains at subsequent steps? We compared average RT and accuracy between blocks organized into task episodes with those that were not. Average RT on 3- and 5-trial task blocks was still significantly lower than on the independent trial blocks (t₁₇ > 4.1, p < 0.001, Cohen’s d = 0.82). This benefit also extended to switch control. Average reaction time switch cost within independent trials was greater than that amongst trials executed as parts of task episodes [176 vs 128 ms; t₁₇ = 2.1, p = 0.06, CI of difference = (− 1, 96), Cohen’s d = 0.49].

Improved performance within task episode trials could have been due to the participants becoming more fatigued during independent trial blocks due to the absence of slightly longer inter-trial intervals that marked task episode boundaries. To examine this we compared the initial 30 trials of the independent trial blocks with trials of the latter half of 3- and 5-trial episode blocks. Participants may be expected to be less tired during the former (as they had only executed a maximum of 30 trials) than the latter (where they would have executed between 35 and 60 trials in 3- and 5-trial episodes, respectively). However, results were identical to the above and participants were still faster during task episode (p < 0.01, Cohen’s d = 0.87) compared to independent trials.

Another possibility may be that during independent trial blocks participants were executing an extended continuous sequence, which may have resulted in attentional breaks. Trials following such breaks can be expected to have very high RTs. In contrast, while executing task episode participants got long temporal breaks after every 3 and 5 trials during episode blocks that could have refreshed and refocused their attention before the next bout of 3 or 5 trials. As per this account the more frequent presence of post-attentional break high RT trials in the continuous trial blocks may have increased their average RTs compared to task episode blocks. This explanation predicts that the RT distribution of independent trials should primarily differ from that of task episode trials in having an additional bump made of these high RT post-attentional break trials in the right sided tail of the distribution. Otherwise, other aspects of their RT distributions, e.g., the mode of their distributions (most frequent RT values), should be identical.

To plot RT distributions we first converted individual RTs to z-scores. We categorized all trials into four categories on the basis of the rule executed (value or font) and whether these rules were switch or repeat from the previous trial. The difference of every RT from the mean of its category was divided by the standard deviation of that category to get a z-score (zRT) corresponding to that RT. The blue lines in Fig. 4 show the distribution of zRTs from trials 2 and beyond of 3- (dotted blue) and 5-trial (dashed blue) episodes. The black line shows the distribution from the independent trial block. Note that while the zRT distributions of trials from 3- and 5-trial episodes are largely congruent, that of the independent trials is very different. Crucially, this difference is not limited to an additional bump in the tail of the distribution (corresponding to the purported post-attentional break very long RT trials), instead the entire distribution of independent trial zRTs is shifted to the right compared to the zRTs from within task episodes. In fact the distribution of zRTs from the independent trial blocks was largely congruent with those of trial 1 of 3- and 5-trial episodes (green plots). Hence, the above observation that trials (subsequent to trial 1), construed as parts of a task episode, were executed faster than identical trials, executed as independent entities, cannot be attributed to more frequent lapses in sustained attention during the independent trial blocks.

The methods here were identical to Experiment 2 with the exceptions that 3- and 5-trial episodes occurred randomly within the same block and there was no continuous block condition. The experimental session lasted half an hour during which participants did an average of 250 task episodes, which consisted of roughly equal number of 3- and 5-trial episodes. Twenty-two healthy participants (11 females) participated.

Tables 5 and 6 summarize the key results. Notably, the trial 1 RT was higher for 5-trial compared to 3-trial episodes [Tables 5 and 6, interaction between serial position and episode length, 95% CI of difference = (15, 96), Cohen’s d = 0.6]. Other results were largely a replication of the previous two experiments. The first step of the conceived episode took longest to execute (Table 5, and main effect of serial position in Table 6). This was the case for both switch and repeat trials at trial 1. Performance on switch trials was poorer than on repeat trials (Table 6 main effect of rule switch); however, the effect of rule switch was not the same across the trials making up the episode (Table 6: rule switch × serial position). Specifically, the switch cost was not significant at the first position [RT: paired t₂₁ = 0.89, p = 0.4, 95% CI of difference = (− 43, 109); accuracy: paired t₂₁ = 0.3, p = 0.7, 95% CI of difference = (− 1.4, 1.9)]. Finally, accuracies were not significantly different between the trial 1 and subsequent trials (main effect of serial position on accuracy in Table 6).

