benchmark_v6.py

#!/usr/bin/env python3
"""
=============================================================================
BENCHMARK v6: THE ADAPTIVE NEURON
=============================================================================

The idea: Don't pick one neuron type. Let each neuron LEARN whether to be
linear or periodic based on the input.

  α(x) = sigmoid(W_route · x)                     # per-input routing
  linear_path  = W_lin · x                         # standard linear
  periodic_path = (W_gate · x) ⊙ sin(ω·W_per · x) # multiplicative periodic
  y = α ⊙ periodic_path + (1-α) ⊙ linear_path     # blend

On periodic tasks → α learns to go high → gets the 162× gains
On OOD inputs   → α learns to go low → falls back to safe linear behavior
On mixed tasks  → different neurons specialize differently

Extra cost: one routing vector W_route per layer — negligible.

We test against: Vanilla, SinGLU, Hybrid, and report:
  - Performance (3 seeds, mean±std)
  - α distribution (what did the routing learn?)
  - OOD generalization (does it auto-fallback?)
  - Gradient norms (is it stable?)

=============================================================================
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import math
import time
import json

DEVICE = 'cpu'
SEEDS = [0, 1, 2]

def set_seed(s):
    torch.manual_seed(s)
    np.random.seed(s)

# ============================================================================
# ALL ARCHITECTURES
# ============================================================================

class VanillaMLP(nn.Module):
    def __init__(self, in_dim, out_dim, hidden_dim, n_hidden):
        super().__init__()
        layers = []
        prev = in_dim
        for _ in range(n_hidden):
            layers.extend([nn.Linear(prev, hidden_dim), nn.ReLU()])
            prev = hidden_dim
        layers.append(nn.Linear(prev, out_dim))
        self.net = nn.Sequential(*layers)
    def forward(self, x):
        return self.net(x)


class SinGLULayer(nn.Module):
    def __init__(self, in_dim, out_dim, mid_dim, omega_0=30.0):
        super().__init__()
        self.W_gate = nn.Linear(in_dim, mid_dim, bias=False)
        self.W_val = nn.Linear(in_dim, mid_dim, bias=False)
        self.W_out = nn.Linear(mid_dim, out_dim, bias=True)
        self.omega_0 = omega_0
        self.ln = nn.LayerNorm(out_dim)
        with torch.no_grad():
            bound = math.sqrt(6.0 / in_dim) / omega_0
            self.W_gate.weight.uniform_(-bound, bound)
            nn.init.xavier_uniform_(self.W_val.weight)
            nn.init.xavier_uniform_(self.W_out.weight)
    def forward(self, x):
        return self.ln(self.W_out(torch.sin(self.omega_0 * self.W_gate(x)) * self.W_val(x)))

class SinGLUNet(nn.Module):
    def __init__(self, in_dim, out_dim, hidden_dim, n_hidden, omega_0=30.0):
        super().__init__()
        mid = max(2, int(hidden_dim * 2 / 3))
        layers = []
        prev = in_dim
        for _ in range(n_hidden):
            layers.append(SinGLULayer(prev, hidden_dim, mid, omega_0))
            prev = hidden_dim
        layers.append(nn.Linear(prev, out_dim))
        self.layers = nn.ModuleList(layers)
    def forward(self, x):
        for l in self.layers: x = l(x)
        return x


class HybridLayer(nn.Module):
    def __init__(self, in_dim, out_dim, mid_dim, omega_0=30.0):
        super().__init__()
        self.W1 = nn.Linear(in_dim, mid_dim, bias=False)
        self.W2 = nn.Linear(in_dim, mid_dim, bias=False)
        self.phase = nn.Parameter(torch.empty(mid_dim))
        self.W3 = nn.Linear(mid_dim, out_dim, bias=True)
        self.omega_0 = omega_0
        self.ln = nn.LayerNorm(out_dim)
        self.residual = nn.Linear(in_dim, out_dim, bias=False) if in_dim != out_dim else nn.Identity()
        with torch.no_grad():
            nn.init.xavier_uniform_(self.W1.weight)
            bound = math.sqrt(6.0 / in_dim) / omega_0
            self.W2.weight.uniform_(-bound, bound)
            self.phase.uniform_(-math.pi, math.pi)
            nn.init.xavier_uniform_(self.W3.weight)
    def forward(self, x):
        lin = self.W1(x)
        per = torch.sin(self.omega_0 * self.W2(x) + self.phase)
        return self.ln(self.W3(lin * per) + self.residual(x))

