RESULT 78: read_lin梯度级验证==fp64(fp32下泄漏塌零,读出舍入=全部病因); 附属谱仪:前沿随训练下压+seed轴证'比值决定非层号'+阈值带2-5e-7; 修复入训练路径; CE级仪器终审在飞(预注册~+100%关闭)

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

1 files changed, 8 insertions, 1 deletions

diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py
index 3cb9494..ed164ce 100644
--- a/ep_run/casc_eq_train.py
+++ b/ep_run/casc_eq_train.py
@@ -699,7 +699,14 @@ def ep_step(x, y):
         gs = torch.autograd.grad(obj, all_params, allow_unused=True)
     elif EST == 'single':
         E = 0.0
-        for z, o in zip(zp, last_outs): E = E + 0.5 * ((z.detach().float() - o.float()) ** 2).sum()   # fp32 accumulation (bf16-safe; no-op in fp32)
+        if args.read_lin:
+            # FP-FLOOR FIX (RESULT 77/78): the read identity z - o = d holds algebraically but not in
+            # fp32 — (z - o) is a ROUNDED copy of d (components below eps·|o| are erased; the C768
+            # top-half leak). The d tensors themselves are full-precision; the linear form -<d,o> has
+            # the identical theta-derivative and bypasses the rounding. Probe-verified == fp64.
+            for dl, o in zip(GOV['_last_d'], last_outs): E = E - (dl.detach().float() * o.float()).sum()
+        else:
+            for z, o in zip(zp, last_outs): E = E + 0.5 * ((z.detach().float() - o.float()) ** 2).sum()   # fp32 accumulation (bf16-safe; no-op in fp32)
         obj = E / (NBT * beta_t) + obj_loss(readout(zp[-1].detach()).reshape(-1, vocab), y.reshape(-1))
         gs = torch.autograd.grad(obj, all_params, allow_unused=True)
     else: