"""
Evaluation harness for the corvid-inspired specialists.

Building the modules isn't the same as confirming they produce the actual dissociation
each species result is named after. Each function here targets ONE specific claim and
measures it directly, on synthetic data engineered so the two competing explanations
(shallow vs. deep rule, item memory vs. relational memory, etc.) are distinguishable:

  eval_nutcracker_few_shot_transfer   -- does training on a handful of exemplar pairs
                                          transfer the same/different RELATION to brand
                                          new item vectors, or does it only memorize the
                                          training items? (Magnotti et al. 2015)

  eval_rook_rule_abstraction          -- when a surface cue that's redundant with the
                                          true rule in training is removed at test time,
                                          does a population of independently-initialized
                                          rook modules split into a shallow-rule majority
                                          and a deep-rule minority, the way real rooks did?
                                          (Bird & Emery 2009)

  eval_magpie_source_monitoring       -- does self/other classification hold up as the
                                          distance between a self-generated event and the
                                          probe grows, or does it degrade with distance?

  eval_raven_delayed_gratification    -- does the RavenForesightBuffer learn to forgo a
                                          low immediate value in favor of a stored high
                                          value item, beating an always-take-immediate
                                          baseline on cumulative realized reward?

None of these need the full CorvidaeAviary model -- they exercise the specialist modules
directly with synthetic data, so they can run (and be trusted) independently of whether
corvidae.py's base architecture is available.

Run `python corvid_eval.py` to execute all four and print a pass/fail summary against
simple, stated thresholds. The thresholds are deliberately loose sanity checks, not
publication-grade statistics -- tighten them once you have real task data.
"""

import torch
import torch.nn.functional as F

from species_memory_bank import NutcrackerConceptMemory
from corvid_extensions import (
    IndividualRookExperts, RavenForesightBuffer, NumerosityModule,
    CrowStatisticalMemory, IndividualRecognitionMemory,
)
from species_memory_bank import MagpieSelfModel
from corvidae import CausalRelationModule, MetatoolPlanningBuffer


# ====================== 1. NUTCRACKER: FEW-SHOT RELATION TRANSFER ======================
def eval_nutcracker_few_shot_transfer(embedding_dim: int = 16, num_train_pairs: int = 8,
                                       num_test_pairs: int = 200, train_steps: int = 300,
                                       seed: int = 0):
    """
    Train on `num_train_pairs` exemplar pairs (half "same" = two noisy copies of one
    vector, half "different" = two independent random vectors). Test on entirely new
    random item vectors never seen in training. If the module learned the RELATION
    (same/different), test accuracy should be well above chance despite the items being
    novel -- item memory alone can't do this, since the test items never appeared in
    training.
    """
    torch.manual_seed(seed)
    module = NutcrackerConceptMemory(embedding_dim)
    opt = torch.optim.Adam(module.parameters(), lr=1e-2)

    def make_pairs(n):
        n_same = n // 2
        base = torch.randn(n_same, embedding_dim)
        same_a = base
        same_b = base + torch.randn(n_same, embedding_dim) * 0.05
        diff_a = torch.randn(n - n_same, embedding_dim)
        diff_b = torch.randn(n - n_same, embedding_dim)
        item_a = torch.cat([same_a, diff_a], dim=0)
        item_b = torch.cat([same_b, diff_b], dim=0)
        labels = torch.cat([torch.ones(n_same), torch.zeros(n - n_same)])
        perm = torch.randperm(n)
        return item_a[perm].unsqueeze(1), item_b[perm].unsqueeze(1), labels[perm].unsqueeze(1)

    train_a, train_b, train_labels = make_pairs(num_train_pairs)
    test_a, test_b, test_labels = make_pairs(num_test_pairs)

    for _ in range(train_steps):
        opt.zero_grad()
        pair = torch.cat([train_a, train_b], dim=-1)
        relation = module.relation_proj(pair)
        relation_n = F.normalize(relation, dim=-1)
        same_sim = torch.einsum('bod,d->bo', relation_n, F.normalize(module.same_prototype, dim=0))
        diff_sim = torch.einsum('bod,d->bo', relation_n, F.normalize(module.diff_prototype, dim=0))
        same_logit = (same_sim - diff_sim)
        loss = module.concept_loss(same_logit, train_labels)
        loss.backward()
        opt.step()

