RESULT 55: 单侧偏置首次直接定量(线性∝β,0.3-1%@3e-3,2.7σ;BBP线真空点补上); 逐样本符号+batch平均=无偏(用户命题四层两β全验证,偏置消到噪声地板,零代价); 预登记推论:若bsign挡不住下沉则cent护甲非消偏

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn

2 files changed, 199 insertions, 0 deletions

diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md
index 6dd93e5..4cd5f3a 100644
--- a/docs/campaign/CASCADE_ABLATION_PLAN.md
+++ b/docs/campaign/CASCADE_ABLATION_PLAN.md
@@ -568,6 +568,23 @@ direction), not training-under-fault; wave-2 = co-training with faults injected
 
 
 
+### RESULT 55 (2026-07-20): SINGLE-SIDED BIAS DIRECTLY QUANTIFIED (the BBP-thread gap) —
+real, linear in beta, 0.3-1% of grad norm @3e-3; PER-SAMPLE SIGN + BATCH AVERAGE = UNBIASED
+(user's claim VERIFIED at zero measurable cost).
+probe_bias @plain2_s35000, 16 batches, beta {3e-3,1e-2}, selfcheck 3e-9: plain rel_bias
+|mean(gEP-gBP)|/|mean(gBP)|: w2_b11 0.0035->0.0114 (x3.26 ~ beta ratio 3.33, 2.7sigma over sem
+0.0013->0.0042); qkv_b0 0.0106->0.0325; all four layers >sem with ~x3.1-3.4 scaling => the
+Laborieux O(beta) term measured directly on a 72M transformer. PSIGN (per-sample random nudge
+sign, read re-flipped): bias -> noise floor in EVERY row (w2_b11: 0.0014 vs sem 0.0014), both
+betas; sem unchanged vs plain => debiasing is FREE (no extra phase, no variance penalty at B=8;
+improves with B — the scale-native estimator, supersedes per-batch bsign in theory).
+Batch size alone does NOT debias (bias survives expectation); it only debiases AFTER the
+per-sample sign flip makes the O(beta) term batch-antisymmetric — user's "sign+大batch=无偏"
+is the correct composition. PRE-REGISTERED corollary: bias is only 0.3-1%, so if the RUNNING
+bsign crown fails to stop the ceiling sink (user predicts it will fail), cent's armor is NOT
+debiasing — it must be a variance/geometry property of the symmetric read; new puzzle.
+In flight: fw72m_bsign (GPU1/3, per-batch sign), fw72m_ratchet (GPU0 queued, halve-only).
+
 ### RESULT 54 (2026-07-20): THE CEILING SINKS ∝1/σ² AS TRAINING SHARPENS — no fixed β survives;
 "fixed β never explodes" was FALSE (it STALLS: endgame skip-storm = frozen training). β must DESCEND.
 User caught it: "固定β永不炸个屁！你一直skip还训不训了？" — dead right.
diff --git a/ep_run/probe_bias.py b/ep_run/probe_bias.py
new file mode 100644
index 0000000..898f8f6
--- /dev/null
+++ b/ep_run/probe_bias.py
@@ -0,0 +1,182 @@
+"""Direct measurement of the systematic single-sided EP bias (the audit gap: BBP probes
+measured NOISE thoroughly, discarded the mean/bias component unreported) + verification of the
+user's claim: PER-SAMPLE random nudge sign makes the batch-averaged estimator unbiased at O(beta).
+Three estimators at fixed ckpt, N batches, two betas:
+  plain      : all samples nudged +beta            -> bias = beta*E[B] (systematic, batch-indep)
+  persample  : sample i nudged s_i*beta, read re-flipped by s_i -> odd-order bias cancels
+               WITHIN each batch (sqrt(B) suppression per step, 0 in expectation)
+  BP         : exact reference.
+Self-check: persample with all s=+1 must equal plain exactly.
+Report per layer family: |mean(ghat-gbp)| / |mean(gbp)| (systematic), across-batch sem (noise
+floor), for beta in {3e-3, 1e-2} — plain's bias should scale ~x3.3, persample's should sit at
+the noise floor. Read-only, coexists on GPU0."""
+import argparse, pickle
+import numpy as np, torch, torch.nn as nn, torch.nn.functional as F
+from pathlib import Path
+
+ap = argparse.ArgumentParser()
+ap.add_argument('--ckpt', default='runs/fw72m_plain2_s35000.pt')
+ap.add_argument('--K', type=int, default=3)
+ap.add_argument('--betas', default='3e-3,1e-2')
+ap.add_argument('--nb', type=int, default=16)
+a = ap.parse_args()
+dev = 'cuda'
+torch.manual_seed(11)
+B, T = 8, 256
+
+class RMSNorm(nn.Module):