class HybridNet(nn.Module):
    def __init__(self, in_dim, out_dim, hidden_dim, n_hidden, omega_0=30.0):
        super().__init__()
        mid = max(2, int(hidden_dim * 0.55))
        layers = []
        prev = in_dim
        for _ in range(n_hidden):
            layers.append(HybridLayer(prev, hidden_dim, mid, omega_0))
            prev = hidden_dim
        layers.append(nn.Linear(prev, out_dim))
        self.layers = nn.ModuleList(layers)
    def forward(self, x):
        for l in self.layers: x = l(x)
        return x


# ============================================================================
# THE ADAPTIVE NEURON
# ============================================================================

class AdaptiveNeuronLayer(nn.Module):
    """
    The adaptive neuron: learns PER-INPUT whether to be linear or periodic.
    
      α(x) = sigmoid(W_route · x + b_route)          # routing signal [0,1]
      linear_path  = ReLU(W_lin · x + b_lin)          # standard linear
      periodic_path = (W_val · x) ⊙ sin(ω·W_per · x) # multiplicative periodic
      y = α ⊙ periodic_path + (1-α) ⊙ linear_path    # input-dependent blend
    
    Param budget:
      W_route(h×d) + W_lin(h×d) + W_per(h×d) + W_val(h×d) + biases
      = 4 matrices. To match budget, use reduced hidden dim.
      
    Alternatively (param-efficient version):
      W_route is a SMALL projection: W_r(h×r) @ W_r2(r×d), r=d//4
      And W_per shares structure with W_lin via low-rank delta.
      
    We implement the efficient version:
      - W_lin: main weight (h×d) — shared backbone
      - W_delta: low-rank perturbation for periodic branch: U(h×r)@V(r×d)
      - W_route: small routing head: W_r(h×r2)@W_r2(r2×d) 
      - W_val: reuse W_lin (same as the linear path)
      
    So periodic branch = (W_lin·x) ⊙ sin(ω·(W_lin + U@V)·x)
    This way W_lin captures shared features, U@V captures what's different
    about the periodic branch. Much more param-efficient.
    """
    def __init__(self, in_dim, out_dim, omega_0=30.0, route_rank=None):
        super().__init__()
        
        r = route_rank or max(2, in_dim // 4)  # rank for routing + delta
        
        # Main linear path
        self.W_lin = nn.Linear(in_dim, out_dim, bias=True)
        
        # Low-rank delta for periodic branch: periodic_W = W_lin.weight + U @ V
        self.U = nn.Linear(r, out_dim, bias=False)
        self.V = nn.Linear(in_dim, r, bias=False)
        self.phase = nn.Parameter(torch.empty(out_dim))
        
        # Routing: small learned gate
        self.route_V = nn.Linear(in_dim, r, bias=False)
        self.route_U = nn.Linear(r, out_dim, bias=True)
        
        self.omega_0 = omega_0
        self.ln = nn.LayerNorm(out_dim)
        
        with torch.no_grad():
            nn.init.xavier_uniform_(self.W_lin.weight)
            # Periodic delta: SIREN init
            bound = math.sqrt(6.0 / in_dim) / omega_0
            self.U.weight.uniform_(-bound, bound)
            self.V.weight.uniform_(-1.0/in_dim, 1.0/in_dim)
            self.phase.uniform_(-math.pi, math.pi)
            # Routing: init near 0.5 (balanced)
            nn.init.zeros_(self.route_U.weight)
            nn.init.zeros_(self.route_U.bias)  # sigmoid(0) = 0.5
            self.route_V.weight.uniform_(-0.01, 0.01)
    
    def forward(self, x):
        # Linear path
        lin = F.relu(self.W_lin(x))                           # (batch, out)
        