    with torch.no_grad():
        pair = torch.cat([test_a, test_b], dim=-1)
        relation = module.relation_proj(pair)
        relation_n = F.normalize(relation, dim=-1)
        same_sim = torch.einsum('bod,d->bo', relation_n, F.normalize(module.same_prototype, dim=0))
        diff_sim = torch.einsum('bod,d->bo', relation_n, F.normalize(module.diff_prototype, dim=0))
        test_logit = same_sim - diff_sim
        preds = (torch.sigmoid(test_logit) > 0.5).float()
        acc = (preds == test_labels).float().mean().item()

    return {"test_accuracy_on_novel_items": acc, "num_train_pairs": num_train_pairs,
            "passed": acc > 0.85}


# ====================== 2. ROOK: RULE ABSTRACTION UNDER CUE REMOVAL ======================
def eval_rook_rule_abstraction(embedding_dim: int = 16, population_size: int = 7,
                                num_train: int = 300, train_steps: int = 200, seed: int = 0):
    """
    Builds `population_size` independently-initialized single-expert modules (a
    population of "individuals"), each trained on data where a surface cue (a specific
    input dimension) and the true abstract rule (parity of a *different* set of
    dimensions) are both perfectly predictive of the label. At test time the surface cue
    dimension is scrambled (replaced with noise uncorrelated with the label), so only a
    model that actually keyed on the abstract rule keeps working.

    Bird & Emery found 6/7 rooks locked onto the shallow cue and only 1/7 generalized.
    We report the fraction of the population whose accuracy holds up post cue-removal
    (mirrors "how many individuals abstracted the deep rule") rather than asserting a
    specific split, since that split is an empirical population statistic, not something
    to hard-code as a pass condition.
    """
    torch.manual_seed(seed)

    def make_batch(n, scramble_cue: bool):
        x = torch.randn(n, embedding_dim)
        # true rule: parity (sign product) of dims 0 and 1
        true_rule = (x[:, 0] * x[:, 1] > 0).float()
        # surface cue: dim 2, set equal to the label (+ small noise) during training,
        # scrambled (independent random) at test time
        if scramble_cue:
            x[:, 2] = torch.randn(n)
        else:
            x[:, 2] = true_rule * 2 - 1 + torch.randn(n) * 0.05
        return x, true_rule

    results = []
    for individual in range(population_size):
        torch.manual_seed(seed * 1000 + individual)
        # a single small "rule hypothesis" network, standing in for one rook -- built
        # directly rather than via IndividualRookExperts' internal K-expert mixture,
        # since here we want K *separate individuals*, not K experts inside one bird.
        net = torch.nn.Sequential(
            torch.nn.Linear(embedding_dim, embedding_dim), torch.nn.ReLU(), torch.nn.Linear(embedding_dim, 1)
        )
        opt = torch.optim.Adam(net.parameters(), lr=5e-3)
        train_x, train_y = make_batch(num_train, scramble_cue=False)
        for _ in range(train_steps):
            opt.zero_grad()
            logit = net(train_x).squeeze(-1)
            loss = F.binary_cross_entropy_with_logits(logit, train_y)
            loss.backward()
            opt.step()

        with torch.no_grad():
            test_x, test_y = make_batch(400, scramble_cue=True)
            test_logit = net(test_x).squeeze(-1)
            test_acc = ((torch.sigmoid(test_logit) > 0.5).float() == test_y).float().mean().item()
        results.append(test_acc)

    generalized = [r > 0.75 for r in results]
    return {
        "per_individual_test_accuracy": results,
        "fraction_generalized": sum(generalized) / len(generalized),
        "passed": 0 < sum(generalized) < population_size,  # expect a SPLIT, not unanimity either way
    }


# ====================== 3. MAGPIE: SOURCE MONITORING OVER DISTANCE ======================
def eval_magpie_source_monitoring(embedding_dim: int = 16, seq_len: int = 64,
                                   train_steps: int = 300, seed: int = 0):
    """
    Builds sequences where each position is either "self" (a noisy copy of the module's
    own self_embedding, simulating re-read of the model's own prior output) or "other"
    (independent random content). Trains the classifier, then checks accuracy in early
    vs. late positions in the sequence -- a real source-monitoring system should hold up
    across the sequence, not just near the start; a system doing something more like
    short-range pattern matching would degrade with distance.
    """
    torch.manual_seed(seed)
    module = MagpieSelfModel(embedding_dim)
    opt = torch.optim.Adam(module.parameters(), lr=1e-2)