On each trial participants saw a centrally presented color word (‘Red’, ‘Blue’, ‘Green’, and ‘White’; in Arial font size 40). The word appeared in congruent font color on 60% of trials and in an incongruent color on the remainder (Fig. 5). Participants chose the color of the print from the two color words presented in Arial black font size 20 below (1 degree away from the center and ½ degree below it) by pressing a button spatially congruent with their choice - left: Numpad 1 (right index finger); right: Numpad 2 (right middle finger). One of the two bottom color words (allocated to left or right at random) always reflected the font color. When the font and word were incongruous, the other word always repeated the stimulus word. On congruous trials the other word was selected randomly from the remaining colors. Between these two options a partially (70%) transparent digit appeared in black. The stimuli remained on screen till a response was made. Erroneous responses elicited a low-pitched feedback tone.

At the start of 4-trial blocks the instruction screen mentioned ‘4—step blocks’, it remained on until the spacebar was pressed. Such blocks consisted of 96 trials. The instruction for the independent trial block mentioned ‘independent trials’. Such blocks consisted of 48 trials. These two blocks were interleaved. Eighteen participants (9 female) each completed 12 blocks.

Tables 7 and 8 summarize the key results. As before, the first trial of the episode took longest to execute [Table 7, and main effect of serial position in Table 8, 95% CI of difference = (54, 146), Cohen’s d = 1.1]. This was the case for both congruent and incongruent trial 1. Expectedly, the performance on incongruent trials was poorer than on congruent trials (Table 8 main effect of Stroop cost), but crucially it did not vary across trials of the episode (Table 8: stroop cost × serial position). Specifically, unlike switch cost in the previous experiments, the cost of incongruence did not disappear on trial 1 of the episode [RT: one-sample t₁₇ = 8.8, p = < 0.001, 95% CI of difference = (127, 206); accuracy: one-sample t₁₇ = 4.2, p ≤ 0.001, 95% CI of difference = (1.7, 5.0)].

RTs on incongruent and congruent trials of the independent trial blocks were higher than the incongruent and congruent trials of task episodes. The average RT on trials executed as parts of task episodes was lower than for independent trials even when RTs of trial 1 of task episodes were included in comparison [congruent trials: t₁₇ = 7, p < 0.001, 95% CI of difference = (89, 165); incongruent trials: t₁₇ = 4.9, p < 0.001, 95% CI of difference = (69, 175)]. While there was a general RT advantage for task episode trials, Stroop cost on RT (RT on incongruent trials − RT on congruent trials) did not differ between task episode and independent trials [t₁₇ = 0.3, p = 0.8, 95% CI of difference = (− 27, 37)]. However, Stroop cost on accuracy (accuracy on Congruent trials − accuracy on incongruent trials) was significantly higher on independent compared to task episode trials [t₁₇ = 2.6, p = 0.02, 95% CI of difference = (0.5, 5.5)].

The results affirmed the general pattern of the previous results in showing elevated RTs on the first trial of episodes in the context Stroop task. They confirmed the prediction that Stroop cost would be present on trial 1 and have the same value as on other trials of the episode. They also replicated the findings of experiment 2 that trials executed as parts of a task episode got executed faster and better controlled than trials executed as independent tasks.

Apart from the additional outer margin, individual trials were identical to Experiments 1–3 (speeded decision as to which of the two numbers had the smaller value or font size as indicated by the color of the inner margin). 27 participants (17 females) did two experimental sessions each lasting 16 min, with a period of rest in between. Within a session the two kinds of conceived episodes occurred randomly.