        # Periodic path: use W_lin + low-rank delta
        base = self.W_lin(x)                                   # shared computation
        delta = self.U(self.V(x))                              # low-rank perturbation
        periodic_input = base + delta
        per = base * torch.sin(self.omega_0 * periodic_input + self.phase)  # multiplicative
        
        # Routing: input-dependent α ∈ [0,1] per neuron
        alpha = torch.sigmoid(self.route_U(self.route_V(x)))  # (batch, out)
        
        # Blend
        out = alpha * per + (1 - alpha) * lin
        return self.ln(out)
    
    def get_alpha(self, x):
        """For analysis: return the routing values"""
        with torch.no_grad():
            return torch.sigmoid(self.route_U(self.route_V(x)))


class AdaptiveNet(nn.Module):
    def __init__(self, in_dim, out_dim, hidden_dim, n_hidden, omega_0=30.0):
        super().__init__()
        layers = []
        prev = in_dim
        for _ in range(n_hidden):
            layers.append(AdaptiveNeuronLayer(prev, hidden_dim, omega_0))
            prev = hidden_dim
        layers.append(nn.Linear(prev, out_dim))
        self.layers = nn.ModuleList(layers)
    
    def forward(self, x):
        for l in self.layers: x = l(x)
        return x
    
    def get_all_alphas(self, x):
        """Get routing values from all adaptive layers"""
        alphas = []
        h = x
        for l in self.layers:
            if isinstance(l, AdaptiveNeuronLayer):
                alphas.append(l.get_alpha(h))
                h = l(h)
            else:
                h = l(h)
        return alphas


# ============================================================================
# UTILS
# ============================================================================

def count_params(m):
    return sum(p.numel() for p in m.parameters() if p.requires_grad)

def find_hidden(in_d, out_d, n_h, target_p, model_cls, **kw):
    lo, hi, best_h = 2, 512, 2
    while lo <= hi:
        mid = (lo + hi) // 2
        p = count_params(model_cls(in_d, out_d, mid, n_h, **kw))
        if abs(p - target_p) < abs(count_params(model_cls(in_d, out_d, best_h, n_h, **kw)) - target_p):
            best_h = mid
        if p < target_p: lo = mid + 1
        else: hi = mid - 1
    return best_h

def train_reg(model, xtr, ytr, xte, yte, epochs, lr, bs=256):
    opt = torch.optim.Adam(model.parameters(), lr=lr)
    sch = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=epochs)
    best = float('inf')
    n = len(xtr)
    for ep in range(epochs):
        model.train()
        perm = torch.randperm(n)
        for i in range(0, n, bs):
            idx = perm[i:i+bs]
            loss = F.mse_loss(model(xtr[idx]), ytr[idx])
            opt.zero_grad(); loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
            opt.step()
        sch.step()
        if (ep+1) % max(1, epochs//10) == 0:
            model.eval()
            with torch.no_grad():
                best = min(best, F.mse_loss(model(xte), yte).item())
    model.eval()
    with torch.no_grad():
        best = min(best, F.mse_loss(model(xte), yte).item())
    return best

def train_clf(model, xtr, ytr, xte, yte, epochs, lr, bs=256):
    opt = torch.optim.Adam(model.parameters(), lr=lr)
    sch = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=epochs)
    best = 0
    n = len(xtr)
    for ep in range(epochs):
        model.train()
        perm = torch.randperm(n)
        for i in range(0, n, bs):
            idx = perm[i:i+bs]
            loss = F.cross_entropy(model(xtr[idx]), ytr[idx])
            opt.zero_grad(); loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
            opt.step()
        sch.step()
        if (ep+1) % max(1, epochs//10) == 0:
            model.eval()
            with torch.no_grad():
                best = max(best, (model(xte).argmax(1) == yte).float().mean().item())
    model.eval()
    with torch.no_grad():
        best = max(best, (model(xte).argmax(1) == yte).float().mean().item())
    return best

