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, 19 insertions, 0 deletions

diff --git a/ep_run/ep_c3_watch.py b/ep_run/ep_c3_watch.py
new file mode 100644
index 0000000..a81fd44
--- /dev/null
+++ b/ep_run/ep_c3_watch.py
@@ -0,0 +1,19 @@
+import time, os, re, subprocess
+os.chdir("/home/yurenh2/ept/ep_run"); LOG="runs/ep_c3_scratch.log"
+def alive(): return subprocess.run(["pgrep","-f","ep_c3_scratch.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
+    m=re.search(r"step\s+(\d+)/.*val CE ([\d.eE+-]+).*res=([\d.eE+-]+)", ls[-1])
+    return (int(m.group(1)),float(m.group(2)),float(m.group(3))) if m else None
+fired=None; t0=time.time()
+while fired is None and time.time()-t0<24*3600:
+    time.sleep(180)
+    d=latest()
+    if not alive(): fired=f"ep_c3 exited; last {d}"; break
+    if not d: continue
+    step,val,res=d
+    if res>0.2 or val>15: fired=f"ep_c3(c=3) DIVERGED step {step} val {val:.2f} res {res:.2e} -> trained-damping did NOT prevent it"; break
+    if val<2.5: fired=f"ep_c3(c=3) reached val {val:.4f} step {step} -> past redx's ~2.7 blow zone, damping fix HOLDING"; break
+print("=== EP_C3 WATCHER ==="); print(fired or "24h timeout"); print("last:", latest())