The primary comparison in the paper is between the three models varying granularity (prototype, exemplar, and adaptive, the last one being the model that infers its own within-class granularity α). The fourth model, shared, varies γ (cross-class sharing). Data are read from the five merged cohort tables in ../model_output/; figures are written to figures/.

library(tidyverse)
library(brms)

dir.create("figures", showWarnings = FALSE)

# Primary comparison = three granularity models (prototype, exemplar, adaptive).
primary_models <- c("prototype", "exemplar", "adaptive")
model_palette  <- c("prototype" = "#39517D", "exemplar" = "#6E685D",
                    "adaptive"  = "#D0742D", "shared"   = "#4E9B8F")
proto_palette  <- c("A" = "#284958", "B" = "#C3824B")
trial_block_scale <- scale_x_continuous(breaks = 0:10, limits = c(0, 10))

# Learner inclusion is a screening decision: keep a
# participant if P(exception-accuracy slope > 0) exceeds PRIMARY_THRESHOLD. 0.75
# sits in the natural gap in the ranked improvements (§2a)
PRIMARY_THRESHOLD <- 0.75
SENSITIVITY_THRESHOLDS <- c(0.75, 0.80, 0.85, 0.90, 0.95)

Import

The five merged cohort tables cover 43 participants and four models (prototype, exemplar, adaptive, shared).

model_output_dir <- "../model_output"

read_table <- function(name)
  read_csv(file.path(model_output_dir, paste0(name, ".csv")), show_col_types = FALSE)

resp_raw   <- read_table("responses_and_model_predictions")
ll_raw     <- read_table("model_segment_log_likelihoods")
params_raw <- read_table("fitted_params")
member_raw <- read_table("cluster_membership")
consen_raw <- read_table("cluster_consensus")

Stimulus bookkeeping and derived variables

a_words <- c("gyvuco", "zyvuco", "gybuco", "gyveco", "gyvuto", "gyvuci")
b_words <- c("zabeti", "gabeti", "zybeti", "zaveti", "zabuti", "zabeto")
exceptions <- c("gavuco", "zabeci")

# suffix classes (the class a word's CORRECT plural belongs to)
re_class <- c(b_words, "gavuco")   # correct suffix -re
ma_class <- c(a_words, "zabeci")   # correct suffix -ma
word_class <- function(w) ifelse(w %in% re_class, "re", "ma")

add_stim_vars <- function(d) {
  d %>% mutate(
    stimulus_type = relevel(
      factor(ifelse(word %in% exceptions, "exception", "prototypical")),
      ref = "prototypical"),
    correct_suffix = case_when(
      word == "gavuco" ~ "re", word == "zabeci" ~ "ma",
      word %in% a_words ~ "ma", word %in% b_words ~ "re"),
    prototype = case_when(
      word == "gavuco" ~ "A", word == "zabeci" ~ "B",
      word %in% a_words ~ "A", word %in% b_words ~ "B"),
    human_correct = as.integer(human_response == correct_suffix)
  )
}

resp <- resp_raw %>%
  mutate(trial_block = Trial / 2) %>%
  add_stim_vars()
ll     <- ll_raw
params <- params_raw
consen <- consen_raw
member <- member_raw

# ALL model-prediction rows (needed for §6); adds fit quantities.
preds <- resp %>%
  filter(!is.na(model)) %>%
  mutate(
    p_human   = ifelse(human_response == "re", P_re, P_ma),
    p_correct = ifelse(correct_suffix == "re", P_re, P_ma),
    log_lik   = log(pmax(p_human, 1e-10)),
    is_right  = as.integer((human_response == "re" & P_re > P_ma) |
                           (human_response == "ma" & P_ma > P_re))
  )

# primary prediction table: three granularity models only.
preds_p <- preds %>%
  filter(model %in% primary_models) %>%
  mutate(model = factor(model, levels = primary_models))

# One row per actual observation (dedupe the 4x model duplication), all 20 blocks.
human <- resp %>%
  distinct(PROLIFIC_PID, Trial, trial_block, word,
           human_response, stimulus_type, prototype, correct_suffix, human_correct)

Fitted parameters for the three granularity models (α applies to adaptive only):

params %>%
  filter(model %in% primary_models) %>%
  mutate(model = factor(model, levels = primary_models)) %>%
  group_by(model) %>%
  summarise(mean_mu = mean(mu), mean_guessing = mean(guessing),
            mean_alpha = mean(mean_alpha, na.rm = TRUE), .groups = "drop")
## # A tibble: 3 × 4
##   model     mean_mu mean_guessing mean_alpha
##   <fct>       <dbl>         <dbl>      <dbl>
## 1 prototype   17.5          0.272    NaN    
## 2 exemplar     1.81         0.151    NaN    
## 3 adaptive    13.5          0.237      0.328

1. Raw participant data (everyone)

The by-word plot shows the progressive polarisation of prototypical stimuli and the slower inversion of the exceptions. The table beneath ranks stimuli by mean accuracy.