    def make_batch(batch_size):
        is_self = (torch.rand(batch_size, seq_len) > 0.5).float()
        self_part = module.self_embedding.detach().unsqueeze(0).unsqueeze(0) + torch.randn(batch_size, seq_len, embedding_dim) * 0.1
        other_part = torch.randn(batch_size, seq_len, embedding_dim)
        x = is_self.unsqueeze(-1) * self_part + (1 - is_self.unsqueeze(-1)) * other_part
        return x, is_self

    for _ in range(train_steps):
        opt.zero_grad()
        x, labels = make_batch(32)
        self_logit, _, _ = module(x)
        loss = module.self_other_loss(self_logit, labels)
        loss.backward()
        opt.step()

    with torch.no_grad():
        x, labels = make_batch(256)
        self_logit, _, _ = module(x)
        preds = (torch.sigmoid(self_logit) > 0.5).float()
        correct = (preds == labels).float()
        early_acc = correct[:, : seq_len // 4].mean().item()
        late_acc = correct[:, -seq_len // 4 :].mean().item()

    return {
        "early_position_accuracy": early_acc,
        "late_position_accuracy": late_acc,
        "accuracy_gap": early_acc - late_acc,
        "passed": late_acc > 0.75 and abs(early_acc - late_acc) < 0.15,
    }


# ====================== 4. RAVEN: DELAYED GRATIFICATION ======================
def eval_raven_delayed_gratification(embedding_dim: int = 16, num_episodes: int = 60,
                                      episode_len: int = 12, seed: int = 0):
    """
    A scripted environment: at each step a candidate item arrives with a small immediate
    value; occasionally (every ~4 steps) a high-value item arrives that's worth much more
    IF the model holds it and cashes it in a few steps later rather than using whatever's
    immediately in front of it. We train the value head with realized-reward targets and
    check whether cumulative realized reward beats an always-take-immediate baseline --
    the direct behavioral signature Kabadayi & Osvath were testing for.
    """
    torch.manual_seed(seed)
    buffer = RavenForesightBuffer(embedding_dim, buffer_size=3, patience_cost=0.01)
    opt = torch.optim.Adam(buffer.parameters(), lr=1e-2)

    high_value_direction = torch.randn(embedding_dim)

    def make_episode():
        items, true_values = [], []
        for t in range(episode_len):
            if t % 4 == 0:
                item = high_value_direction + torch.randn(embedding_dim) * 0.05
                value = torch.tensor(5.0)
            else:
                item = torch.randn(embedding_dim)
                value = torch.tensor(0.5)
            items.append(item)
            true_values.append(value)
        return torch.stack(items), torch.stack(true_values)

    # NOTE: computing "realized reward for the actual policy" requires tracking which
    # true_value corresponds to whatever got cashed in each step (immediate vs. stored);
    # for this sanity-check version we approximate policy quality via the learned
    # candidate_value's correlation with true value, and via store utilization -- a full
    # credit-assignment loop (matching stored items back to their original true_value on
    # the step they're eventually used) is straightforward but more code than belongs in
    # a smoke-level eval; flagged here rather than silently faked.
    correlations = []
    for ep in range(num_episodes):
        buffer.reset(batch_size=1, device="cpu")
        items, true_values = make_episode()
        opt.zero_grad()
        pred_values = []
        for t in range(episode_len):
            candidate = items[t : t + 1]
            _, info = buffer.step(candidate)
            pred_values.append(info["candidate_value"])
        pred_values = torch.cat(pred_values)
        loss = buffer.value_calibration_loss(pred_values, true_values)
        loss.backward()
        opt.step()
        with torch.no_grad():
            if pred_values.std() > 1e-6:
                corr = torch.corrcoef(torch.stack([pred_values, true_values]))[0, 1].item()
                correlations.append(corr)

    final_corr = sum(correlations[-10:]) / max(1, len(correlations[-10:]))
    return {
        "value_head_true_value_correlation": final_corr,
        "passed": final_corr > 0.6,
        "caveat": "Tests whether the value head learns to recognize high-value items; "
                  "does NOT yet close the loop on full episode-level realized-reward "
                  "credit assignment for the wait/use decision -- see NOTE in source.",
    }


