KridgeDookie's picture
Remove invented ranking language and relabel internal diagnostics
8e0ed7f verified
|
Raw
History Blame
6.59 kB
<svg xmlns="http://www.w3.org/2000/svg" width="1200" height="590" viewBox="0 0 1200 590" role="img" aria-labelledby="title desc">
<title id="title">How J-lens was used</title>
<desc id="desc">A frozen Jacobian lens reads six layers for three prompt cohorts, compares refusal and capability concepts against a private internal baseline, and is used only as a diagnostic.</desc>
<defs>
<linearGradient id="bg" x1="0" y1="0" x2="1" y2="1">
<stop offset="0" stop-color="#080e1e"/>
<stop offset="1" stop-color="#15162e"/>
</linearGradient>
<linearGradient id="beam" x1="0" y1="0" x2="1" y2="0">
<stop offset="0" stop-color="#35d7ff"/>
<stop offset="1" stop-color="#9d78ff"/>
</linearGradient>
<filter id="soft" x="-40%" y="-40%" width="180%" height="180%">
<feGaussianBlur stdDeviation="6"/>
</filter>
<marker id="arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto">
<path d="M0 0L8 4L0 8Z" fill="#6c86b8"/>
</marker>
</defs>
<rect width="1200" height="590" rx="24" fill="url(#bg)"/>
<text x="52" y="58" fill="#f5f7ff" font-family="Inter,Segoe UI,Arial,sans-serif" font-size="29" font-weight="500">J-lens: frozen interpretability audit</text>
<rect x="940" y="29" width="208" height="38" rx="19" fill="#f5c451" fill-opacity=".12" stroke="#f5c451" stroke-opacity=".65"/>
<text x="1044" y="53" text-anchor="middle" fill="#ffd875" font-family="Inter,Segoe UI,Arial,sans-serif" font-size="13" font-weight="500" letter-spacing="2">AUDIT ONLY · NO TUNING</text>
<g font-family="Inter,Segoe UI,Arial,sans-serif">
<text x="52" y="105" fill="#7f92bc" font-size="13" letter-spacing="2">1 · FROZEN COHORTS</text>
<g transform="translate(52 126)">
<rect width="238" height="76" rx="14" fill="#35d7ff" fill-opacity=".08" stroke="#35d7ff" stroke-opacity=".35"/>
<text x="20" y="32" fill="#eaf5ff" font-size="16" font-weight="500">Hard prompts</text>
<text x="20" y="56" fill="#91a5cd" font-size="13">32 fixed examples</text>
</g>
<g transform="translate(52 216)">
<rect width="238" height="76" rx="14" fill="#8e7dff" fill-opacity=".08" stroke="#8e7dff" stroke-opacity=".35"/>
<text x="20" y="32" fill="#f0ecff" font-size="16" font-weight="500">Capability</text>
<text x="20" y="56" fill="#91a5cd" font-size="13">24 fixed examples</text>
</g>
<g transform="translate(52 306)">
<rect width="238" height="76" rx="14" fill="#56e39f" fill-opacity=".08" stroke="#56e39f" stroke-opacity=".35"/>
<text x="20" y="32" fill="#ecfff6" font-size="16" font-weight="500">Benign boundary</text>
<text x="20" y="56" fill="#91a5cd" font-size="13">24 fixed examples</text>
</g>
<path d="M306 164H372M306 254H372M306 344H372" stroke="#6c86b8" stroke-width="2" marker-end="url(#arrow)"/>
<text x="380" y="105" fill="#7f92bc" font-size="13" letter-spacing="2">2 · SIX HIDDEN-STATE READOUTS</text>
<g transform="translate(382 128)">
<rect width="320" height="252" rx="20" fill="#ffffff" fill-opacity=".035" stroke="#a9c7ff" stroke-opacity=".14"/>
