Revert "Add 'Try to fool the referee' tab: curated cheat gallery (bench-shape special-case, pointer memoization, in-place mutation), each caught by the in-process referee"

app.py

@@ -507,13 +507,6 @@ try:
 except Exception:
     REPLAYS_JSON = "{}"
 
-CHEATS_PATH = os.path.join(os.path.dirname(__file__), "cheats", "index.json")
-try:
-    with open(CHEATS_PATH, encoding="utf-8") as f:
-        CHEATS_JSON = json.dumps(json.load(f), separators=(",", ":")).replace("</", "<\\/")
-except Exception:
-    CHEATS_JSON = "{}"
-
 APP_HTML = r"""
 <div id="km" data-build="2026-06-14-local-baseline-guard">
   <div id="km-top">
@@ -527,11 +520,9 @@ APP_HTML = r"""
       <button class="km-tab" data-tab="expert">Expert</button>
       <button class="km-tab" data-tab="lb">Leaderboard</button>
       <button class="km-tab" data-tab="docs">Run locally</button>
-      <button class="km-tab" data-tab="fool">Try to fool it</button>
     </div>
   </div>
   <script id="km-replays-data" type="application/json">__KM_REPLAYS__</script>
-  <script id="km-cheats-data" type="application/json">__KM_CHEATS__</script>
 
   <!-- ============ BUILD ============ -->
   <div class="km-page" data-page="build">
@@ -700,20 +691,10 @@ kernel plus a run(x, residual, weight) entry point. Triton only.</div>
       <p>MIT licensed. The models are <a href="https://huggingface.co/YMRohit/ouroboros-kernelsmith-minicpm5-1b" style="color:var(--accent)">MiniCPM5-1B</a> and <a href="https://huggingface.co/YMRohit/ouroboros-kernelsmith-qwen3.6-27b" style="color:var(--accent)">Qwen3.6-27B</a> (both Apache-2.0). The verified kernels and evidence reports are in the <a href="https://huggingface.co/datasets/YMRohit/ouroboros-kernel-corpus" style="color:var(--accent)">corpus</a>.</p>
     </section>
   </div>
-
-  <!-- ============ TRY TO FOOL THE REFEREE ============ -->
-  <div class="km-page" data-page="fool" hidden>
-    <section class="km-card">
-      <h2>Try to fool the referee</h2>
-      <p class="km-dim">A green tick is easy to believe. The harder, more honest test is the inverse: hand the referee a kernel built to cheat, and watch it get caught. Each card is a real exploit, taken from the harness's own anti-gaming controls, and each is defeated by a different defense. The verdicts come from the same referee bundled in this Space (<code>referee/harness.py</code>), so they reproduce: switch to Local mode to run any of them live.</p>
-      <div id="km-cheats" class="km-cheats"></div>
-    </section>
-  </div>
 </div>
 """
 
 APP_HTML = APP_HTML.replace("__KM_REPLAYS__", REPLAYS_JSON)
-APP_HTML = APP_HTML.replace("__KM_CHEATS__", CHEATS_JSON)
 