databyword_all <- human %>%
  filter(Trial %% 2 == 0) %>%
  group_by(trial_block, prototype, stimulus_type, word) %>%
  summarise(mean.acc.binary = mean(human_correct), .groups = "drop") %>%
  mutate(mean.str1.hits = case_when(
    stimulus_type == "exception" & prototype == "A" ~ 1 - mean.acc.binary,
    stimulus_type == "exception" & prototype == "B" ~ mean.acc.binary,
    prototype == "A" ~ mean.acc.binary,
    TRUE ~ 1 - mean.acc.binary))

ggplot(databyword_all, aes(trial_block, mean.str1.hits,
                           group = word, color = prototype, shape = stimulus_type)) +
  geom_point() +
  geom_line(data = filter(databyword_all, stimulus_type == "prototypical"),
            linetype = "dashed", alpha = 0.3) +
  geom_line(data = filter(databyword_all, stimulus_type == "exception")) +
  scale_color_manual(values = proto_palette) +
  trial_block_scale +
  labs(y = "% 'gyvuco' suffix", shape = "is.exception", x = "Trial block") +
  theme_minimal(22)

ggsave("figures/raw-data-everyone.pdf", width = 9, height = 7)

databyword_all %>%
  group_by(word, stimulus_type, prototype) %>%
  summarise(x = mean(mean.acc.binary), .groups = "drop") %>%
  arrange(x)
## # A tibble: 14 × 4
##    word   stimulus_type prototype     x
##    <chr>  <fct>         <chr>     <dbl>
##  1 gavuco exception     A         0.407
##  2 zabeci exception     B         0.563
##  3 zabeto prototypical  B         0.570
##  4 gyvuci prototypical  A         0.693
##  5 gyvuto prototypical  A         0.763
##  6 zyvuco prototypical  A         0.777
##  7 gyvuco prototypical  A         0.798
##  8 zabuti prototypical  B         0.8  
##  9 gybuco prototypical  A         0.805
## 10 gyveco prototypical  A         0.807
## 11 zybeti prototypical  B         0.807
## 12 gabeti prototypical  B         0.821
## 13 zabeti prototypical  B         0.849
## 14 zaveti prototypical  B         0.856

2. Learner identification and threshold sensitivity

We keep participants who show improvement on the exceptions: a logistic model on exception trials with a by-participant random slope for centred Trial. We read P(slope > 0) for each participant off the posterior draws, so any threshold applies from the single fit.

exceptions_df <- human %>%
  filter(stimulus_type == "exception") %>%
  mutate(trial_c = Trial - mean(Trial))
learning_model <- brm(
  human_correct ~ 0 + (trial_c | PROLIFIC_PID),
  data = exceptions_df, family = bernoulli(),
  cores = 4, iter = 4000, silent = 2, refresh = 0
)
# With no population-level term, coef() == ranef() (the random slope IS the slope).
sl_draws <- coef(learning_model, summary = FALSE)$PROLIFIC_PID[, , "trial_c"]

learner_status <- tibble(
  PROLIFIC_PID = colnames(sl_draws),
  slope = colMeans(sl_draws),
  p_pos = colMeans(sl_draws > 0)          # P(slope > 0)
) %>% arrange(desc(p_pos))

pick_learners <- function(th = PRIMARY_THRESHOLD)
  learner_status$PROLIFIC_PID[learner_status$p_pos > th]
learner_ids <- pick_learners(PRIMARY_THRESHOLD)

# retained count at each threshold
tibble(threshold = SENSITIVITY_THRESHOLDS) %>%
  mutate(n_retained = map_int(threshold, ~ length(pick_learners(.x))))
## # A tibble: 5 × 2
##   threshold n_retained
##       <dbl>      <int>
## 1      0.75         17
## 2      0.8          17
## 3      0.85         13
## 4      0.9          12
## 5      0.95         10

2a. Is there a natural gap between learners and non-learners?

The cut-off falls in a natural break in the data. Ranked by P(slope > 0), the retained participants sit above a visible gap before the exceedances decay toward zero. The table lists the largest gaps and the plot marks where the 0.75 cutoff is.

gap_df <- learner_status %>%
  mutate(rank = row_number(), gap_below = p_pos - lead(p_pos))

gap_df %>% filter(!is.na(gap_below)) %>%
  slice_max(gap_below, n = 5) %>%
  transmute(between_rank = paste0(rank, "-", rank + 1),
            upper = round(p_pos, 3),
            lower = round(p_pos - gap_below, 3),
            gap = round(gap_below, 3))
## # A tibble: 5 × 4
##   between_rank upper lower   gap
##   <chr>        <dbl> <dbl> <dbl>
## 1 35-36        0.373 0.279 0.094
## 2 36-37        0.279 0.22  0.06 
## 3 41-42        0.103 0.046 0.057
## 4 17-18        0.803 0.75  0.054
## 5 34-35        0.426 0.373 0.053
ggplot(gap_df, aes(rank, p_pos)) +
  geom_hline(yintercept = PRIMARY_THRESHOLD, linetype = "dashed",
             color = "#C3824B", linewidth = 1) +
  geom_point(aes(color = p_pos > PRIMARY_THRESHOLD), size = 2.4) +
  scale_color_manual(values = c(`TRUE` = "#39517D", `FALSE` = "grey65"), guide = "none") +
  scale_y_continuous(limits = c(0, 1), breaks = seq(0, 1, 0.25)) +
  labs(x = "Participant (ranked by P(slope>0))", y = "P(exception slope > 0)",
       subtitle = paste0("dashed = ", PRIMARY_THRESHOLD, " cut  |  retained = ",
                         length(learner_ids), " of ", nrow(learner_status))) +
  theme_minimal(16)

ggsave("figures/learner-gap.pdf", width = 9, height = 5)

