File size: 5,017 Bytes
9e9bfb7 | 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 | <svg xmlns="http://www.w3.org/2000/svg" width="1200" height="520" viewBox="0 0 1200 520" role="img" aria-labelledby="title desc">
<title id="title">PreFailureNet UI-PreFail-v2 release visual</title>
<desc id="desc">A compact technical diagram contrasting the v1 mid-frame shortcut flaw with the v2 ordered-event counterfactual benchmark where matched pairs share static and unordered evidence but differ in event order.</desc>
<defs>
<style>
.bg { fill: #f8fafc; }
.panel { fill: #ffffff; stroke: #cbd5e1; stroke-width: 1.5; }
.title { fill: #0f172a; font: 700 30px Arial, sans-serif; }
.subtitle { fill: #475569; font: 16px Arial, sans-serif; }
.h { fill: #0f172a; font: 700 18px Arial, sans-serif; }
.txt { fill: #334155; font: 14px Arial, sans-serif; }
.small { fill: #64748b; font: 12px Arial, sans-serif; }
.metric { fill: #0f172a; font: 700 22px Arial, sans-serif; }
.frame { fill: #f1f5f9; stroke: #94a3b8; stroke-width: 1.3; }
.same { fill: #ecfdf5; stroke: #34d399; stroke-width: 1.3; }
.weak { fill: #fff7ed; stroke: #fb923c; stroke-width: 1.3; }
.bad { fill: #fef2f2; stroke: #f87171; stroke-width: 1.3; }
.good { fill: #eff6ff; stroke: #60a5fa; stroke-width: 1.3; }
.arrow { stroke: #64748b; stroke-width: 2; fill: none; marker-end: url(#arrow); }
.eventA { fill: #2563eb; }
.eventB { fill: #f97316; }
</style>
<marker id="arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="7" markerHeight="7" orient="auto-start-reverse">
<path d="M 0 0 L 10 5 L 0 10 z" fill="#64748b"/>
</marker>
</defs>
<rect class="bg" width="1200" height="520"/>
<text x="42" y="58" class="title">PreFailureNet / UI-PreFail-v2</text>
<text x="42" y="86" class="subtitle">Synthetic ordered temporal counterfactual benchmark, not a production browser-agent guardrail</text>
<rect x="40" y="120" width="350" height="330" rx="8" class="panel"/>
<text x="68" y="155" class="h">v1 flaw</text>
<text x="68" y="184" class="txt">Matched final evidence, but the label leaked through</text>
<text x="68" y="204" class="txt">localized mid-sequence visual patches.</text>
<rect x="78" y="234" width="62" height="54" rx="5" class="frame"/>
<rect x="155" y="234" width="62" height="54" rx="5" class="bad"/>
<rect x="232" y="234" width="62" height="54" rx="5" class="frame"/>
<circle cx="186" cy="260" r="12" fill="#ef4444"/>
<path d="M 143 261 L 151 261" class="arrow"/>
<path d="M 220 261 L 228 261" class="arrow"/>
<text x="84" y="312" class="small">t0</text>
<text x="166" y="312" class="small">mid-frame shortcut</text>
<text x="238" y="312" class="small">final</text>
<text x="68" y="363" class="metric">LR AP 0.9249</text>
<text x="68" y="391" class="small">Negative result: v1 should not support a temporal-reasoning claim.</text>
<rect x="425" y="120" width="390" height="330" rx="8" class="panel"/>
<text x="453" y="155" class="h">v2 ordered-event counterfactual</text>
<text x="453" y="184" class="txt">Safe and unsafe pairs share state, context, actions,</text>
<text x="453" y="204" class="txt">final frame, and unordered frame bag.</text>
<text x="458" y="246" class="small">safe order</text>
<rect x="538" y="224" width="52" height="44" rx="5" class="good"/>
<circle cx="564" cy="246" r="9" class="eventA"/>
<path d="M 596 246 L 626 246" class="arrow"/>
<rect x="638" y="224" width="52" height="44" rx="5" class="good"/>
<circle cx="664" cy="246" r="9" class="eventB"/>
<text x="702" y="250" class="small">label: stable</text>
<text x="458" y="318" class="small">unsafe order</text>
<rect x="538" y="296" width="52" height="44" rx="5" class="weak"/>
<circle cx="564" cy="318" r="9" class="eventB"/>
<path d="M 596 318 L 626 318" class="arrow"/>
<rect x="638" y="296" width="52" height="44" rx="5" class="weak"/>
<circle cx="664" cy="318" r="9" class="eventA"/>
<text x="702" y="322" class="small">label: failure</text>
<rect x="455" y="376" width="320" height="42" rx="6" class="same"/>
<text x="474" y="402" class="txt">Only event order changes inside the matched pair.</text>
<rect x="850" y="120" width="310" height="330" rx="8" class="panel"/>
<text x="878" y="155" class="h">Held-out v2 result</text>
<text x="878" y="198" class="small">Final-frame LR AP</text>
<text x="1034" y="198" class="metric">0.5005</text>
<text x="878" y="238" class="small">Unordered bag LR AP</text>
<text x="1034" y="238" class="metric">0.5048</text>
<text x="878" y="278" class="small">Frame-delta LR AP</text>
<text x="1034" y="278" class="metric">0.8525</text>
<text x="878" y="318" class="small">Transformer AP</text>
<text x="1034" y="318" class="metric">0.9926</text>
<text x="878" y="358" class="small">Transformer pair-rank</text>
<text x="1034" y="358" class="metric">1.0000</text>
<text x="878" y="404" class="small">Claim boundary: ordered temporal evidence under synthetic controls.</text>
</svg>
|