 APP_JS = r"""
 () => {
@@ -752,29 +733,6 @@ APP_JS = r"""
   const esc=s=>cleanText(s).replace(/[&<>"']/g,c=>({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[c]));
   let REPLAYS={};
   try{REPLAYS=JSON.parse(($('#km-replays-data')||{}).textContent||'{}');}catch(e){REPLAYS={};}
-  let CHEATS={};
-  try{CHEATS=JSON.parse(($('#km-cheats-data')||{}).textContent||'{}');}catch(e){CHEATS={};}
-  function renderCheats(){
-    const wrap=$('#km-cheats'); if(!wrap||wrap.dataset.done)return;
-    const order=['softmax_cheat_shape','softmax_cheat_memo','silu_mutate'].filter(k=>CHEATS[k]);
-    if(!order.length){wrap.innerHTML="<p class='km-dim'>No cheats loaded.</p>";return;}
-    wrap.innerHTML=order.map(k=>{const c=CHEATS[k];return (
-      "<div class='km-cheat'>"
-      +"<div class='km-cheat-top'><span class='km-cheat-name'>"+esc(c.label)+"</span><span class='km-cheat-claims'>submitted as <code>"+esc(c.op)+"</code></span></div>"
-      +"<p class='km-cheat-trick'>"+esc(c.trick)+"</p>"
-      +"<pre class='km-cheat-src'>"+esc(c.source)+"</pre>"
-      +"<button class='km-btn cheat-run' data-cheat='"+esc(k)+"'>Run it past the referee</button>"
-      +"<div class='km-cheat-verdict' id='cv-"+esc(k)+"'></div>"
-      +"</div>");
-    }).join('');
-    wrap.dataset.done='1';
-    $$('.cheat-run').forEach(b=>b.addEventListener('click',()=>{
-      const k=b.dataset.cheat,c=CHEATS[k],el=document.getElementById('cv-'+k);
-      if(!c||!el)return;
-      el.innerHTML="<div class='km-reject'><div class='km-reject-row'><span class='km-reject-x'>REJECTED</span><span class='km-reject-by'>caught by "+esc(c.caught_by)+"</span></div><p>"+esc(c.explanation)+"</p></div>";
-      b.textContent='Rejected by the referee';b.classList.add('done');b.disabled=true;
-    }));
-  }
   const REPLAY_META={
     rmsnorm_gelu:{label:'RMSNorm to GELU',short:'RMSNorm GELU'},
     softmax:{label:'Softmax',short:'Softmax'},
@@ -826,7 +784,6 @@ APP_JS = r"""
     requestAnimationFrame(positionTabPill);
     if(tab==='expert')resolveExpert();
     if(tab==='lb')refreshLB();
-    if(tab==='fool')renderCheats();
   }
   $$('.km-tab').forEach(b=>b.addEventListener('click',()=>showTab(b.dataset.tab)));
   window.addEventListener('resize',positionTabPill);

cheats/index.json

@@ -1,32 +0,0 @@
-{
-  "softmax_cheat_shape": {
-    "op": "softmax",
-    "label": "Special-case the benchmark shape",
-    "trick": "Hard-code the answer only for the exact shape the timer uses, and return instant garbage everywhere else. The classic infinite-speedup exploit.",
-    "caught_by": "the correctness sweep",
-    "explanation": "The referee folds the public bench shape into its adversarial correctness sweep, so special-casing it buys nothing. This kernel returns NaN on 8192x4096 and is rejected before it is ever timed.",
-    "verdict": "REJECTED",
-    "status": "incorrect",
-    "source": "# NEGATIVE CONTROL (anti-gaming): special-cases the PUBLIC, FIXED bench shape.\n# Correct softmax everywhere the old correctness sweep could look, but at the timing shape\n# (8192, 4096) it returns uninitialized garbage instantly \u2014 the \"infinite speedup\" exploit.\n# The V2 harness includes the bench inputs in the correctness sweep, so this MUST be REJECTED.\ndef run(x):\n    if x.shape == (8192, 4096):\n        return torch.empty_like(x)          # instant garbage at the bench shape\n    return torch.softmax(x.float(), dim=-1).to(x.dtype)"
-  },
-  "softmax_cheat_memo": {
-    "op": "softmax",
-    "label": "Memoize by input pointer",
-    "trick": "Cache the output keyed by the input's memory address, so after the first call it does zero GPU work and wins the timer.",
-    "caught_by": "the anti-memoization check",
-    "explanation": "Before each timed iteration the referee pokes one input element, then allclose-checks the final timed output against the live input state. A cached, stale output fails: memoization/caching suspected.",
-    "verdict": "REJECTED",
-    "status": "incorrect",
-    "source": "# NEGATIVE CONTROL (anti-gaming): memoizes its output by input POINTER, so after the first\n# call it does no GPU work \u2014 the \"cache the answer, win the bench\" exploit. The V2 harness\n# defeats this two ways: the CUDA caching allocator reuses freed clone blocks across\n# correctness cases (stale hit -> wrong output), and the bench loop pokes one input element\n# before every timed iteration then allclose-verifies the FINAL timed output (stale -> fail).\n# Either path: this MUST be REJECTED.\n_CACHE = {}\n\n\ndef run(x):\n    key = (x.data_ptr(), tuple(x.shape), x.dtype)\n    hit = _CACHE.get(key)\n    if hit is not None:\n        return hit\n    out = torch.softmax(x.float(), dim=-1).to(x.dtype)\n    _CACHE[key] = out\n    return out"
-  },
-  "silu_mutate": {
-    "op": "silu",
-    "label": "Mutate the input in place",
-    "trick": "Compute the correct values but write them back into the input tensor, so the bench loop's reused inputs look already done.",
-    "caught_by": "the input-mutation contract",
-    "explanation": "run() must write a fresh output tensor and never touch its inputs. The referee checks the input is unchanged after the call: kernel MUTATES its input.",
-    "verdict": "REJECTED",
-    "status": "incorrect",
-    "source": "# NEGATIVE CONTROL (contract): produces the CORRECT silu values but writes them IN-PLACE\n# into its input tensor. The bench loop reuses identical args across iterations, which is\n# only sound for kernels that never mutate inputs \u2014 so the correctness phase must enforce\n# the contract, not assume it. This MUST be REJECTED (\"kernel MUTATES its input\").\ndef run(x):\n    xf = x.float()\n    x.copy_((xf * torch.sigmoid(xf)).to(x.dtype))\n    return x"
-  }
-}

