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, 26 insertions, 0 deletions

diff --git a/ep_run/ep_jacreg_spike.py b/ep_run/ep_jacreg_spike.py
new file mode 100644
index 0000000..a72b89c
--- /dev/null
+++ b/ep_run/ep_jacreg_spike.py
@@ -0,0 +1,26 @@
+import time, os, re, subprocess
+os.chdir("/home/yurenh2/ept/ep_run"); LOG="runs/ep_jacreg.log"
+def alive(): return subprocess.run(["pgrep","-f","ckpt runs/ep_jacreg.pt"],capture_output=True).returncode==0
+def steps():
+    out=[]
+    try:
+        for l in open(LOG):
+            if not l.startswith("step"): continue
+            ms=re.search(r"step\s+(\d+)",l); mv=re.search(r"val CE ([\d.]+)",l)
+            mj=re.search(r"jr=([\d.eE+-]+)",l); mr=re.search(r"res=([\d.eE+-]+)",l)
+            if ms and mv and mj: out.append((int(ms.group(1)),float(mv.group(1)),float(mj.group(1)),float(mr.group(1)) if mr else 0))
+    except Exception: pass
+    return out
+fired=None; t0=time.time(); hi=0; prev=None; seen=set([r[0] for r in steps()])  # ignore already-seen (incl the 6250 spike)
+while fired is None and time.time()-t0<4*3600:
+    time.sleep(60)
+    if not alive(): fired=f"EXITED last={steps()[-1] if steps() else None}"; break
+    for r in steps():
+        if r[0] in seen: continue
+        seen.add(r[0]); step,val,jr,res=r
+        if val>15 or res>0.3: fired=f"DIVERGED @{step} val{val:.2f} res{res:.1e} jr{jr:.1f}"; break
+        hi = hi+1 if jr>=8 else 0
+        if jr<2.0 and val<2.55: fired=f"SUPPRESSED @{step}: jr relaxed to {jr:.1f}, CE recovered {val:.3f} (best2.4381) res{res:.1e} -> controller WON the spike"; break
+        if hi>=3: fired=f"jr SATURATING @{step}: jr>=8 for {hi} logged-steps (now {jr:.1f}), val{val:.3f} res{res:.1e} -> controller maxed, not relaxing (early hijack/saturation)"; break
+    if fired: break
+print("=== EP_JACREG SPIKE-RECOVERY ==="); print(fired or "4h timeout"); print("last5:", steps()[-5:])