# ====================== 5. RAVEN NUMEROSITY: RELATIVE NUMBER + ADDITION ======================
def eval_numerosity(embedding_dim: int = 16, num_train: int = 300, train_steps: int = 300, seed: int = 0):
    """
    Pika et al. (2020): ravens matched great apes on both relative-number discrimination
    and addition-of-hidden-quantities. We build synthetic "quantity" embeddings whose
    true magnitude is an underlying scalar baked into the embedding via a fixed random
    projection (so the model must learn to extract it, not just read it off directly),
    train the magnitude head on relative-number comparisons, then test BOTH relative
    number on novel magnitude pairs and addition (m1 + m2 vs. a third quantity) using
    magnitude combinations never seen during training.
    """
    torch.manual_seed(seed)
    module = NumerosityModule(embedding_dim)
    opt = torch.optim.Adam(module.parameters(), lr=1e-2)

    proj = torch.randn(1, embedding_dim)  # fixed encoding direction "hiding" the magnitude

    def encode(magnitude):
        return magnitude.unsqueeze(-1) * proj + torch.randn(*magnitude.shape, embedding_dim) * 0.05

    def make_pairs(n, max_mag=10.0):
        mag_a = torch.rand(n) * max_mag
        mag_b = torch.rand(n) * max_mag
        item_a, item_b = encode(mag_a), encode(mag_b)
        labels = (mag_a > mag_b).float()
        return item_a, item_b, labels

    train_a, train_b, train_labels = make_pairs(num_train)
    for _ in range(train_steps):
        opt.zero_grad()
        _, _, _, choose_a_logit = module(train_a, train_b)
        loss = module.relative_number_loss(choose_a_logit, train_labels)
        loss.backward()
        opt.step()

    with torch.no_grad():
        test_a, test_b, test_labels = make_pairs(300, max_mag=20.0)  # novel magnitude range
        _, _, _, test_logit = module(test_a, test_b)
        rel_acc = ((torch.sigmoid(test_logit) > 0.5).float() == test_labels).float().mean().item()

        # addition: sum two never-jointly-seen magnitudes and compare to a third
        m1 = torch.rand(300) * 10.0
        m2 = torch.rand(300) * 10.0
        mc = torch.rand(300) * 20.0
        item1, item2, itemc = encode(m1), encode(m2), encode(mc)
        add_labels = ((m1 + m2) > mc).float()
        _, _, add_logit = module.addition_forward(item1, item2, itemc)
        add_acc = ((torch.sigmoid(add_logit) > 0.5).float() == add_labels).float().mean().item()

    return {
        "relative_number_accuracy": rel_acc, "addition_accuracy": add_acc,
        "passed": rel_acc > 0.8 and add_acc > 0.7,
    }


# ====================== 6. STATISTICAL INFERENCE FROM MEMORIZED PROBABILITIES ======================
def eval_statistical_inference(embedding_dim: int = 16, num_stimuli: int = 9, seed: int = 0):
    """
    Johnston, Brecht & Nieder (2023): crows chose the higher-REWARD-PROBABILITY stimulus
    even when it was shown less often (lower absolute frequency) during the choice test
    than the alternative -- a sample-to-population inference from memorized associations,
    not a simple frequency-matching heuristic. We give each of `num_stimuli` fixed
    embeddings a true reward probability, "train" the memory with a number of learn()
    exposures per stimulus that's INVERSELY related to its true probability (mirroring
    the actual test design where the higher-probability stimulus is shown less), then
    check whether compare_and_choose still favors the higher-probability stimulus.
    """
    torch.manual_seed(seed)
    memory = CrowStatisticalMemory(embedding_dim, memory_size=num_stimuli * 2)

    stimuli = torch.randn(num_stimuli, embedding_dim)
    true_probs = torch.linspace(0.1, 0.9, num_stimuli)
    # inverse exposure count: the higher-probability stimuli are seen FEWER times,
    # mirroring the actual experimental design's deliberate frequency/probability split
    exposure_counts = (50 - 40 * true_probs).long()

    for i in range(num_stimuli):
        for _ in range(int(exposure_counts[i])):
            observed = 1.0 if torch.rand(1).item() < true_probs[i].item() else 0.0
            memory.learn(stimuli[i], observed)

    correct, total = 0, 0
    for i in range(num_stimuli):
        for j in range(num_stimuli):
            if i == j:
                continue
            logit = memory.compare_and_choose(stimuli[i], stimuli[j])
            predicted_prefers_i = logit.item() > 0
            true_prefers_i = true_probs[i] > true_probs[j]
            correct += int(predicted_prefers_i == true_prefers_i.item())
            total += 1

    accuracy = correct / total
    return {
        "pairwise_choice_accuracy": accuracy,
        "note": "Higher-probability stimuli were shown FEWER times during learning, so "
                "high accuracy here means the model tracked probability, not raw frequency.",
        "passed": accuracy > 0.75,
    }


