/* Results tab: every number below is taken from the SchemeArena paper (arXiv:2609.08126). */
(function () {
"use strict";
var MODELS = [
{ key: "q8", name: "Qwen3-8B", color: "var(--m-q8)", hex: "#a78bdb" },
{ key: "q32", name: "Qwen3-32B", color: "var(--m-q32)", hex: "#5b3fb8" },
{ key: "o4", name: "o4-mini", color: "var(--m-o4)", hex: "#23a585" },
{ key: "o1", name: "o1", color: "var(--m-o1)", hex: "#0c6a52" },
{ key: "claude", name: "Claude-3.7-Sonnet", color: "var(--m-claude)", hex: "#d2714f" }
];
// Figure 4: scheming propensity (%), Setups 1-4.
var PROP_SETUPS = [
{ name: "Setup 1", desc: "All factors off", hex: "#d9d5ce" },
{ name: "Setup 2", desc: "Pressure only", hex: "#ebbf98" },
{ name: "Setup 3", desc: "Goal only", hex: "#d9783d" },
{ name: "Setup 4", desc: "Goal + pressure", hex: "#97320f" }
];
var PROP = {
q8: [31.9, 37.2, 81.2, 84.8],
q32: [29.6, 38.1, 79.0, 82.9],
o4: [31.1, 38.3, 75.4, 76.7],
o1: [29.1, 34.5, 74.4, 78.8],
claude: [21.7, 23.7, 75.6, 76.8]
};
// Section 5.3 text: capability (%). null = not reported numerically in the text.
var CAP = {
q8: [99.47, null, null],
q32: [98.94, null, null],
o4: [77.04, 81.89, 96.43],
o1: [70.85, 76.46, 97.98],
claude: [63.12, 68.57, 68.68]
};
// Figure 6: % trajectories with scheming reasoning but no scheming action.
var GAP_COLS = ["Setup 1
All off", "Setup 2
Pressure", "Setup 3
Goal", "Setup 4
Goal + Pres.", "Setup 5
Hint", "Setup 6
w/ Action", "Setup 6
w/ All", "Setup 7
Detailed hint"];
var GAP = {
q8: [55.3, 53.9, 48.0, 44.0, 5.0, 6.0, 5.7, 0.3],
q32: [42.4, 52.9, 39.2, 35.1, 4.3, 4.7, 4.8, 0.5],
o4: [47.2, 57.9, 49.2, 50.3, 50.2, 46.9, 49.6, 13.2],
o1: [62.4, 59.4, 50.7, 53.9, 62.9, 49.0, 57.7, 18.0],
claude: [44.2, 64.3, 50.4, 50.3, 34.3, 32.3, 44.2, 19.6]
};
var tip;
function showTip(e, html) { tip.innerHTML = html; tip.style.left = e.clientX + "px"; tip.style.top = e.clientY + "px"; tip.style.opacity = 1; }
function hideTip() { tip.style.opacity = 0; }
function propChart() {
var el = document.getElementById("propChart");
var W = 960, H = 380, m = { t: 26, r: 8, b: 44, l: 44 };
var iw = W - m.l - m.r, ih = H - m.t - m.b;
var gw = iw / MODELS.length, pad = gw * 0.16, bw = (gw - 2 * pad) / 4;
var y = function (v) { return m.t + ih - (v / 100) * ih; };
var s = '";
el.innerHTML = s;
el.querySelectorAll(".bar").forEach(function (b) {
b.addEventListener("mousemove", function (e) { showTip(e, b.getAttribute("data-tip")); });
b.addEventListener("mouseleave", hideTip);
});
document.getElementById("propLegend").innerHTML = PROP_SETUPS.map(function (p) {
return '' + p.name + ": " + p.desc + "";
}).join("");
}
function dot(md) { return '' + md.name; }
function propTable() {
var h = "