File size: 6,508 Bytes
25f3b2e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
<!doctype html>
<html lang="en">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>TimeRCD Reproduction Poster</title>
<style>
  :root {
    --accent: #5f2878;
    --accent-deep: #321444;
    --accent-soft: #eee4f3;
    --emph: #2d5f3e;
    --emph-soft: #e4f0e8;
    --ink: #172033;
    --muted: #5e687a;
    --line: #d8deea;
    --paper: #fbfbfd;
    --card: #ffffff;
  }
  * { box-sizing: border-box; }
  body { margin: 0; background: var(--paper); color: var(--ink); font-family: Inter, ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; }
  .poster { width: min(1200px, 100vw); aspect-ratio: 5 / 3; margin: 0 auto; padding: 28px; display: grid; grid-template-rows: auto 1fr auto; gap: 18px; background: linear-gradient(180deg, #fff 0%, #f7f4fa 100%); border: 1px solid var(--line); }
  header { display: grid; grid-template-columns: 1fr auto; gap: 18px; align-items: end; border-bottom: 4px solid var(--accent); padding-bottom: 12px; }
  h1 { margin: 0; font-size: clamp(28px, 3.1vw, 48px); line-height: 1.02; letter-spacing: 0; color: var(--accent-deep); }
  .subtitle { margin-top: 8px; color: var(--muted); font-size: clamp(13px, 1.35vw, 18px); }
  .badge { background: var(--accent); color: white; padding: 10px 12px; border-radius: 6px; font-weight: 800; text-align: center; min-width: 132px; }
  main { display: grid; grid-template-columns: 1.15fr 1fr 1fr 1fr; gap: 14px; min-height: 0; }
  section { background: var(--card); border: 1px solid var(--line); border-radius: 8px; padding: 16px; display: flex; flex-direction: column; gap: 12px; box-shadow: 0 12px 26px rgba(35, 26, 48, 0.08); position: relative; }
  section[data-logbook-target]::after { content: "Open details ↗"; position: absolute; top: 10px; right: 10px; font-size: 11px; background: var(--accent-soft); color: var(--accent-deep); border: 1px solid #dac8e7; padding: 4px 7px; border-radius: 999px; }
  h2 { margin: 0; color: var(--accent); font-size: clamp(16px, 1.55vw, 22px); line-height: 1.1; }
  p, li { font-size: clamp(12px, 1.1vw, 16px); line-height: 1.34; margin: 0; }
  ul { margin: 0; padding-left: 18px; display: grid; gap: 6px; }
  .statgrid { display: grid; grid-template-columns: repeat(2, 1fr); gap: 8px; }
  .stat { background: var(--accent-soft); border: 1px solid #dac8e7; border-radius: 6px; padding: 10px; min-height: 68px; display: flex; flex-direction: column; justify-content: center; }
  .num { font-size: clamp(21px, 2.2vw, 34px); line-height: 1; color: var(--accent-deep); font-weight: 900; white-space: nowrap; }
  .lab { font-size: 11px; color: var(--muted); margin-top: 4px; }
  .warn { background: #fff3df; border-color: #edd4a7; }
  .ok { background: var(--emph-soft); border-color: #bdd7c5; }
  .mini { width: 100%; border-collapse: collapse; font-size: clamp(11px, 1vw, 14px); }
  .mini th, .mini td { border-top: 1px solid var(--line); padding: 6px 4px; text-align: left; }
  .mini th { color: var(--muted); font-weight: 700; }
  footer { display: flex; justify-content: space-between; gap: 12px; color: var(--muted); font-size: 12px; border-top: 1px solid var(--line); padding-top: 8px; }
  a { color: inherit; text-decoration: none; }
  section[data-logbook-target]:hover { outline: 3px solid rgba(95, 40, 120, 0.22); }
</style>
<div class="poster">
  <header data-logbook-target="executive-summary">
    <div>
      <h1>TimeRCD Zero-Shot TSAD Reproduction</h1>
      <div class="subtitle">Evidence logbook for Relative Context Discrepancy, synthetic pretraining, and reported benchmark claims.</div>
    </div>
    <div class="badge">ICML 2026<br>paper #12052</div>
  </header>
  <main>
    <section data-logbook-target="claim-1-claim-1">
      <h2>Reported rankings changed from the prompt</h2>
      <div class="statgrid">
        <div class="stat warn"><div class="num">38 / 14</div><div class="lab">current paper zero-shot first / second, 64 cases</div></div>
        <div class="stat warn"><div class="num">41 / 6</div><div class="lab">prompt claim, 56 cases</div></div>
        <div class="stat warn"><div class="num">35 / 3</div><div class="lab">current paper full-shot first / second</div></div>
        <div class="stat warn"><div class="num">28</div><div class="lab">prompt full-shot first claim</div></div>
      </div>
      <p>The extracted camera-ready paper text does not match the user-supplied ranking counts, so the ranking claims are partially supported only for the current-paper version.</p>
    </section>
    <section data-logbook-target="claim-3-claim-3">
      <h2>Mechanism verified in released code</h2>
      <ul>
        <li>8-layer, 8-head Transformer-style encoder with RoPE.</li>
        <li>Separate reconstruction and anomaly heads are present.</li>
        <li>Zero-shot wrapper returns anomaly logits, not reconstruction error alone.</li>
      </ul>
      <table class="mini">
        <tr><th>Head</th><th>Loss</th></tr>
        <tr><td>reconstruction</td><td>MSE</td></tr>
        <tr><td>anomaly</td><td>cross entropy</td></tr>
      </table>
    </section>
    <section data-logbook-target="claim-4-claim-4">
      <h2>Generator stages present</h2>
      <ul>
        <li>Univariate templates with trend, seasonality, noise, and local changes.</li>
        <li>Multivariate DAG sampling with ARX-style parent dynamics.</li>
        <li>Endogenous/exogenous anomaly paths and token-level labels.</li>
      </ul>
      <div class="stat ok"><div class="num">15,360</div><div class="lab">local generated public-code points; T4 job generated 34,816</div></div>
    </section>
    <section data-logbook-target="claim-6-claim-6">
      <h2>Reduced cloud cross-check</h2>
      <p>A T4 Hugging Face Job cloned the public repos, generated official synthetic samples, and ran a transparent contextual-anomaly proxy.</p>
      <div class="statgrid">
        <div class="stat ok"><div class="num">65.8k/s</div><div class="lab">T4 generator throughput</div></div>
        <div class="stat ok"><div class="num">0.134</div><div class="lab">toy RCD proxy F1</div></div>
      </div>
      <p>The proxy favors relative context over one-step reconstruction, but it is not a substitute for the released checkpoint benchmark.</p>
    </section>
  </main>
  <footer>
    <span>Assets: thu-sail-lab/Time-RCD, thu-sail-lab/TSAD_dataset_gen_public, thu-sail-lab/Time_RCD Space</span>
    <span>Bundle: scripts, logs, generated CSV/JSON/HTML, paper PDF/text</span>
  </footer>
</div>
</html>