# ====================== 7. CAUSAL ANALOGY TRANSFER (TRAP-TUBE -> TRAP-TABLE) ======================
def eval_causal_analogy_transfer(embedding_dim: int = 16, num_surface_contexts: int = 4,
                                  train_steps: int = 400, seed: int = 0):
    """
    Taylor et al. (2009): NC crows solved a trap-tube then immediately transferred to a
    trap-TABLE sharing no visual features. We build synthetic contexts where the true
    outcome depends only on an abstract "causal" feature (e.g. hole-relative-position,
    encoded as one direction in embedding space) while a "surface" feature (apparatus
    identity/appearance, encoded as a different, context-specific direction) is
    correlated with outcome only within each training surface context. At test time we
    introduce a BRAND NEW surface context (never seen in training) where only the causal
    feature still predicts the outcome. A model that leaned on surface features should
    fail on the new context; one with genuine causal invariance should not.
    """
    torch.manual_seed(seed)
    module = CausalRelationModule(embedding_dim, num_surface_contexts=num_surface_contexts)
    opt = torch.optim.Adam(module.parameters(), lr=1e-2)

    causal_direction = torch.randn(embedding_dim)
    surface_directions = torch.randn(num_surface_contexts + 1, embedding_dim)  # +1 = held-out test context

    def make_batch(n, surface_context: int):
        causal_val = torch.randn(n)
        outcome = (causal_val > 0).float()
        x = (causal_val.unsqueeze(-1) * causal_direction
             + torch.randn(n) .unsqueeze(-1) * surface_directions[surface_context] * 0.8
             + torch.randn(n, embedding_dim) * 0.1)
        surface_label = torch.full((n,), surface_context, dtype=torch.long)
        return x.unsqueeze(1), outcome.unsqueeze(1), surface_label.unsqueeze(1)  # add a seq_len=1 dim

    for _ in range(train_steps):
        opt.zero_grad()
        ctx = torch.randint(0, num_surface_contexts, (1,)).item()
        x, outcome, surface_label = make_batch(32, ctx)
        _ = module(x)
        loss = (module.outcome_loss(module.last_outcome_logit, outcome)
                + module.surface_adversary_loss(module.last_surface_logits, surface_label))
        loss.backward()
        opt.step()

    with torch.no_grad():
        # held-out, never-seen surface context (index num_surface_contexts)
        x, outcome, _ = make_batch(300, num_surface_contexts)
        _ = module(x)
        preds = (torch.sigmoid(module.last_outcome_logit) > 0.5).float()
        transfer_acc = (preds == outcome).float().mean().item()

    return {
        "transfer_accuracy_novel_surface_context": transfer_acc,
        "passed": transfer_acc > 0.75,
    }


# ====================== 8. METATOOL SUB-GOAL / DISTRACTOR SUPPRESSION ======================
def eval_metatool_subgoal_distractor(embedding_dim: int = 16, train_steps: int = 300, seed: int = 0):
    """
    Gruber et al. (2019): crows kept a functional sub-goal AND a distractor sub-goal in
    mind simultaneously across out-of-sight stages, but specifically suppressed the
    distractor's influence on behavior. We give the buffer a sequence where a "functional
    subgoal" signal and a "distractor subgoal" signal are both written, then check that
    the buffer's read-out correlates with the functional signal much more than with the
    distractor signal after training the distractor gate to fire when distractor content
    is present.
    """
    torch.manual_seed(seed)
    buffer = MetatoolPlanningBuffer(embedding_dim, num_slots=3)
    opt = torch.optim.Adam(buffer.parameters(), lr=1e-2)

    functional_direction = torch.randn(embedding_dim)
    distractor_direction = torch.randn(embedding_dim)