2b. Threshold sensitivity (§4.4)

For each learner set (all participants, then P(slope>0) > 0.75 / 0.80 / 0.85 / 0.90 / 0.95) we recompute the best-fitting-model counts and the summed log-likelihood ordering of the three models. The ordering is stable across thresholds: prototype is always worst, and adaptive/exemplar trade the top spot depending on the aggregate-vs-per-participant tension discussed in §3.

best_fit_dist <- function(ids) {
  preds_p %>% filter(PROLIFIC_PID %in% ids) %>%
    group_by(PROLIFIC_PID, model) %>% summarise(total_ll = sum(log_lik), .groups = "drop") %>%
    group_by(PROLIFIC_PID) %>% slice_max(total_ll, n = 1, with_ties = FALSE) %>%
    ungroup() %>% count(best_model = model, .drop = FALSE)
}
ll_ordering <- function(ids) {
  preds_p %>% filter(PROLIFIC_PID %in% ids) %>%
    group_by(model) %>% summarise(summed_ll = sum(log_lik), .groups = "drop") %>%
    arrange(desc(summed_ll))
}
sets <- c(list(`keep all` = unique(preds_p$PROLIFIC_PID)),
          setNames(lapply(SENSITIVITY_THRESHOLDS, pick_learners),
                   paste0("p>", format(SENSITIVITY_THRESHOLDS))))

# best-fit model counts by learner set
imap(sets, ~ best_fit_dist(.x) %>% mutate(set = .y)) %>%
  bind_rows() %>%
  pivot_wider(names_from = set, values_from = n, values_fill = 0)
## # A tibble: 3 × 7
##   best_model `keep all` `p>0.75` `p>0.80` `p>0.85` `p>0.90` `p>0.95`
##   <fct>           <int>    <int>    <int>    <int>    <int>    <int>
## 1 prototype           6        0        0        0        0        0
## 2 exemplar           18        4        4        2        2        0
## 3 adaptive           19       13       13       11       10       10
# summed-LL model ordering by learner set
imap(sets, ~ ll_ordering(.x) %>% mutate(set = .y, rank = row_number())) %>%
  bind_rows() %>% select(set, rank, model, summed_ll) %>% arrange(set, rank)
## # A tibble: 18 × 4
##    set       rank model     summed_ll
##    <chr>    <int> <fct>         <dbl>
##  1 keep all     1 exemplar     -3109.
##  2 keep all     2 adaptive     -3182.
##  3 keep all     3 prototype    -3453.
##  4 p>0.75       1 exemplar     -1056.
##  5 p>0.75       2 adaptive     -1062.
##  6 p>0.75       3 prototype    -1286.
##  7 p>0.80       1 exemplar     -1056.
##  8 p>0.80       2 adaptive     -1062.
##  9 p>0.80       3 prototype    -1286.
## 10 p>0.85       1 adaptive      -710.
## 11 p>0.85       2 exemplar      -736.
## 12 p>0.85       3 prototype     -911.
## 13 p>0.90       1 adaptive      -628.
## 14 p>0.90       2 exemplar      -654.
## 15 p>0.90       3 prototype     -828.
## 16 p>0.95       1 adaptive      -521.
## 17 p>0.95       2 exemplar      -546.
## 18 p>0.95       3 prototype     -684.
tibble(threshold = names(sets), n_learners = lengths(sets)) %>%
  mutate(threshold = factor(threshold, levels = names(sets))) %>%
  ggplot(aes(threshold, n_learners)) +
  geom_col(fill = "#39517D") +
  geom_text(aes(label = n_learners), vjust = -0.4) +
  labs(x = "Inclusion threshold", y = "Participants retained") +
  theme_minimal(16)

ggsave("figures/threshold-sensitivity.pdf", width = 8, height = 5)

2c. Restrict to learners

preds_pL <- preds_p %>% filter(PROLIFIC_PID %in% learner_ids)   # 3 models, learners
preds_L  <- preds   %>% filter(PROLIFIC_PID %in% learner_ids)   # all 4 models, learners (§6)
human_L  <- human   %>% filter(PROLIFIC_PID %in% learner_ids)
ll_L     <- ll      %>% filter(PROLIFIC_PID %in% learner_ids)
consen_L <- consen  %>% filter(PROLIFIC_PID %in% learner_ids)