# ============================================================================
# DATA
# ============================================================================

def data_complex(n=1000):
    x = torch.rand(n,4)*2-1
    y = torch.exp(torch.sin(x[:,0]**2+x[:,1]**2)+torch.sin(x[:,2]**2+x[:,3]**2))
    return x, y.unsqueeze(1)

def data_nested(n=1000):
    x = torch.rand(n,2)*2-1
    y = torch.sin(math.pi*(x[:,0]**2+x[:,1]**2))*torch.cos(3*math.pi*x[:,0]*x[:,1])
    return x, y.unsqueeze(1)

def data_spiral(n=1000):
    t = torch.linspace(0, 4*np.pi, n//2)
    r = torch.linspace(0.3, 2, n//2)
    x1 = torch.stack([r*torch.cos(t), r*torch.sin(t)], 1)
    x2 = torch.stack([r*torch.cos(t+np.pi), r*torch.sin(t+np.pi)], 1)
    x = torch.cat([x1,x2]) + torch.randn(n,2)*0.05
    y = torch.cat([torch.zeros(n//2), torch.ones(n//2)]).long()
    p = torch.randperm(n); return x[p], y[p]

def data_checker(n=1000, freq=3):
    x = torch.rand(n,2)*2-1
    y = ((torch.sin(freq*math.pi*x[:,0])*torch.sin(freq*math.pi*x[:,1])) > 0).long()
    return x, y

def data_highfreq(n=1000):
    x = torch.linspace(-1,1,n).unsqueeze(1)
    y = torch.sin(20*x)+torch.sin(50*x)+0.5*torch.sin(100*x)
    return x, y

def data_memorize(n=200):
    return torch.randn(n, 8), torch.randn(n, 4)

def data_ood_train(n=800):
    x = torch.rand(n,2)*2-1
    y = torch.sin(3*math.pi*x[:,0]) * torch.cos(3*math.pi*x[:,1]) + x[:,0]*x[:,1]
    return x, y.unsqueeze(1)

def data_ood_test(n=300):
    x = torch.rand(n,2) + 1
    y = torch.sin(3*math.pi*x[:,0]) * torch.cos(3*math.pi*x[:,1]) + x[:,0]*x[:,1]
    return x, y.unsqueeze(1)

# ============================================================================
# MAIN
# ============================================================================

def main():
    print("="*80)
    print("  BENCHMARK v6: THE ADAPTIVE NEURON")
    print("  α(x)·periodic + (1-α(x))·linear — learns WHEN to use each")
    print("  3 seeds | vs Vanilla, SinGLU, Hybrid | + OOD + α analysis")
    print("="*80)
    
    N_H = 3
    models = {
        'Vanilla':  (VanillaMLP,  {}),
        'SinGLU':   (SinGLUNet,   {'omega_0': None}),
        'Hybrid':   (HybridNet,   {'omega_0': None}),
        'Adaptive': (AdaptiveNet, {'omega_0': None}),
    }
    
    tasks = [
        ("Complex Fn (4D)",  "reg", data_complex,  4,1, 5000, 300, 1e-3, 30.0, 750),
        ("Nested Fn (2D)",   "reg", data_nested,   2,1, 3000, 300, 1e-3, 20.0, 750),
        ("Spiral",           "clf", data_spiral,    2,2, 3000, 250, 1e-3, 15.0, 700),
        ("Checkerboard",     "clf", data_checker,   2,2, 3000, 250, 1e-3, 20.0, 700),
        ("High-Freq",        "reg", data_highfreq,  1,1, 8000, 300, 1e-3, 60.0, 700),
        ("Memorization",     "reg", data_memorize,  8,4, 5000, 400, 1e-3, 10.0, 200),
    ]
    
    all_results = {}
    alpha_analysis = {}
    
    for tname, ttype, dfn, ind, outd, budget, epochs, lr, omega, split in tasks:
        print(f"\n{'━'*80}")
        print(f"  {tname}  |  budget ~{budget:,}  |  {len(SEEDS)} seeds")
        print(f"{'━'*80}")
        