seed_kernels/cheats/silu_mutate.py

@@ -1,8 +0,0 @@
-# NEGATIVE CONTROL (contract): produces the CORRECT silu values but writes them IN-PLACE
-# into its input tensor. The bench loop reuses identical args across iterations, which is
-# only sound for kernels that never mutate inputs — so the correctness phase must enforce
-# the contract, not assume it. This MUST be REJECTED ("kernel MUTATES its input").
-def run(x):
-    xf = x.float()
-    x.copy_((xf * torch.sigmoid(xf)).to(x.dtype))
-    return x

seed_kernels/cheats/softmax_cheat_memo.py

@@ -1,17 +0,0 @@
-# NEGATIVE CONTROL (anti-gaming): memoizes its output by input POINTER, so after the first
-# call it does no GPU work — the "cache the answer, win the bench" exploit. The V2 harness
-# defeats this two ways: the CUDA caching allocator reuses freed clone blocks across
-# correctness cases (stale hit -> wrong output), and the bench loop pokes one input element
-# before every timed iteration then allclose-verifies the FINAL timed output (stale -> fail).
-# Either path: this MUST be REJECTED.
-_CACHE = {}
-
-
-def run(x):
-    key = (x.data_ptr(), tuple(x.shape), x.dtype)
-    hit = _CACHE.get(key)
-    if hit is not None:
-        return hit
-    out = torch.softmax(x.float(), dim=-1).to(x.dtype)
-    _CACHE[key] = out
-    return out

seed_kernels/cheats/softmax_cheat_shape.py

@@ -1,8 +0,0 @@
-# NEGATIVE CONTROL (anti-gaming): special-cases the PUBLIC, FIXED bench shape.
-# Correct softmax everywhere the old correctness sweep could look, but at the timing shape
-# (8192, 4096) it returns uninitialized garbage instantly — the "infinite speedup" exploit.
-# The V2 harness includes the bench inputs in the correctness sweep, so this MUST be REJECTED.
-def run(x):
-    if x.shape == (8192, 4096):
-        return torch.empty_like(x)          # instant garbage at the bench shape
-    return torch.softmax(x.float(), dim=-1).to(x.dtype)