Key results of the previous experiments were again replicated (Tables 9, 10). The first trial of the episode took longest to execute (F_2,52 = 28.1, p < 0.001). This trial 1 RT was longer before longer episodes [95% CI of difference = (14, 39), Cohen’s d = 0.81; t₂₆ = 4.2, p < 0.001]. Switch cost was substantially reduced at the first trial of the episode (Smaller episodes: F_2,52 = 18; Longer episodes: F_6,156 = 8.1, p < 0.001 for both). The execution of trials construed as parts of a larger task episode again showed key behavioral signatures that suggested that these trials were executed not as independent entities but through a common program. This was the case even though participants were unaware of the number and sequence of task items they would execute as well as when those items would be executed and the precise duration of the episode.

Note that the trial 1 RT in this experiment was delayed by only 64 ms compared to subsequent trial RTs. In comparison, in experiments 1 to 3 this was 175–250 ms. This is likely to be due to greater uncertainty related to the timings of later trials compared to trial 1. While trial 1 would predictably occur 2 s from the offset of the previous task episode margin, subsequent trials would appear variably between 50 ms to 5 s after the previous trial. This can be expected to decrease trial 1 RT while increasing RTs on subsequent trials (Table 9).

Stimuli and trial rules were identical to Experiment 1. The key differences in this experiment concerned iTis and the additional requirement to press a ‘task end’ button (key ‘Z’ on QWERTY keyboard) at the completion of each task episode. In 3-trial-short and 5-trial-short episodes, the stimuli for the next trial followed immediately after the response to the previous trial (Fig. 7a, b). Within 3-step-long episodes the onset of the next trial occurred 2 s after the response to the previous trial during which the monitor was blank (Fig. 7c). In all conditions there was a 2 s delay between the end of an episode and the beginning of the next, hence trial 1 of all task episodes was preceded by identical iTis. The experiment consisted of blocks made of 60 trials. Nineteen Participants (11 females) completed 30–40 blocks. Each of these blocks consisted of one of the three task episodes (i.e., 20 task episodes in 3-trial-short and 3-trial-long blocks, and 12 task episodes in 5-trial-short blocks). The type of task episode to be executed in the block was cued by the instruction screen at the beginning of the block. Different task episode blocks were randomly interleaved with each other.

As predicted trial 1 RT of 3-trial-long episodes was slower than trial 1 RT of 3-trial-short episodes [paired t₁₈ = 2.9, p ≤ 0.01, 95% CI of difference = (17, 103), Cohen’s d = 0.67]. Thus, the magnitude of the episode-related program was affected not just by the number of constituent trials or steps but also by the temporal duration of the episode. Second, difference in RTs between 3-trial-short and 5-trial-short episodes was not limited to trial 1, instead RTs on all trials of 3-trial-short episodes was slower than the corresponding trials of 5-trial-short episodes. Hence, the main effect of episode was significant (F_1,18 = 10.5, p = 0.004) but its interaction with serial position was not (F_2,36 = 0.4, p = 0.6), suggesting that maintaining larger program did decrease the amount of cognitive reserves available for trial execution. In other aspects results were identical to the previous experiments (Tables 11 and 12).

Most attempts at modeling hierarchical behavior follow the assumption that execution of hierarchical behavior requires a parallel hierarchy of cognitive processing units (Cooper & Shallice, 2000; Estes, 1972; Norman & Shallice, 1986; Rumelhart & Norman, 1982). Typically, these models have discrete units for temporally smaller, lower level acts (e.g., ‘pick spoon up’) which are activated by units corresponding to temporally longer, higher level acts (e.g., ‘add sugar from packet’), which in turn are activated by the unit related to the entire task/goal (e.g., ‘prepare coffee’). Such models, however, can only work for situations, where identity and position of component steps of the task are explicitly known. In addition, the deterministic relation between their higher and lower level units creates a problem for generalizing actions across different task contexts as well as utilizing shared aspects of related actions. For example, across different coffee making situations sugar may be added before or after milk. Because the lower level acts making up the task of preparing coffee have fixed identities and ordinal positions, a small change such as this may require a totally different coffee making routine. In contrast to these models, the current study evidenced hierarchical cognition in situations, where higher level entities did not determine the identity and position of lower level steps.