+    def __init__(self, C, eps=1e-6):
+        super().__init__(); self.g = nn.Parameter(torch.ones(C)); self.eps = eps
+    def forward(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.g
+
+class SwiGLU(nn.Module):
+    def __init__(self, C):
+        super().__init__()
+        h = ((8 * C // 3) + 63) // 64 * 64
+        self.w1 = nn.Linear(C, h, bias=False); self.w3 = nn.Linear(C, h, bias=False)
+        self.w2 = nn.Linear(h, C, bias=False)
+    def forward(self, x):
+        return self.w2(F.silu(self.w1(x)) * self.w3(x))
+
+class Olmo2Attn(nn.Module):
+    def __init__(self, C, H, T):
+        super().__init__()
+        self.H, self.hd = H, C // H
+        self.qkv = nn.Linear(C, 3 * C, bias=False); self.proj = nn.Linear(C, C, bias=False)
+        self.qn, self.kn = RMSNorm(C), RMSNorm(C)
+        inv = 1.0 / (500000.0 ** (torch.arange(0, self.hd, 2).float() / self.hd))
+        fr = torch.outer(torch.arange(T).float(), inv)
+        self.register_buffer('rc', fr.cos(), persistent=False)
+        self.register_buffer('rs', fr.sin(), persistent=False)
+    def rope(self, x):
+        Tn = x.shape[2]
+        x1, x2 = x[..., ::2], x[..., 1::2]
+        c, s = self.rc[None, None, :Tn].to(x.dtype), self.rs[None, None, :Tn].to(x.dtype)
+        return torch.stack((x1 * c - x2 * s, x1 * s + x2 * c), dim=-1).flatten(-2)
+    def forward(self, x):
+        Bn, Tn, C = x.shape
+        q, k, v = self.qkv(x).split(C, dim=2)
+        q, k = self.qn(q), self.kn(k)
+        q = self.rope(q.view(Bn, Tn, self.H, self.hd).transpose(1, 2))
+        k = self.rope(k.view(Bn, Tn, self.H, self.hd).transpose(1, 2))
+        v = v.view(Bn, Tn, self.H, self.hd).transpose(1, 2)
+        y = F.scaled_dot_product_attention(q, k, v, is_causal=True)
+        return self.proj(y.transpose(1, 2).contiguous().view(Bn, Tn, C))
+
+class Olmo2Block(nn.Module):
+    def __init__(self, C, H, T):
+        super().__init__()
+        self.attn = Olmo2Attn(C, H, T); self.ff = SwiGLU(C)
+        self.na, self.nf = RMSNorm(C), RMSNorm(C)
+    def forward(self, z):
+        z = z + self.na(self.attn(z))
+        return z + self.nf(self.ff(z))
+
+ck = torch.load(a.ckpt, map_location=dev, weights_only=False)
+cfg = ck['config']; C, H, L = cfg['C'], cfg['H'], cfg['L']
+DD = Path('/home/yurenh2/ept/ep_run/data') / cfg.get('data', 'fineweb_edu')
+vocab = pickle.load(open(DD / 'meta.pkl', 'rb'))['vocab_size']
+data = np.memmap(DD / 'val.bin', dtype=np.uint16, mode='r')
+tok = nn.Embedding(vocab, C).to(dev); tok.load_state_dict(ck['tok'])
+blocks = nn.ModuleList([Olmo2Block(C, H, T) for _ in range(L)]).to(dev)
+blocks.load_state_dict(ck['blocks'], strict=False)
+W_out = ck['wout'].to(dev); ln_f = RMSNorm(C).to(dev); ln_f.load_state_dict(ck['lnf'])
+NBT = B * T
+params = list(blocks.parameters())
+names = [n for n, _ in blocks.named_parameters()]
+
+def readout(z): return ln_f(z) @ W_out.t()
+
+def get_batch(g):
+    ix = torch.randint(len(data) - T - 1, (B,), generator=g)
+    x = torch.stack([torch.from_numpy(data[i:i + T].astype(np.int64)) for i in ix])
+    y = torch.stack([torch.from_numpy(data[i + 1:i + 1 + T].astype(np.int64)) for i in ix])
+    return x.to(dev), y.to(dev)
+
+def grads_from(outs_list, cots):
+    obj = sum((o * c.detach()).sum() for o, c in zip(outs_list, cots))
+    gs = torch.autograd.grad(obj, params, allow_unused=True, retain_graph=True)
+    return [g.float() if g is not None else torch.zeros_like(p) for g, p in zip(gs, params)]
+
+def estimators(x, y, beta, signs):
+    """signs: (B,) tensor of +-1. Returns (g_bp, g_ep) where g_ep uses per-sample sign nudges."""
+    z = tok(x)
+    zs_bp = []
+    for b in blocks:
+        z = b(z); zs_bp.append(z)
+    ce = F.cross_entropy(readout(zs_bp[-1]).reshape(-1, vocab), y.reshape(-1))