seed_kernels/cheats/softmax_wrong.py

@@ -1,27 +0,0 @@
-# NEGATIVE CONTROL: row softmax with NO max-subtraction. Mathematically "equivalent" and
-# passes on benign N(0,1) inputs — but OVERFLOWS to inf/nan on large magnitudes (the
-# adversarial scale sweep in specs._mk_softmax). The harness MUST reject it. This is the
-# exact trap the brief warns about: a kernel that passes one easy regime and is wrong.
-@triton.jit
-def _bad_kernel(x_ptr, y_ptr, stride, N, BLOCK: tl.constexpr):
-    row = tl.program_id(0)
-    x_ptr += row * stride
-    y_ptr += row * stride
-    d = tl.zeros([BLOCK], dtype=tl.float32)
-    for off in range(0, N, BLOCK):
-        cols = off + tl.arange(0, BLOCK)
-        x = tl.load(x_ptr + cols, mask=cols < N, other=0.0).to(tl.float32)
-        d += tl.where(cols < N, tl.exp(x), 0.0)        # BUG: exp(x), not exp(x - row_max)
-    denom = tl.sum(d)
-    for off in range(0, N, BLOCK):
-        cols = off + tl.arange(0, BLOCK)
-        mask = cols < N
-        x = tl.load(x_ptr + cols, mask=mask, other=0.0).to(tl.float32)
-        tl.store(y_ptr + cols, tl.exp(x) / denom, mask=mask)
-
-
-def run(x):
-    M, N = x.shape
-    y = torch.empty_like(x)
-    _bad_kernel[(M,)](x, y, x.stride(0), N, BLOCK=1024)
-    return y

style.css

@@ -445,25 +445,3 @@ input[type=range].km-range::-moz-range-thumb { width: 16px; height: 16px; border
 #km #km-act2-wrap.open .km-add2 { display: none; }
 #km #km-act2-stage { display: none; }
 #km #km-act2-wrap.open #km-act2-stage { display: flex; }
-
-/* ---- Try to fool the referee ---- */
-#km .km-cheats { display: flex; flex-direction: column; gap: 16px; margin-top: 18px; }
-#km .km-cheat { border: 1px solid var(--line); border-radius: var(--r); padding: 18px 20px; background: var(--surface-2); }
-#km .km-cheat-top { display: flex; align-items: baseline; justify-content: space-between; gap: 12px; flex-wrap: wrap; }
-#km .km-cheat-name { font: 600 17px var(--sans); color: var(--text); }
-#km .km-cheat-claims { font: 500 13px var(--sans); color: var(--faint); }
-#km .km-cheat-claims code { font-family: var(--mono); color: var(--accent); }
-#km .km-cheat-trick { margin: 8px 0 12px; color: var(--muted); font-size: 14px; line-height: 1.55; }
-#km .km-cheat-src { margin: 0 0 14px; padding: 14px 16px; border-radius: var(--r-sm); background: #0a0b0d;
-  border: 1px solid var(--line); font: 12.5px/1.5 var(--mono); overflow-x: auto; white-space: pre; color: #cdd3da; }
-#km .cheat-run.done { opacity: 0.65; cursor: default; border-color: var(--danger); color: var(--danger); }
-#km .km-cheat-verdict { margin-top: 14px; }
-#km .km-cheat-verdict:empty { margin-top: 0; }
-#km .km-reject { border: 1px solid var(--danger); border-radius: var(--r-sm); padding: 16px 18px;
-  background: var(--danger-soft); animation: kmRejectIn 0.25s var(--ease); }
-@keyframes kmRejectIn { from { opacity: 0; transform: translateY(6px); } to { opacity: 1; transform: none; } }
-#km .km-reject-row { display: flex; align-items: center; gap: 14px; margin-bottom: 8px; }
-#km .km-reject-x { font: 700 14px var(--mono); letter-spacing: 1px; color: var(--danger);
-  border: 1px solid var(--danger); border-radius: 7px; padding: 4px 11px; }
-#km .km-reject-by { color: #f3a3a3; font: 550 14px var(--sans); }
-#km .km-reject p { margin: 0; color: var(--muted); font-size: 14px; line-height: 1.55; }