Botvinick and Plaut (2004) suggested that a hierarchically organized processing system is not a prerequisite for hierarchical tasks. Instead, their model had perceptual input layer connected to the motor output layer through an internal hidden layer with recurrent connections between its component units. This allowed the units of this layer to not only link the perceptual units to motor units but to also represent the current behavioral context generated by the previous step. This got integrated with incoming information from the perceptual layer to choose the correct context appropriate next step. Through practice, the model learnt to represent the sequential structure of the task in the patterns of activations across units of the internal layer.

Contrary to the predictions of this model, the current results suggested that even behavioral sequences that have no external hierarchical structure or pre-existing internal hierarchical task representations get hierarchically executed if those sequences are conceived as defined task entities. Because this model executes one step at a time, in absence of any entity that corresponds to the overarching task episode, it cannot predict long step 1 RTs that are longer before longer task episodes. Likewise, because processings related to individual steps are not subsumed by anything related to goal or task episode, this model is unlikely to show absent step-related switch costs at episode boundaries.

Existing accounts of cognition have proposed constructs like plans, scripts, schema and frames that subsume and control extended behavior but can be selected and instantiated as one entity (Cooper & Shallice, 2000; Miller et al., 1960; Minsky, 1974; Schank & Abelson, 1977). While the knowledge of the identity and sequence of component steps is a requisite in these accounts, the current study suggested that even unpredictable task episodes, where such knowledge is absent, are executed through subsuming cognitive entities that correspond to extended behavior. Such entities may be better thought of as embodying the executive commands for organizing, controlling and executing task episodes, and not as mere representations of the steps to be executed.

Hierarchical instantiation of executive commands is well known in motor cognition (Henry & Rogers, 1960; Keele, Cohen, & Ivry, 1990; Rosenbaum et al., 1983; Rosenbaum, Cohen, Jax, Weiss, & van der Wel, 2007). Motor actions typically consist of a sequence of smaller acts, e.g., articulating a word may consist of a sequence of phonemic articulations. Instead of individually instantiating higher level executive commands for each of these component acts, a motor program embodying the commands for the entire sequence is instantiated in one-go. This then unfolds across time into the seamless chain of small acts making up the overall action. While motor programs have typically been characterized in situations, where behavior seems to get executed ballistically, programs in general need not be limited to such instances.

When participants are given a memorized list to execute (e.g., CCSS, where C and S may, respectively, stand for color and shape decisions to be made across sequential trials), such that only the sequence of task items to be executed across time, and not the actual sequence of motor acts to be made, is known in advance, the behavior evidences a cognitive program, and not merely a recall of the sequence in working memory. Hence, not only is the item 1 RT high, it remains high even when the same sequence is iteratively executed, suggesting that something additional has to be done to the sequence already in working memory. Furthermore, this item 1 RT correlates with the number of item-level switches in the sequence (e.g., Schneider & Logan, 2006; Desrochers & Badre, 2015). Again, merely recalling a sequence like CSCS (with three item-level switches) need not take longer than a sequence like CCSS (with one item-level switch), but prospectively preparing for item switches to come may lead to longer item 1 RTs. The cognitive entity assembled at trial 1 thus embodies prospective control-related commands that will be needed during the ensuing episode, and has to be reassembled every time the episode has to be executed, even when the same episode is being iteratively executed.

Such programs related to memorized task-sequence execution did not directly translate into behavior, instead they translated into the sequence of rule-related cognitive set changes through which the correct motor act was selected in response to the stimuli (see discussion of Schneider & Logan, 2006). Analogously, the program in the current study could have translated into a sequence of control-related cognitive changes that facilitated the search for the correct rule-related cognitive set through which the correct motor act was subsequently selected.