Raw data, learners only:

databyword <- human_L %>%
  filter(Trial %% 2 == 0) %>%
  group_by(trial_block, prototype, stimulus_type, word) %>%
  summarise(mean.acc.binary = mean(human_correct), .groups = "drop") %>%
  mutate(mean.str1.hits = case_when(
    stimulus_type == "exception" & prototype == "A" ~ 1 - mean.acc.binary,
    stimulus_type == "exception" & prototype == "B" ~ mean.acc.binary,
    prototype == "A" ~ mean.acc.binary,
    TRUE ~ 1 - mean.acc.binary))

ggplot(databyword, aes(trial_block, mean.str1.hits,
                       group = word, color = prototype, shape = stimulus_type)) +
  geom_point() +
  geom_line(data = filter(databyword, stimulus_type == "prototypical"),
            linetype = "dashed", alpha = 0.3) +
  geom_line(data = filter(databyword, stimulus_type == "exception")) +
  scale_color_manual(values = proto_palette) +
  trial_block_scale +
  labs(y = "% -ma suffix", shape = "is.exception", x = "Trial block") +
  theme_minimal(25)

ggsave("figures/raw-data-successful-learners.pdf", width = 9, height = 5)

Accuracy over time by stimulus type (positive Trial slope; large negative exception offset; positive interaction — exceptions start worse but improve faster):

mod1 <- brm(human_correct ~ Trial * stimulus_type + (1 | PROLIFIC_PID),
            data = human_L, family = bernoulli(),
            cores = 4, iter = 4000, silent = 2, refresh = 0)
summary(mod1)
##  Family: bernoulli 
##   Links: mu = logit 
## Formula: human_correct ~ Trial * stimulus_type + (1 | PROLIFIC_PID) 
##    Data: human_L (Number of observations: 4760) 
##   Draws: 4 chains, each with iter = 4000; warmup = 2000; thin = 1;
##          total post-warmup draws = 8000
## 
## Multilevel Hyperparameters:
## ~PROLIFIC_PID (Number of levels: 17) 
##               Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
## sd(Intercept)     0.79      0.16     0.55     1.17 1.00     1369     2657
## 
## Regression Coefficients:
##                              Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS
## Intercept                        0.94      0.20     0.53     1.35 1.00     1252
## Trial                            0.07      0.01     0.06     0.09 1.00     4439
## stimulus_typeexception          -1.68      0.19    -2.05    -1.31 1.00     3063
## Trial:stimulus_typeexception     0.06      0.02     0.03     0.09 1.00     2852
##                              Tail_ESS
## Intercept                        1715
## Trial                            5453
## stimulus_typeexception           4805
## Trial:stimulus_typeexception     4379
## 
## Draws were sampled using sampling(NUTS). For each parameter, Bulk_ESS
## and Tail_ESS are effective sample size measures, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).
conditional_effects(mod1, effect = "Trial:stimulus_type")[[1]] %>%
  mutate(trial_block = Trial / 2) %>%
  ggplot(aes(trial_block, estimate__, fill = stimulus_type, color = stimulus_type)) +
  geom_ribbon(aes(ymax = upper__, ymin = lower__), alpha = .5, color = NA) +
  geom_line(linewidth = 1.4) +
  trial_block_scale +
  labs(y = "is.correct", x = "Trial block",
       fill = "is.exception", color = "is.exception") +
  theme_minimal(22)

ggsave("figures/model-inference.pdf", width = 9, height = 7)

3. Aggregate model fit (three granularity models)

Predicted probability of the intended suffix. All models start strong on prototypes; the prototype model’s exception accuracy declines with training, while the adaptive model’s improves.

preds_pL %>%
  filter(Trial %% 2 == 0) %>%
  group_by(model, stimulus_type, trial_block) %>%
  summarise(avg_pcorrect = mean(p_correct), .groups = "drop") %>%
  ggplot(aes(trial_block, avg_pcorrect, color = model, linetype = stimulus_type)) +
  geom_point() + geom_line() +
  scale_color_manual(values = model_palette) +
  trial_block_scale + theme_minimal(25) +
  labs(x = "Trial block", y = "Predicted probability of\nintended suffix",
       color = "Model", linetype = "Stimulus type")

ggsave("figures/overall-model-prediction.pdf", width = 9, height = 5)

Summed log-likelihood by stimulus type. The three models are close on prototypes; on exceptions the prototype model gets progressively less accurate as it grows confidently wrong.

preds_pL %>%
  filter(Trial %% 2 == 0) %>%
  group_by(model, trial_block, stimulus_type) %>%
  summarise(mean_ll = sum(log_lik), .groups = "drop") %>%
  ggplot(aes(trial_block, mean_ll, colour = model)) +
  geom_point() + geom_line() +
  scale_color_manual(values = model_palette) +
  trial_block_scale +
  facet_wrap(~stimulus_type, scales = "free_y") +
  labs(y = "Summed Log Likelihood", x = "Trial block") +
  theme_minimal(22)

ggsave("figures/ll-by-stimulus.pdf", width = 11, height = 7)