+    g_bp = [g.float() for g in torch.autograd.grad(ce, params, retain_graph=True)]
+    f_ins, f_outs = [], []
+    prev = tok(x).detach()
+    for b in blocks:
+        i = prev.detach().requires_grad_(True)
+        o = b(i)
+        f_ins.append(i); f_outs.append(o)
+        prev = o.detach()
+    ins, outs = f_ins, f_outs
+    zs = [o.detach().float() for o in f_outs]
+    d = [None] * L
+    sB = signs.view(B, 1, 1).float()
+    for k in range(a.K):
+        zc = zs[L - 1].detach().requires_grad_(True)
+        # per-sample signed top objective: sum_i s_i * (sum_t loss_it)  (token-sum, matches beta*NBT*mean scaling)
+        ce_tok = F.cross_entropy(readout(zc).reshape(-1, vocab), y.reshape(-1), reduction='none').view(B, T)
+        obj_top = (signs.float()[:, None] * ce_tok).sum()
+        d[L - 1] = (-beta * torch.autograd.grad(obj_top, zc)[0]).detach().float()
+        for l in range(L - 2, -1, -1):
+            d[l] = torch.autograd.grad(outs[l + 1], ins[l + 1], grad_outputs=d[l + 1],
+                                       retain_graph=True)[0].detach().float()
+        prev = tok(x).detach()
+        n_ins, n_outs = [], []
+        for l in range(L):
+            i = prev.detach().requires_grad_(True)
+            o = blocks[l](i)
+            n_ins.append(i); n_outs.append(o)
+            zs[l] = o.detach().float() + d[l]
+            prev = zs[l]
+        ins, outs = n_ins, n_outs
+    md = [-(sB * di) / (beta * NBT) for di in d]   # re-flip per sample; scale matches trainer
+    g_ep = grads_from(outs, md)
+    return g_bp, g_ep
+
+SEL = {'qkv_b0': names.index('0.attn.qkv.weight'), 'qkv_b8': names.index('8.attn.qkv.weight'),
+       'w2_b6': names.index('6.ff.w2.weight'), 'w2_b11': names.index('11.ff.w2.weight')}
+betas = [float(s) for s in a.betas.split(',')]
+
+# self-check: persample with all +1 == plain
+g0 = torch.Generator().manual_seed(99)
+x, y = get_batch(g0)
+_, gA = estimators(x, y, 3e-3, torch.ones(B, dtype=torch.long, device=dev))
+_, gB = estimators(x, y, 3e-3, torch.ones(B, dtype=torch.long, device=dev))
+same = max(float((a_ - b_).abs().max()) for a_, b_ in zip(gA, gB))
+print(f'selfcheck determinism max|diff|={same:.2e} (must be ~0)', flush=True)
+
+gbatch = torch.Generator().manual_seed(1234)
+sgen = torch.Generator(device='cpu').manual_seed(777)
+batches = [get_batch(gbatch) for _ in range(a.nb)]
+for beta in betas:
+    acc = {k: {'plain': [], 'psign': [], 'bp': []} for k in SEL}
+    for (x, y) in batches:
+        ones = torch.ones(B, dtype=torch.long, device=dev)
+        s = (torch.randint(0, 2, (B,), generator=sgen) * 2 - 1).to(dev)
+        g_bp, g_pl = estimators(x, y, beta, ones)
+        _, g_ps = estimators(x, y, beta, s)
+        for k, i in SEL.items():
+            acc[k]['plain'].append((g_pl[i] - g_bp[i]).detach().cpu())
+            acc[k]['psign'].append((g_ps[i] - g_bp[i]).detach().cpu())
+            acc[k]['bp'].append(g_bp[i].detach().cpu())
+        del g_bp, g_pl, g_ps
+        torch.cuda.empty_cache()
+    print(f'\n=== beta={beta:g}  (nb={a.nb}; rel_bias = |mean delta| / |mean g_bp|; sem = noise floor) ===', flush=True)
+    print(f'{"layer":>8} {"plain rel_bias":>15} {"plain sem":>10} {"psign rel_bias":>15} {"psign sem":>10}', flush=True)
+    for k in SEL:
+        gb = torch.stack(acc[k]['bp']).mean(0); gnorm = float(gb.norm())
+        out = []
+        for est in ('plain', 'psign'):
+            D = torch.stack(acc[k][est])
+            mean = D.mean(0); rel = float(mean.norm()) / max(gnorm, 1e-30)
+            sem = float((D - mean).pow(2).sum(dim=(1, 2)).mean().sqrt()) / (a.nb ** 0.5) / max(gnorm, 1e-30)
+            out += [rel, sem]
+        print(f'{k:>8} {out[0]:>15.4f} {out[1]:>10.4f} {out[2]:>15.4f} {out[3]:>10.4f}', flush=True)
+print('\nBIAS_DONE', flush=True)