        # Find hidden dims
        hdims = {}
        for mn, (mc, mk) in models.items():
            kw = {k: (omega if v is None else v) for k,v in mk.items()}
            hdims[mn] = find_hidden(ind, outd, N_H, budget, mc, **kw)
        
        task_res = {}
        task_alphas = []
        
        for mn, (mc, mk) in models.items():
            kw = {k: (omega if v is None else v) for k,v in mk.items()}
            h = hdims[mn]
            scores = []
            
            for seed in SEEDS:
                set_seed(seed)
                x, y = dfn()
                if split >= len(x): xtr,ytr,xte,yte = x,y,x,y
                else: xtr,ytr,xte,yte = x[:split],y[:split],x[split:],y[split:]
                
                set_seed(seed + 100)
                model = mc(ind, outd, h, N_H, **kw)
                
                if ttype == 'reg':
                    s = train_reg(model, xtr, ytr, xte, yte, epochs, lr)
                else:
                    s = train_clf(model, xtr, ytr, xte, yte, epochs, lr)
                scores.append(s)
                
                # Collect α values from Adaptive model (last seed only)
                if mn == 'Adaptive' and seed == SEEDS[-1] and isinstance(model, AdaptiveNet):
                    model.eval()
                    with torch.no_grad():
                        alphas = model.get_all_alphas(xte[:100])
                        mean_alpha = torch.cat([a.mean(dim=0) for a in alphas]).mean().item()
                        std_alpha = torch.cat([a.std(dim=0) for a in alphas]).mean().item()
                        task_alphas = [mean_alpha, std_alpha]
            
            p = count_params(mc(ind, outd, h, N_H, **kw))
            task_res[mn] = {
                'mean': np.mean(scores), 'std': np.std(scores),
                'scores': scores, 'params': p, 'hidden': h
            }
        
        alpha_analysis[tname] = task_alphas
        
        # Print
        is_reg = ttype == 'reg'
        metric = "MSE ↓" if is_reg else "Acc ↑"
        
        if is_reg:
            best_mn = min(task_res, key=lambda k: task_res[k]['mean'])
        else:
            best_mn = max(task_res, key=lambda k: task_res[k]['mean'])
        
        print(f"\n  {'Model':<12} {'H':>4} {'Params':>7} {metric+' (mean±std)':>28}")
        print(f"  {'─'*56}")
        for mn, r in task_res.items():
            m, s = r['mean'], r['std']
            if is_reg:
                if m < 0.001: ms = f"{m:.2e}±{s:.1e}"
                else: ms = f"{m:.4f}±{s:.4f}"
            else:
                ms = f"{m:.1%}±{s:.3f}"
            mark = " ★" if mn == best_mn else ""
            print(f"  {mn:<12} {r['hidden']:>4} {r['params']:>7,} {ms:>28}{mark}")
        
        print(f"  → Winner: {best_mn}")
        
        if task_alphas:
            print(f"  → Adaptive α: mean={task_alphas[0]:.3f}, std={task_alphas[1]:.3f}"
                  f"  ({'mostly periodic' if task_alphas[0] > 0.6 else 'mostly linear' if task_alphas[0] < 0.4 else 'balanced'})")
        
        all_results[tname] = task_res
    
    # ================================================================
    # OOD TEST
    # ================================================================
    print(f"\n{'━'*80}")
    print(f"  OOD: Train [-1,1] → Test [1,2]")
    print(f"  f(x1,x2) = sin(3π·x1)·cos(3π·x2) + x1·x2")
    print(f"  Does Adaptive auto-fallback to linear on OOD?")
    print(f"{'━'*80}")
    
    budget_ood = 5000
    ood_res = {}
    ood_alphas = {}
    
    for mn, (mc, mk) in models.items():
        kw = {k: (20.0 if v is None else v) for k,v in mk.items()}
        h = find_hidden(2, 1, N_H, budget_ood, mc, **kw)
        