Log-likelihood is asymmetric — it punishes confident errors without bound but bounds the reward for confident hits — so it rewards hedging near 0.5:

tibble(p = seq(0.001, 0.999, length.out = 100)) %>%
  mutate(value = log(p)) %>%
  ggplot(aes(p, value)) +
  geom_line(size = 1) +
  geom_hline(yintercept = 0, color = "black", size = 0.2) +
  geom_vline(xintercept = 0.5, color = "gray", linetype = "dotted") +
  labs(x = "Probability (p)", y = "Log-Likelihood Value") +
  theme_minimal(22)

ggsave("figures/log-likelihood-plot.pdf", width = 7, height = 5)

Mean probability assigned to the human’s actual choice (symmetric complement to LL): the adaptive model leads on exceptions.

preds_pL %>%
  filter(Trial %% 2 == 0) %>%
  group_by(trial_block, model, stimulus_type) %>%
  summarise(avg_p_human = mean(p_human), .groups = "drop") %>%
  ggplot(aes(trial_block, avg_p_human, color = model, linetype = stimulus_type)) +
  geom_point() + geom_line() +
  scale_color_manual(values = model_palette) +
  trial_block_scale +
  labs(y = "Mean P(human choice)", x = "Trial block",
       linetype = "Stimulus type", color = "Model") +
  theme_minimal(25)

ggsave("figures/probability-of-human-choice.pdf", width = 9, height = 5)

3. Fit summary by model and stimulus type

Mean log-likelihood and mean probability assigned to the human’s response (higher is better), plus MAP accuracy, by model and stimulus type. This is the source for the model-fit table in the paper: the three models are near-identical on the prototypical stimuli, while on the exceptions the adaptive model leads on every measure and the prototype model does poorly.

preds_pL %>%
  group_by(model, stimulus_type) %>%
  summarise(mean_ll = mean(log_lik),
            mean_p_human = mean(p_human),
            acc_MAP = mean(is_right), .groups = "drop") %>%
  arrange(stimulus_type, desc(mean_p_human))
## # A tibble: 6 × 5
##   model     stimulus_type mean_ll mean_p_human acc_MAP
##   <fct>     <fct>           <dbl>        <dbl>   <dbl>
## 1 adaptive  prototypical   -0.426        0.759   0.826
## 2 exemplar  prototypical   -0.415        0.740   0.826
## 3 prototype prototypical   -0.433        0.727   0.826
## 4 adaptive  exception      -0.571        0.624   0.691
## 5 exemplar  exception      -0.615        0.579   0.644
## 6 prototype exception      -1.18         0.396   0.35

4. Individual differences (three granularity models)

Overall best-fitting model per participant (the table below): no one is best fit by the prototype model; participants split between exemplar and adaptive.

participant_best_overall <- preds_pL %>%
  group_by(PROLIFIC_PID, model) %>% summarise(total_ll = sum(log_lik), .groups = "drop") %>%
  group_by(PROLIFIC_PID) %>% slice_max(total_ll, n = 1, with_ties = FALSE) %>%
  ungroup() %>% select(PROLIFIC_PID, best_model = model)

preds_pL <- preds_pL %>% left_join(participant_best_overall, by = "PROLIFIC_PID")
participant_best_overall %>% count(best_model, .drop = FALSE)
## # A tibble: 3 × 2
##   best_model     n
##   <fct>      <int>
## 1 prototype      0
## 2 exemplar       4
## 3 adaptive      13

Best-fitting model shifts from prototype toward adaptive/exemplar as evidence accumulates:

cumulative_best <- preds_pL %>%
  filter(Trial %% 2 == 0) %>%
  arrange(PROLIFIC_PID, model, iteration) %>%
  group_by(PROLIFIC_PID, model, iteration) %>%
  summarise(segment_ll = sum(log_lik), .groups = "drop") %>%
  group_by(PROLIFIC_PID, model) %>%
  mutate(cumulative_ll = cumsum(segment_ll)) %>%
  group_by(PROLIFIC_PID, iteration) %>%
  slice_max(cumulative_ll, n = 1, with_ties = FALSE) %>%
  ungroup()

cumulative_best %>%
  mutate(trial_block = iteration) %>%
  count(trial_block, model, .drop = FALSE) %>%
  ggplot(aes(trial_block, n, fill = model)) +
  geom_col(position = "stack") +
  scale_fill_manual(values = model_palette, limits = primary_models) +
  scale_x_continuous(breaks = 0:10) + coord_cartesian(xlim = c(0, 10)) +
  labs(x = "Trial block", y = "Participants", fill = "Best model\n(cumulative)") +
  theme_minimal(25)

ggsave("figures/best-cumulative-model.pdf", width = 9, height = 5)

Behaviour of participants grouped by best-fitting model, and their mean guessing rate (exemplar-best participants guess more and hover near chance on exceptions):

human_L %>%
  left_join(participant_best_overall, by = "PROLIFIC_PID") %>%
  filter(Trial %% 2 == 0) %>%
  group_by(best_model, trial_block, stimulus_type) %>%
  summarise(accuracy = mean(human_correct), n = n(), .groups = "drop") %>%
  ggplot(aes(trial_block, accuracy, color = best_model, shape = stimulus_type)) +
  geom_point() + geom_line() +
  scale_color_manual(values = model_palette) +
  trial_block_scale + facet_wrap(~best_model) +
  labs(title = "Participant best fit by...", x = "Trial block",
       y = "Human accuracy", color = "Best-fit model", shape = "Stimulus type") +
  theme_minimal(16)

ggsave("figures/performance-by-best-model.pdf", width = 11, height = 6)