The instantiation of attention and control is always linked to the goal in operation. Goals, typically, require some extended period of thought and behavior (i.e., task episodes) for their completion. It is possible that the executive commands for instantiating the myriad higher level goal-related control changes in cognition for a task episode is executed in one-go. If, for example, the goal requires searching for visual targets across 20 s, the higher level commands that instantiate the search and its related attentional changes in the brain need not be instantiated through separate commands every millisecond or every second. It is possible that these may be achieved through a program that subsumes the search period bringing about relevant changes across time.

Every task episode requires organization of cognition across time. Various irrelevant processes and representations need to be cleared out and maintained in abeyance across time so that they don’t compete for cognitive reserves and disrupt task execution. Various task-relevant learnings, memories, skills, dispositions, knowledge and expectancies, and the corresponding configurational changes in various perceptual, attentional, mnemonic, and motor processes may be brought to fore (Bartlett, 1932; Logan & Gordon, 2001; Mayr & Keele, 2000; Miller et al., 1960; Rogers & Monsell, 1995). The program may achieve these widespread changes across time. For example, in current task episodes processing related to mind-wandering, ongoing unconscious goals, task-irrelevant sensory and motor processing etc. had to be relegated. At the same time, the predictiveness of the episode was to be utilized to make anticipatory changes; e.g., knowledge that responses would be right handed, visual attention limited to area around fixation, along with an implicit idea of iTis, may have been used to increase preparations and attention at times when a stimulus was expected and decrease when iTi was expected.

Entities that are selected and instantiated as one unit but correspond to a sequence of processes and actions have been proposed and evidenced in other domains. Ullman (1984) suggested that the immediate and effortless perception of spatial properties and relations of objects may be achieved through automatic instantiation of a visual routine—a fixed set of elemental processes on early visual representations (for a related construct see Sprites, Cavanagh, Labianca, & Thornton, 2001). Logan (1999) suggested that attention during perceptual tasks be considered as a routine (attentional routine) consisting of the set of processes that intervene between perception and reaching a goal relevant propositional conclusion, e.g., during identification of object X this routine would intervene between early perceptual processes and the reaching of conscious conclusion that the ‘visible object is X’. Component motor acts of a larger motor action come about through a common program assembled at the beginning of execution (Henry & Rogers, 1960; Schmidt, 1975). Fan et al. (2012) found that creating a means-ends relation between two sequential tasks created an additional delay in the execution of the first task, which they interpreted as related to the time taken by the supervisory control system in chaining subunits into longer goal-directed behavior (see also Sacker & Dehaene, 2009). Botvinick et al. (2009) proposed that extended sequence of behavior (e.g., ‘open laptop’, ‘mouse to browser icon’, ‘double-click’, ‘enter URL’, among others) may be chunked into larger routines (‘check email’) through hierarchical reinforcement learning, wherein the reward signals strengthen the entire sequence of behavior, and not just one act, allowing instantiation of the entire sequence as a routine. In comparison, current results suggested that abstract, unpredictable, higher level sequential behavior can be executed as one-routine in absence means-ends relation and without receiving hierarchical reinforcing reward signals if that behavior is construed as one task entity.

We suggest that our phenomenal experience of executing task identities (e.g., ‘prepare breakfast’, ‘boil water’) as a whole, and not of individually executing its very many components, may be real (Vallacher & Wegner, 1989). These typically correspond to an episode and consist of numerous sequential acts (‘prepare breakfast’, ‘boil water’). We suggest that our phenomenal experience of executing such task identities as a whole, and not of individually executing its very many components, may be real. At the beginning of task episodes we prepare for the entire episode as one unit. This preparation is embodied in a program. The actual contents of this program, and the prospective preparation it embodies, may depend on the familiarity and predictability of the task episode related to that goal. Completely predictable tasks may be prepared down to the level of individual motor acts to be made across time (e.g., Henry & Rogers, 1960; Rosenbaum et al., 1983). During less predictable task episodes like memorized task-sequence execution the program may involve preparing for the sequence of stimulus—response rules to be applied across time (e.g., Schneider & Logan, 2006). During unpredictable task episodes, the program may only instantiate goal-directed attention and other control processes, and their related changes in cognition that enable the search for implementation strategies that will culminate in goal completion.