Initial commit: ept — backprop-free equilibrium transformer (EP)

Code (ep_run/), organized docs (docs/{method,campaign,hardware,outreach,paper}),
analysis scripts (scripts/), ONBOARDING.md entry point. Large data/checkpoints
git-ignored (share separately).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn

1 files changed, 23 insertions, 0 deletions

diff --git a/ep_run/ep_fast_check.py b/ep_run/ep_fast_check.py
new file mode 100644
index 0000000..fe82955
--- /dev/null
+++ b/ep_run/ep_fast_check.py
@@ -0,0 +1,23 @@
+import time, os, re, subprocess
+os.chdir("/home/yurenh2/ept/ep_run"); LOG="runs/ep_resreg_fast.log"; TARGET=2.30
+def alive(): return subprocess.run(["pgrep","-f","ckpt runs/ep_resreg_fast.pt"],capture_output=True).returncode==0
+def latest():
+    try: ls=[l for l in open(LOG) if l.startswith("step")]
+    except Exception: return None
+    if not ls: return None
+    l=ls[-1]
+    ms=re.search(r"step\s+(\d+)",l); mv=re.search(r"val CE ([\d.eE+-]+)",l)
+    mb=re.search(r"best ([\d.eE+-]+)",l); mr=re.search(r"res=([\d.eE+-]+)",l); mj=re.search(r"jr=([\d.eE+-]+)",l)
+    if not (ms and mv and mb and mr): return None
+    return (int(ms.group(1)),float(mv.group(1)),float(mb.group(1)),float(mr.group(1)),float(mj.group(1)) if mj else 0.0)
+fired=None; t0=time.time(); hires=0.0
+while fired is None and time.time()-t0<12*3600:
+    time.sleep(90)
+    d=latest()
+    if not alive(): fired=f"ep_resreg_fast EXITED; last {d}"; break
+    if not d: continue
+    step,val,best,res,jr=d
+    hires=max(hires,res)
+    if val>12 or res>0.3: fired=f"ep_resreg_fast BLEW @step{step} val{val:.2f} res{res:.1e} best{best:.4f} peakres{hires:.1e}"; break
+    if best<=TARGET: fired=f"ep_resreg_fast reached {TARGET}: step{step} best{best:.4f} val{val:.4f} res{res:.1e} peakres{hires:.1e} -> resreg broke past the jacreg-stall region"; break
+print(f"=== EP_RESREG_FAST -> {TARGET} ==="); print(fired or "12h timeout"); print("last:",latest())