# Mean guessing rate by best-fit model (learners; averaged over the 3 primary
# models per participant, then over participants in each best-fit group).
# Supplies [X] (exemplar) and [Y] (adaptive) for the text.
params %>%
  filter(model %in% primary_models, PROLIFIC_PID %in% learner_ids) %>%
  group_by(PROLIFIC_PID) %>%
  summarise(mean_guessing = mean(guessing), .groups = "drop") %>%
  inner_join(participant_best_overall, by = "PROLIFIC_PID") %>%
  group_by(best_model) %>%
  summarise(mean_guessing = round(mean(mean_guessing), 3),
            n = n(), .groups = "drop") %>%
  arrange(desc(mean_guessing))
## # A tibble: 2 × 3
##   best_model mean_guessing     n
##   <fct>              <dbl> <int>
## 1 exemplar           0.222     4
## 2 adaptive           0.179    13

5. Cluster structure (adaptive model)

Number of clusters the adaptive model infers over learning — starting near the two prototype clusters and growing as the exceptions are accommodated:

ll_L %>%
  filter(model == "adaptive") %>%
  group_by(iteration) %>%
  summarise(mean_clusters = mean(total_clusters),
            se = sd(total_clusters) / sqrt(n()), .groups = "drop") %>%
  mutate(trial_block = iteration) %>%
  ggplot(aes(trial_block, mean_clusters)) +
  geom_ribbon(aes(ymin = mean_clusters - se, ymax = mean_clusters + se),
              alpha = 0.2, fill = model_palette["adaptive"]) +
  geom_point(color = model_palette["adaptive"]) +
  geom_line(color = model_palette["adaptive"]) +
  trial_block_scale +
  labs(x = "Trial block", y = "Total clusters") +
  theme_minimal(16)

ggsave("figures/clusters-over-training.pdf", width = 9, height = 7)

5a. Consensus co-assignment — which words cluster together

cluster_consensus gives, per (participant, segment) and word pair, the fraction of posterior samples in which the two words share a dish (label-switching invariant); we average over learners and show early / mid / late segments.

word_order <- c(a_words, "gavuco", b_words, "zabeci")

consensus_mean <- consen_L %>%
  filter(model == "adaptive") %>%
  group_by(iteration, word_a, word_b) %>%
  summarise(p_coassign = mean(p_coassign), .groups = "drop")

# symmetric long form (incl. diagonal = 1) for one segment
heat_seg <- function(seg) {
  d <- consensus_mean %>% filter(iteration == seg)
  bind_rows(d,
            d %>% rename(word_a = word_b, word_b = word_a),
            tibble(iteration = seg, word_a = word_order,
                   word_b = word_order, p_coassign = 1))
}

seg_labels <- c(`1` = "Segment 1 (early)",
                `5` = "Segment 5 (mid)",
                `10` = "Segment 10 (late)")

heat <- bind_rows(lapply(c(1, 5, 10), heat_seg)) %>%
  mutate(word_a = factor(word_a, levels = word_order),
         word_b = factor(word_b, levels = rev(word_order)),
         panel  = factor(seg_labels[as.character(iteration)],
                         levels = unname(seg_labels)))

ggplot(heat, aes(word_a, word_b, fill = p_coassign)) +
  geom_tile(color = "white") +
  scale_fill_viridis_c(limits = c(0, 1)) +
  facet_wrap(~panel) +
  coord_equal() +
  labs(x = NULL, y = NULL, fill = "P(co-assign)") +
  theme_minimal(11) +
  theme(axis.text.x = element_text(angle = 90, hjust = 1, vjust = 0.5))

ggsave("figures/consensus-heatmap.pdf", width = 13, height = 5)

5b. Do the exceptions split off into their own cluster?

An exception and its phonological group take different suffixes, so within the adaptive model (no cross-class sharing) they cannot share a dish; but we can ask whether the exception first gets lumped with the prototypical words of its own suffix class and then splits off. We track (i) co-assignment with the words that share a suffix class with it (gavuco↔︎b_words, zabeci↔︎a_words) and (ii) an isolation index (1 − mean co-assignment with all partners). The first falls and the second rises, i.e. the exceptions do progressively form their own cluster.

pair_long <- consen_L %>%
  filter(model == "adaptive") %>%
  bind_rows(consen_L %>% filter(model == "adaptive") %>%
              rename(word_a = word_b, word_b = word_a))

is_suffix_mate <- function(ex, w)
  (ex == "gavuco" & w %in% b_words) | (ex == "zabeci" & w %in% a_words)

detach_own <- pair_long %>%
  filter(word_a %in% exceptions, is_suffix_mate(word_a, word_b)) %>%
  group_by(iteration, exception = word_a) %>%
  summarise(coassign_suffix_mates = mean(p_coassign), .groups = "drop")

isolation <- pair_long %>%
  filter(word_a %in% exceptions) %>%
  group_by(iteration, exception = word_a) %>%
  summarise(isolation = 1 - mean(p_coassign), .groups = "drop")

left_join(detach_own, isolation, by = c("iteration", "exception")) %>%
  pivot_longer(c(coassign_suffix_mates, isolation),
               names_to = "metric", values_to = "value") %>%
  mutate(trial_block = iteration,
         metric = recode(metric,
                         coassign_suffix_mates = "co-assign w/ suffix-class mates",
                         isolation = "isolation index")) %>%
  ggplot(aes(trial_block, value, color = exception)) +
  geom_point() + geom_line() +
  facet_wrap(~metric) + trial_block_scale +
  labs(x = "Trial block", y = NULL, color = "Exception") +
  theme_minimal(16)

ggsave("figures/exception-detachment.pdf", width = 11, height = 6)

6. Supplementary: the shared model (allowing cluster sharing)

The shared model is identical to adaptive except that γ is sampled, so a cluster may be shared across the two suffix classes. It is not a competitor on granularity — it varies γ, not α — and it does not fit human responses any better than adaptive: summed log-likelihood over learners differs by only a few units, and across participant × segment cells shared assigns higher probability to the human’s response in barely over half of them. Still, the shared model allows us to investigate cluster structure.

Paired shared − adaptive fit differences (near zero; shared wins ≈ half the cells):

paired <- preds_L %>%
  filter(model %in% c("adaptive", "shared"), Trial %% 2 == 0) %>%
  group_by(PROLIFIC_PID, iteration, stimulus_type, model) %>%
  summarise(ll = sum(log_lik), .groups = "drop") %>%
  pivot_wider(names_from = model, values_from = ll) %>%
  mutate(d_ll = shared - adaptive)

paired %>%
  summarise(mean_d_ll = mean(d_ll, na.rm = TRUE),
            frac_shared_better_ll = mean(d_ll > 0, na.rm = TRUE))
## # A tibble: 1 × 2
##   mean_d_ll frac_shared_better_ll
##       <dbl>                 <dbl>
## 1    0.0153                 0.526
paired %>%
  group_by(iteration, stimulus_type) %>%
  summarise(mean_d_ll = mean(d_ll, na.rm = TRUE),
            se = sd(d_ll, na.rm = TRUE) / sqrt(n()), .groups = "drop") %>%
  mutate(trial_block = iteration) %>%
  ggplot(aes(trial_block, mean_d_ll, color = stimulus_type)) +
  geom_hline(yintercept = 0, color = "grey50") +
  geom_ribbon(aes(ymin = mean_d_ll - se, ymax = mean_d_ll + se, fill = stimulus_type),
              alpha = 0.2, color = NA) +
  geom_point() + geom_line() +
  trial_block_scale +
  labs(x = "Trial block", y = "LL(shared) − LL(adaptive)",
       color = "Stimulus type", fill = "Stimulus type") +
  theme_minimal(16)

ggsave("figures/shared-vs-adaptive-delta.pdf", width = 9, height = 6)

Sampled γ and the number of dishes shared across both classes, over learning:

shared_diag <- ll_L %>%
  filter(model == "shared") %>%
  group_by(iteration) %>%
  summarise(mean_gamma = mean(mean_gamma, na.rm = TRUE),
            se_gamma = sd(mean_gamma, na.rm = TRUE) / sqrt(n()),
            mean_shared = mean(shared_dishes, na.rm = TRUE),
            se_shared = sd(shared_dishes, na.rm = TRUE) / sqrt(n()),
            .groups = "drop") %>%
  mutate(trial_block = iteration)

ggplot(shared_diag, aes(trial_block, mean_gamma)) +
  geom_ribbon(aes(ymin = mean_gamma - se_gamma, ymax = mean_gamma + se_gamma),
              alpha = 0.2, fill = model_palette["shared"]) +
  geom_point(color = model_palette["shared"]) + geom_line(color = model_palette["shared"]) +
  trial_block_scale + labs(x = "Trial block", y = "Sampled gamma (shared model)") +
  theme_minimal(16)

ggsave("figures/gamma-over-time.pdf", width = 9, height = 6)

ggplot(shared_diag, aes(trial_block, mean_shared)) +
  geom_ribbon(aes(ymin = mean_shared - se_shared, ymax = mean_shared + se_shared),
              alpha = 0.2, fill = model_palette["shared"]) +
  geom_point(color = model_palette["shared"]) + geom_line(color = model_palette["shared"]) +
  trial_block_scale + labs(x = "Trial block", y = "Mean # dishes shared across classes") +
  theme_minimal(16)

ggsave("figures/shared-dishes.pdf", width = 9, height = 6)