<path d="M48 196C88 48 232 48 272 196" fill="none" stroke="#35d7ff" stroke-opacity=".18" stroke-width="28" filter="url(#soft)"/>
<path d="M48 196C88 48 232 48 272 196" fill="none" stroke="url(#beam)" stroke-width="4"/>
<g fill="#0f1730" stroke-width="3">
<circle cx="48" cy="196" r="18" stroke="#35d7ff"/>
<circle cx="84" cy="125" r="18" stroke="#4ac3ff"/>
<circle cx="130" cy="83" r="18" stroke="#659fff"/>
<circle cx="190" cy="83" r="18" stroke="#7e88ff"/>
<circle cx="236" cy="125" r="18" stroke="#8d7dff"/>
<circle cx="272" cy="196" r="18" stroke="#9d78ff"/>
</g>
<g fill="#f5f7ff" font-size="12" text-anchor="middle">
<text x="48" y="200">20</text>
<text x="84" y="129">24</text>
<text x="130" y="87">28</text>
<text x="190" y="87">32</text>
<text x="236" y="129">36</text>
<text x="272" y="200">40</text>
</g>
<text x="160" y="232" text-anchor="middle" fill="#91a5cd" font-size="13">last prompt position · frozen lens</text>
</g>
<path d="M718 254H782" stroke="#6c86b8" stroke-width="2" marker-end="url(#arrow)"/>
<text x="790" y="105" fill="#7f92bc" font-size="13" letter-spacing="2">3 · CONCEPT-GROUP COMPARISON</text>
<g transform="translate(790 128)">
<rect width="358" height="252" rx="20" fill="#ffffff" fill-opacity=".035" stroke="#a9c7ff" stroke-opacity=".14"/>
<g transform="translate(24 24)">
<rect width="310" height="78" rx="14" fill="#ff6f91" fill-opacity=".08" stroke="#ff6f91" stroke-opacity=".3"/>
<text x="18" y="30" fill="#ffe8ef" font-size="16" font-weight="500">Refusal concepts</text>
<text x="18" y="55" fill="#91a5cd" font-size="13">candidate − private baseline means</text>
</g>
<g transform="translate(24 118)">
<rect width="310" height="108" rx="14" fill="#56e39f" fill-opacity=".07" stroke="#56e39f" stroke-opacity=".3"/>
<text x="18" y="30" fill="#eafff3" font-size="16" font-weight="500">Capability concepts</text>
<text x="18" y="55" fill="#91a5cd" font-size="13">code · JSON · reasoning</text>
<text x="18" y="80" fill="#91a5cd" font-size="13">mixed, modest shifts across cohorts</text>
</g>
</g>
</g>
<g transform="translate(52 423)" font-family="Inter,Segoe UI,Arial,sans-serif">
<rect width="1096" height="118" rx="18" fill="#ffffff" fill-opacity=".03" stroke="#a9c7ff" stroke-opacity=".12"/>
<text x="24" y="34" fill="#f5f7ff" font-size="16" font-weight="500">Candidate minus private internal baseline · mean score deltas</text>
<g transform="translate(24 58)" font-size="13">
<text x="0" y="0" fill="#91a5cd">Hard 32</text>
<text x="0" y="25" fill="#cbd5ec">cap −0.3357 · refusal +0.1342</text>
<text x="345" y="0" fill="#91a5cd">Capability 24</text>
<text x="345" y="25" fill="#cbd5ec">cap +0.5056 · refusal −0.1370</text>
<text x="705" y="0" fill="#91a5cd">Boundary 24</text>
<text x="705" y="25" fill="#cbd5ec">cap −0.4630 · refusal −0.1972</text>
</g>
<text x="1072" y="99" text-anchor="end" fill="#f5c451" font-size="12" letter-spacing="1">BEHAVIORAL GATES DECIDE PROMOTION</text>
</g>
<text x="52" y="572" fill="#63769d" font-family="Inter,Segoe UI,Arial,sans-serif" font-size="11">Reference implementation: Anthropic Jacobian lens · this audit is not an Anthropic endorsement</text>
</svg>