        id_scores, ood_scores = [], []
        for seed in SEEDS:
            set_seed(seed)
            xtr, ytr = data_ood_train()
            set_seed(seed + 50)
            xid = torch.rand(200,2)*2-1
            yid = (torch.sin(3*math.pi*xid[:,0]) * torch.cos(3*math.pi*xid[:,1]) + xid[:,0]*xid[:,1]).unsqueeze(1)
            set_seed(seed + 50)
            xood, yood = data_ood_test()
            
            set_seed(seed + 100)
            model = mc(2, 1, h, N_H, **kw)
            s_id = train_reg(model, xtr, ytr, xid, yid, 300, 1e-3)
            model.eval()
            with torch.no_grad():
                s_ood = F.mse_loss(model(xood), yood).item()
            id_scores.append(s_id)
            ood_scores.append(s_ood)
            
            # Get α on ID vs OOD for adaptive
            if mn == 'Adaptive' and seed == SEEDS[-1] and isinstance(model, AdaptiveNet):
                model.eval()
                with torch.no_grad():
                    a_id = model.get_all_alphas(xid[:100])
                    a_ood = model.get_all_alphas(xood[:100])
                    ood_alphas['ID'] = torch.cat([a.mean(dim=0) for a in a_id]).mean().item()
                    ood_alphas['OOD'] = torch.cat([a.mean(dim=0) for a in a_ood]).mean().item()
        
        p = count_params(mc(2, 1, h, N_H, **kw))
        ood_res[mn] = {
            'id_mean': np.mean(id_scores), 'id_std': np.std(id_scores),
            'ood_mean': np.mean(ood_scores), 'ood_std': np.std(ood_scores),
            'params': p,
            'degradation': np.mean(ood_scores) / max(np.mean(id_scores), 1e-10),
        }
    
    best_ood_mn = min(ood_res, key=lambda k: ood_res[k]['ood_mean'])
    
    print(f"\n  {'Model':<12} {'Params':>7} {'ID MSE':>14} {'OOD MSE':>14} {'Degrad.':>9}")
    print(f"  {'─'*58}")
    for mn, r in ood_res.items():
        mark = " ★" if mn == best_ood_mn else ""
        print(f"  {mn:<12} {r['params']:>7,} {r['id_mean']:>9.4f}±{r['id_std']:.3f}"
              f" {r['ood_mean']:>9.4f}±{r['ood_std']:.3f} {r['degradation']:>8.1f}x{mark}")
    
    print(f"  → Best OOD: {best_ood_mn}")
    
    if ood_alphas:
        print(f"\n  ADAPTIVE α SHIFT (the key test):")
        print(f"    α on in-distribution data:  {ood_alphas.get('ID', 'N/A'):.3f}")
        print(f"    α on out-of-distribution:   {ood_alphas.get('OOD', 'N/A'):.3f}")
        shift = ood_alphas.get('OOD', 0.5) - ood_alphas.get('ID', 0.5)
        if shift < -0.05:
            print(f"    → α DROPPED by {abs(shift):.3f} on OOD → model learned to reduce periodic! ✅")
        elif shift > 0.05:
            print(f"    → α INCREASED by {shift:.3f} on OOD → model did NOT learn fallback ❌")
        else:
            print(f"    → α roughly stable (shift={shift:+.3f}) → routing not input-sensitive here")
    
    all_results['OOD'] = {mn: {'mean': r['ood_mean'], 'std': r['ood_std']} 
                          for mn, r in ood_res.items()}
    
    # ================================================================
    # α ANALYSIS SUMMARY
    # ================================================================
    print(f"\n{'━'*80}")
    print(f"  WHAT DID α LEARN? (Routing analysis per task)")
    print(f"{'━'*80}")
    
    print(f"\n  {'Task':<22} {'Mean α':>8} {'Std α':>8} {'Interpretation':<30}")
    print(f"  {'─'*70}")
    for tname, aa in alpha_analysis.items():
        if aa:
            interp = ("→ mostly PERIODIC" if aa[0] > 0.6 
                      else "→ mostly LINEAR" if aa[0] < 0.4 
                      else "→ BALANCED mix")
            print(f"  {tname:<22} {aa[0]:>8.3f} {aa[1]:>8.3f} {interp:<30}")
    
    if ood_alphas:
        id_val = ood_alphas.get('ID', 0)
        ood_val = ood_alphas.get('OOD', 0)
        print(f"  {'OOD (out-of-dist)':<22} {ood_val:>8.3f} {'--':>8} "
              f"shifted from {id_val:.3f}")
    