What is shared across classes: the top cross-class co-assignment pairs are each exception with its own phonological group across the suffix boundary (gavuco with the gyvuco family, zabeci with the zabeti family

consen_L %>%
  filter(model == "shared") %>%
  mutate(cls_a = word_class(word_a), cls_b = word_class(word_b)) %>%
  filter(cls_a != cls_b) %>%
  group_by(word_a, word_b) %>%
  summarise(mean_coassign = mean(p_coassign),
            max_coassign = max(p_coassign), .groups = "drop") %>%
  arrange(desc(mean_coassign)) %>% head(15)
## # A tibble: 15 × 4
##    word_a word_b mean_coassign max_coassign
##    <chr>  <chr>          <dbl>        <dbl>
##  1 gavuco zyvuco       0.182           1   
##  2 gavuco gyvuto       0.182           1   
##  3 gavuco gyvuci       0.182           1   
##  4 gavuco gyveco       0.181           1   
##  5 gavuco gybuco       0.181           1   
##  6 gavuco gyvuco       0.180           1   
##  7 zabeci zybeti       0.170           1   
##  8 zabeci zabuti       0.170           1   
##  9 gabeti zabeci       0.169           1   
## 10 zabeci zabeto       0.169           1   
## 11 zabeci zaveti       0.169           1   
## 12 zabeci zabeti       0.168           1   
## 13 gavuco zabeci       0.148           1   
## 14 zabeto zyvuco       0.00894         0.24
## 15 zybeti zyvuco       0.00847         0.12
consen_L %>%
  filter(model == "shared") %>%
  mutate(cross = word_class(word_a) != word_class(word_b)) %>%
  filter(cross) %>%
  group_by(iteration) %>%
  summarise(mean_cross_coassign = mean(p_coassign), .groups = "drop") %>%
  mutate(trial_block = iteration) %>%
  ggplot(aes(trial_block, mean_cross_coassign)) +
  geom_point(color = model_palette["shared"]) + geom_line(color = model_palette["shared"]) +
  trial_block_scale + labs(x = "Trial block", y = "Mean cross-class co-assignment") +
  theme_minimal(16)

ggsave("figures/cross-class-sharing.pdf", width = 9, height = 6)

7. Session info

sessionInfo()
## R version 4.6.1 (2026-06-24)
## Platform: aarch64-apple-darwin23
## Running under: macOS Sequoia 15.4.1
## 
## Matrix products: default
## BLAS:   /Library/Frameworks/R.framework/Versions/4.6/Resources/lib/libRblas.0.dylib 
## LAPACK: /Library/Frameworks/R.framework/Versions/4.6/Resources/lib/libRlapack.dylib;  LAPACK version 3.12.1
## 
## locale:
## [1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
## 
## time zone: Europe/Rome
## tzcode source: internal
## 
## attached base packages:
## [1] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
##  [1] brms_2.23.0     Rcpp_1.1.2      lubridate_1.9.5 forcats_1.0.1  
##  [5] stringr_1.6.0   dplyr_1.2.1     purrr_1.2.2     readr_2.2.0    
##  [9] tidyr_1.3.2     tibble_3.3.1    ggplot2_4.0.3   tidyverse_2.0.0
## 
## loaded via a namespace (and not attached):
##  [1] gtable_0.3.6          tensorA_0.36.2.1      xfun_0.60            
##  [4] bslib_0.11.0          lattice_0.22-9        tzdb_0.5.0           
##  [7] vctrs_0.7.3           tools_4.6.1           generics_0.1.4       
## [10] parallel_4.6.1        pkgconfig_2.0.3       Matrix_1.7-5         
## [13] checkmate_2.3.4       RColorBrewer_1.1-3    S7_0.2.2             
## [16] distributional_0.8.1  RcppParallel_5.1.11-2 lifecycle_1.0.5      
## [19] compiler_4.6.1        farver_2.1.2          textshaping_1.0.5    
## [22] Brobdingnag_1.2-9     htmltools_0.5.9       sass_0.4.10          
## [25] bayesplot_1.15.0      yaml_2.3.12           pillar_1.11.1        
## [28] crayon_1.5.3          jquerylib_0.1.4       cachem_1.1.0         
## [31] bridgesampling_1.2-1  abind_1.4-8           nlme_3.1-169         
## [34] posterior_1.7.0       tidyselect_1.2.1      digest_0.6.39        
## [37] mvtnorm_1.4-1         stringi_1.8.7         labeling_0.4.3       
## [40] fastmap_1.2.0         grid_4.6.1            cli_3.6.6            
## [43] magrittr_2.0.5        loo_2.10.0            utf8_1.2.6           
## [46] withr_3.0.3           scales_1.4.0          backports_1.5.1      
## [49] bit64_4.8.2           timechange_0.4.0      rmarkdown_2.31       
## [52] matrixStats_1.5.0     bit_4.6.0             otel_0.2.0           
## [55] ragg_1.5.2            hms_1.1.4             coda_0.19-4.1        
## [58] evaluate_1.0.5        knitr_1.51            viridisLite_0.4.3    
## [61] rstantools_2.6.0      rlang_1.3.0           glue_1.8.1           
## [64] rstudioapi_0.19.0     vroom_1.7.1           jsonlite_2.0.0       
## [67] R6_2.6.1              systemfonts_1.3.2