gpt-5.6-sol

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/submission/curate.py\\n@@\\n-    return q, ds, max(2,int(len(raw)/4.05)+1), fp\\n+    return q, ds, max(2,int(len(raw)/4.05)+1), fp, manual\\n@@\\n-          q,ds,est,fp=z; rows.append({'id':o['id'],'q':q,'d':ds,'est':est,'fp':fp})\\n-    # Assign specialized registers only when evidence clears a useful floor;\\n-    # remaining good prose forms the web bucket.\\n-    buckets={x:[] for x in ('encyclopedic','web','news','technical')}\\n+          q,ds,est,fp,manual=z; rows.append({'id':o['id'],'q':q,'d':ds,'m':manual,'est':est,'fp':fp})\\n+\\n+    names=('encyclopedic','web','news','technical')\\n+    # Ranker B: register assignment and relevance learned from the disclosed\\n+    # four-quarter target.  It is particularly effective for news and Q&A.\\n+    buckets={x:[] for x in names}\\n     for r in rows:\\n       dom=max(r['d'],key=r['d'].get); strength=r['d'][dom]\\n-      # Domain relevance is a modest bonus; quality remains dominant.\\n       r['score']=r['q'] + .32*max(-2.0,min(3.0,strength))\\n       if r['q']>-1.0: buckets[dom].append(r)\\n     for b in buckets.values(): b.sort(key=lambda r:(-r['score'],r['id']))\\n-\\n-    # Greedy equal quotas with cross-domain normalized-prefix deduplication.\\n-    quota=int(a.budget*1.35/4); chosen={k:[] for k in buckets}; seen=set()\\n-    for dom in ('technical','news','encyclopedic','web'):\\n-      total=0\\n-      for r in buckets[dom]:\\n-        if r['fp'] in seen: continue\\n-        seen.add(r['fp']); chosen[dom].append(r); total+=r['est']\\n-        if total>=quota: break\\n-    # Weighted round-robin by estimated tokens, so every prefix is balanced.\\n-    order=[]; idx={k:0 for k in chosen}; used={k:0 for k in chosen}; active=set(chosen)\\n-    target=a.budget*1.50\\n-    while sum(used.values())<target and active:\\n-      dom=min(active,key=lambda k: used[k])\\n-      if idx[dom]>=len(chosen[dom]):\\n-        active.remove(dom); continue\\n-      r=chosen[dom][idx[dom]];idx[dom]+=1;order.append(r['id']);used[dom]+=r['est']\\n-      if all(idx[k]>=len(chosen[k]) for k in chosen): break\\n+\\n+    def ordered_prefix(bs, quota, target):\\n+      chosen={k:[] for k in bs}; seen=set()\\n+      for dom in ('technical','news','encyclopedic','web'):\\n+        total=0\\n+        for r in bs[dom]:\\n+          if r['fp'] in seen: continue\\n+          seen.add(r['fp']);chosen[dom].append(r);total+=r['est']\\n+          if total>=quota: break\\n+      out=[];idx={k:0 for k in chosen};used={k:0 for k in chosen};active=set(chosen)\\n+      while sum(used.values())<target and active:\\n+        dom=min(active,key=lambda k:(used[k],k))\\n+        if idx[dom]>=len(chosen[dom]): active.remove(dom);continue\\n+        r=chosen[dom][idx[dom]];idx[dom]+=1;out.append(r);used[dom]+=r['est']\\n+      return out\\n+\\n+    B=ordered_prefix(buckets,int(a.budget*2.0/4),a.budget*1.50)\\n+\\n+    # Ranker A: conservative manual semantic cues.  It proved more robust for\\n+    # encyclopedia/general prose than fine target-style matching.\\n+    mb={x:[] for x in names}\\n+    for r in rows:\\n+      d3={k:r['m'][k] for k in ('technical','news','encyclopedic')}\\n+      best=max(d3,key=d3.get); strength=d3[best]\\n+      if (best=='technical' and strength>=1.35) or (best=='news' and strength>=1.12) or (best=='encyclopedic' and strength>=.92): dom=best\\n+      else: dom='web'\\n+      r['mscore']=r['q']+(0 if dom=='web' else .42*min(3.0,strength))\\n+      if r['q']>-1.0: mb[dom].append(r)\\n+    for b in mb.values(): b.sort(key=lambda r:(-r['mscore'],r['id']))\\n+    A=ordered_prefix(mb,int(a.budget*1.35/4),a.budget*1.18)\\n+\\n+    # Hybrid by target-assigned register, then token-balanced interleaving.\\n+    src={'encyclopedic':A,'web':A,'news':B,'technical':B}; chosen={}\\n+    for dom in names:\\n+      chosen[dom]=[r for r in src[dom] if max(r['d'],key=r['d'].get)==dom]\\n+      have={r['id'] for r in chosen[dom]}\\n+      if sum(r['est'] for r in chosen[dom])<4_500_000:\\n+        chosen[dom]+=[r for r in B if r['id'] not in have and max(r['d'],key=r['d'].get)==dom]\\n+    order=[];idx={k:0 for k in chosen};used={k:0 for k in chosen};active=set(chosen);seen=set()\\n+    while sum(used.values())<16_000_000 and active:\\n+      dom=min(active,key=lambda k:(used[k],k))\\n+      if idx[dom]>=len(chosen[dom]):active.remove(dom);continue\\n+      r=chosen[dom][idx[dom]];idx[dom]+=1\\n+      if r['id'] in seen:continue\\n+      seen.add(r['id']);order.append(r['id']);used[dom]+=r['est']\\n@@\\n-    json.dump({'criterion':'quality + four-register balance + prefix dedup','ids':len(order),'estimated_tokens':sum(used.values()),'estimated_by_domain':used,'available_by_domain':{k:len(v) for k,v in buckets.items()}},open(a.report,'w'),indent=2)\\n+    json.dump({'criterion':'hybrid quality/semantic and disclosed-target log-odds rankers; four-register balance; prefix dedup','ids':len(order),'estimated_tokens':sum(used.values()),'estimated_by_domain':used,'available_by_domain':{k:len(v) for k,v in buckets.items()}},open(a.report,'w'),indent=2)\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"python submission/curate.py\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "70",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":61316,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "71",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":61316,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "72",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":61316,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "73",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":61316,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "74",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":61316,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"python - <<'PY'\\nimport json\\nA=json.load(open('/tmp/hybridsel.json'));B=json.load(open('submission/selection.json'))\\nprint(len(A),len(B),'same order',A==B,'same set',set(A)==set(B),'prefix equal',next((i for i,(a,b) in enumerate(zip(A,B)) if a!=b),len(A)))\\nprint('unique',len(B)==len(set(B)),'range',min(B),max(B))\\nPY\\npython pack_selection.py submission/selection.json /tmp/final.npy\\nsha256sum submission/selection.json submission/curate.py\\ncat /workspace/.timer/remaining_secs\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "76",
  "yield_time_ms": 30000,
  "max_tokens": 5000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":79044,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":5000}); text(JSON.stringify(r))\n"
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"PYTHONPATH=/workspace python /tmp/train_domains.py --train_npy /tmp/final.npy --heldout_npy data/multi_dev.npy --out_json /tmp/final.json --max_iters 3000 --seed 1337\",\"workdir\":\"/workspace\",\"yield_time_ms\":30000,\"max_output_tokens\":3000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "78",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.write_stdin({\"session_id\":91539,\"chars\":\"\",\"yield_time_ms\":30000,\"max_output_tokens\":3000}); text(JSON.stringify(r))\n"
}