    # ================================================================
    # GRAND SUMMARY
    # ================================================================
    print(f"\n{'='*80}")
    print(f"  GRAND SUMMARY")
    print(f"{'='*80}")
    
    win_counts = {k: 0 for k in models}
    
    print(f"\n  {'Task':<20}", end="")
    for mn in models: print(f" {mn:>12}", end="")
    print(f" {'Winner':>10}")
    print(f"  {'─'*72}")
    
    for tname, tr in all_results.items():
        if tname == 'OOD':
            scores = {k: v['mean'] for k,v in tr.items()}
            winner = min(scores, key=scores.get)
            win_counts[winner] += 1
            row = f"  {tname:<20}"
            for mn in models: row += f" {scores[mn]:>12.4f}"
            row += f" {'→'+winner:>10}"
            print(row)
            continue
            
        scores = {k: v['mean'] for k,v in tr.items()}
        max_s = max(scores.values())
        is_clf = max_s > 0.5 and max_s <= 1.0 and min(scores.values()) >= 0
        if min(scores.values()) < 0.001: is_clf = False
        
        if is_clf: winner = max(scores, key=scores.get)
        else: winner = min(scores, key=scores.get)
        win_counts[winner] += 1
        
        row = f"  {tname:<20}"
        for mn in models:
            s = scores[mn]
            if is_clf: row += f" {s:>11.1%}"
            elif s < 0.001: row += f" {s:>11.2e}"
            else: row += f" {s:>11.4f}"
        row += f"  {'→'+winner:>10}"
        print(row)
    
    print(f"\n  {'─'*72}")
    print(f"  WIN COUNTS:")
    for mn, c in sorted(win_counts.items(), key=lambda x: -x[1]):
        print(f"    {mn:<14} {c} wins  {'█'*c*3}")
    
    print(f"""
  ╔════════════════════════════════════════════════════════════════════════════╗
  ║  THE ADAPTIVE NEURON: VERDICT                                            ║
  ║                                                                          ║
  ║  The question: can a neuron that LEARNS when to be periodic vs linear    ║
  ║  get the best of both worlds?                                            ║
  ║                                                                          ║
  ║  Check the α values above — they tell the whole story:                   ║
  ║  • If α is high on periodic tasks and low on linear tasks → it works    ║
  ║  • If α is ~0.5 everywhere → routing didn't learn, it's just averaging  ║
  ║  • If α shifts down on OOD → it learned to auto-fallback → that's huge  ║
  ╚════════════════════════════════════════════════════════════════════════════╝
""")
    
    # Save
    save = {'main_tasks': {}, 'ood': {}, 'alpha_analysis': {}, 'ood_alphas': ood_alphas}
    for tname, tr in all_results.items():
        save['main_tasks'][tname] = {
            mn: {'mean': float(r['mean']), 'std': float(r.get('std', 0)),
                 'scores': [float(s) for s in r.get('scores', [r['mean']])],
                 'params': r.get('params', 0), 'hidden': r.get('hidden', 0)}
            for mn, r in tr.items()
        }
    save['alpha_analysis'] = {k: v for k,v in alpha_analysis.items()}
    save['ood'] = {mn: {k: float(v) if isinstance(v, (float, np.floating)) else v
                        for k,v in r.items()} for mn, r in ood_res.items()}
    
    with open('/app/results_v6.json', 'w') as f:
        json.dump(save, f, indent=2, default=str)
    print("  Results saved to /app/results_v6.json")


if __name__ == "__main__":
    main()