    def make_sequence(batch_size, seq_len=6):
        # step 0: "observe" functional subgoal; step 1: "observe" distractor subgoal;
        # remaining steps: neutral context (subgoals now out of sight)
        seq = torch.randn(batch_size, seq_len, embedding_dim) * 0.1
        seq[:, 0, :] += functional_direction
        seq[:, 1, :] += distractor_direction
        return seq

    for _ in range(train_steps):
        opt.zero_grad()
        buffer.clear()
        x = make_sequence(16)
        read = buffer(x)
        final_read = read[:, -1, :]  # read-out at the last (out-of-sight) step
        # train distractor gate to suppress: push final read AWAY from distractor
        # direction and TOWARD functional direction
        target = functional_direction.unsqueeze(0).expand(final_read.size(0), -1)
        loss = F.mse_loss(final_read, target)
        loss.backward()
        opt.step()

    with torch.no_grad():
        buffer.clear()
        x = make_sequence(64)
        read = buffer(x)
        final_read = F.normalize(read[:, -1, :], dim=-1)
        func_sim = (final_read @ F.normalize(functional_direction, dim=0)).mean().item()
        distractor_sim = (final_read @ F.normalize(distractor_direction, dim=0)).mean().item()

    return {
        "functional_subgoal_similarity": func_sim,
        "distractor_subgoal_similarity": distractor_sim,
        "passed": func_sim > distractor_sim + 0.2,
    }


# ====================== 9. ASYMMETRIC ONE-SHOT THREAT LEARNING ======================
def eval_individual_recognition_asymmetric_learning(embedding_dim: int = 16, seed: int = 0):
    """
    Marzluff et al. (2012): a SINGLE capture event is enough to teach a crow a face is
    dangerous; positive/neutral associations seem to build more gradually. We check that
    IndividualRecognitionMemory's valence estimate after ONE threatening exposure is much
    larger in magnitude than after ONE caring exposure, directly from the asymmetric
    learning rate.
    """
    torch.manual_seed(seed)
    memory = IndividualRecognitionMemory(embedding_dim, capacity=8)

    face_threat = torch.randn(embedding_dim)
    face_caring = torch.randn(embedding_dim)

    memory.update(face_threat, event_valence=-1.0)   # one capture event
    memory.update(face_caring, event_valence=1.0)     # one feeding event

    with torch.no_grad():
        valence_threat, _, known_threat = memory.recognize(face_threat)
        valence_caring, _, known_caring = memory.recognize(face_caring)

    return {
        "valence_after_one_threat_event": valence_threat.item(),
        "valence_after_one_caring_event": valence_caring.item(),
        "passed": abs(valence_threat.item()) > abs(valence_caring.item()) and known_threat.item() and known_caring.item(),
    }


if __name__ == "__main__":
    print("=== Nutcracker: few-shot same/different transfer ===")
    r1 = eval_nutcracker_few_shot_transfer()
    print(r1)

    print("\n=== Rook: rule abstraction under cue removal ===")
    r2 = eval_rook_rule_abstraction()
    print(r2)

    print("\n=== Magpie: source monitoring over distance ===")
    r3 = eval_magpie_source_monitoring()
    print(r3)

    print("\n=== Raven: delayed gratification (value calibration) ===")
    r4 = eval_raven_delayed_gratification()
    print(r4)

    print("\n=== Raven numerosity: relative number + addition ===")
    r5 = eval_numerosity()
    print(r5)

    print("\n=== Statistical inference from memorized reward probabilities ===")
    r6 = eval_statistical_inference()
    print(r6)

    print("\n=== Causal analogy transfer (trap-tube -> trap-table style) ===")
    r7 = eval_causal_analogy_transfer()
    print(r7)

    print("\n=== Metatool sub-goal / distractor suppression ===")
    r8 = eval_metatool_subgoal_distractor()
    print(r8)

    print("\n=== Asymmetric one-shot threat learning ===")
    r9 = eval_individual_recognition_asymmetric_learning()
    print(r9)

    print("\n=== summary ===")
    results = [
        ("nutcracker", r1), ("rook", r2), ("magpie", r3), ("raven_foresight", r4),
        ("numerosity", r5), ("statistical_inference", r6), ("causal_analogy", r7),
        ("metatool_distractor", r8), ("identity_asymmetric_learning", r9),
    ]
    for name, r in results:
        print(f"{name}: {'PASS' if r['passed'] else 'FAIL'}")