{
  "cell_id": "79",
  "yield_time_ms": 30000,
  "max_tokens": 3000
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"python - <<'PY'\\nimport json\\nA=json.load(open('/tmp/hybridsel.json'));B=json.load(open('submission/selection.json'))\\nprint('EXP',A[:20]);print('DET',B[:20])\\nPY\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":3000}); text(r.output)\n"
}

{
  "input": "const r = await tools.exec_command({\"cmd\":\"python - <<'PY'\\nimport json\\nA=json.load(open('/tmp/hybridsel.json'));B=json.load(open('submission/selection.json'))\\nn=min(len(A),len(B)); print('equal positions',sum(a==b for a,b in zip(A,B)),n)\\nfor k in [10,100,1000,5000,10000,15000,20000]: print(k,sum(A[i]==B[i] for i in range(k))/k, len(set(A[:k])^set(B[:k])))\\nPY\",\"workdir\":\"/workspace\",\"yield_time_ms\":10000,\"max_output_tokens\":3000}); text(r.output)\n"
}

{
  "input": "const patch = \"*** Begin Patch\\n*** Update File: /workspace/submission/curate.py\\n@@\\n     order=[];idx={k:0 for k in chosen};used={k:0 for k in chosen};active=set(chosen);seen=set()\\n+    tie_priority={'web':0,'news':1,'technical':2,'encyclopedic':3}\\n     while sum(used.values())<16_000_000 and active:\\n-      dom=min(active,key=lambda k:(used[k],k))\\n+      dom=min(active,key=lambda k:(used[k],tie_priority[k]))\\n*** End Patch\";\ntext(await tools.apply_patch(patch));\n"
}