{"container_html":"<div class=\"mercury\">\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"1\" class=\"lead\"><h1>BXRL: Behavior-Explainable Reinforcement Learning</h1>\n<p>Ram Rachum, &nbsp; Yotam Amitai,&nbsp; Yonatan Nakar,<br>Reuth Mirsky, &nbsp; Cameron Allen</p>\n<div class=\"lead-footer-right\">\n  <img class=\"dark-include\" src=\"/bxrl-10m/images/gold-lab-logo.png?v=1786868460\">\n  </img>\n</div>\n<aside class=\"saturn-venus-note\" data-index=\"0\" class=\"saturn-venus-note\" data-index=\"0\">\n  Bittersweet giving this talk, the center of it is the failure of Experience Breakdown\n  Post-mortem, suspense\n</aside><div class=\"blouse-slide-number\">1</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"2\" class=\"overview overview-1\">\n<h1>We’ll talk about:</h1>\n<ul>\n  <li><i>Experience Breakdown</i>: The method that didn’t work 😩</li>\n  <li>Behavior as a new explanation target</li>\n  <li>Explanation targets in XRL</li>\n  <li>HighJax: Our environment for BXRL experiments</li>\n</ul><div class=\"blouse-slide-number\">2</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"3\"><h2>Multi-Agent Reinforcement Learning</h2>\n<div style=\"position:absolute; top:105px; left:0; right:0; bottom:70px; display:flex; align-items:center; justify-content:center\">\n  <img src=\"/bxrl-10m/images/gridworld.png?v=1786868460\" style=\"max-height:297px; max-width:54%; object-fit:contain\">\n  </img>\n</div>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:30px; right:30px; font-size:11pt; color:#ababab; line-height:1.35\">\n  Carvalho, D. S. &amp; Sengupta, B. (2022). <i>Hierarchically Structured Scheduling and Execution of Tasks in a Multi-Agent Environment</i>. arXiv:2203.03021.\n</div><div class=\"blouse-slide-number\">3</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"4\">\n<h2>Experience Breakdown: Strategy</h2>\n<aside class=\"saturn-venus-note\" data-index=\"3\" class=\"saturn-venus-note\" data-index=\"3\">\n  <p>I’ll tell you about the explainability method that I tried to develop and it didn’t work. This is kind of a tangent to this talk, and I’m careful not to dive into it, so I’ll just do a couple of slides about it and if anyone wants to know more, talk to me after the meeting.</p>\n  <p>Background on MARL research. I developed EB in ignorance, trying to think from first principles. Spinning it off to its own paper, wanted to know if anyone’s done that before.</p>\n  <p>Why is $m(\\pi)$ high: give example of overtaking.</p>\n</aside>\n<div class=\"fc\" style=\"position:relative;height:460px\">\n  <svg style=\"position:absolute;width:0;height:0\">\n    <defs>\n      <marker id=\"fc-ah\" markerWidth=\"10\" markerHeight=\"8\" refX=\"10\" refY=\"4\" orient=\"auto\">\n        <polygon points=\"0,0 10,4 0,8\" fill=\"#999\"></polygon>\n      </marker>\n    </defs>\n  </svg>\n  <div class=\"bx question\" style=\"left:10px;top:0;width:340px\" data-anim=\"show@1\">\n    <b>Q:</b>\n     Why did the agent choose action $a$ in state $s$?\n  </div>\n  <div class=\"bx answer\" style=\"left:430px;top:93px;width:430px\" data-anim=\"show@2\">\n    <b>A:</b>\n     In training, agent reached $s' \\sim s$, chose $a' \\sim a$ and got a high advantage.\n  </div>\n  <div style=\"position:absolute;left:150px;top:75px\" data-anim=\"show@3\">\n    <svg width=\"20\" height=\"134\">\n      <line x1=\"10\" y1=\"0\" x2=\"10\" y2=\"134\" stroke=\"#999\" stroke-width=\"1.5\" marker-end=\"url(#fc-ah)\"></line>\n    </svg>\n  </div>\n  <div class=\"lbl\" style=\"left:175px;top:133px\" data-anim=\"show@3\">Remove sampling</div>\n  <div class=\"bx question\" style=\"left:10px;top:209px;width:270px\" data-anim=\"show@3\">\n    <b>Q:</b>\n     Why is $\\pi(a|s)$ high?\n  </div>\n  <div style=\"position:absolute;left:270px;top:222px;pointer-events:none\" data-anim=\"show@4\">\n    <svg width=\"140\" height=\"120\">\n      <line x1=\"130\" y1=\"114\" x2=\"24\" y2=\"21\" stroke=\"#999\" stroke-width=\"1.5\" marker-end=\"url(#fc-ah)\"></line>\n    </svg>\n  </div>\n  <div style=\"position:absolute;left:590px;top:160px;pointer-events:none\" data-anim=\"show@4\">\n    <svg width=\"100\" height=\"170\">\n      <line x1=\"10\" y1=\"158\" x2=\"83\" y2=\"10\" stroke=\"#999\" stroke-width=\"1.5\" marker-end=\"url(#fc-ah)\"></line>\n    </svg>\n  </div>\n  <div class=\"bx strategy\" style=\"left:380px;top:290px;width:470px\" data-anim=\"show@4\">\n    <b>Strategy:</b>\n     Break down training to timesteps. Find timesteps that increase $\\pi(a|s)$, see whether they match the answer pattern.\n  </div>\n  <div style=\"position:absolute;left:150px;top:252px\" data-anim=\"show@5\">\n    <svg width=\"20\" height=\"46\">\n      <line x1=\"10\" y1=\"0\" x2=\"10\" y2=\"46\" stroke=\"#999\" stroke-width=\"1.5\" marker-end=\"url(#fc-ah)\"></line>\n    </svg>\n  </div>\n  <div class=\"lbl\" style=\"left:175px;top:266px\" data-anim=\"show@5\">Generalize</div>\n  <div class=\"bx question\" style=\"left:10px;top:298px;width:270px\" data-anim=\"show@5\">\n    <b>Q:</b>\n     Why is $m(\\pi)$ high?\n  </div>\n  <div style=\"position:absolute;left:270px;top:312px;pointer-events:none\" data-anim=\"show@5\">\n    <svg width=\"120\" height=\"39\">\n      <line x1=\"110\" y1=\"31\" x2=\"24\" y2=\"11\" stroke=\"#999\" stroke-width=\"1.5\" marker-end=\"url(#fc-ah)\"></line>\n    </svg>\n  </div>\n</div><div class=\"blouse-slide-number\">4</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"5\">\n\n<h2>Experience Breakdown: Method</h2>\n<ol>\n  <li>Define $m(\\pi) = \\pi(a|s)$ as our explanation target</li>\n  <li>Break gradient down into per-timestep gradients $g = \\nabla_\\theta \\frac{1}{N}\\sum_t L_t = \\frac{1}{N}\\sum_t \\nabla_\\theta L_t$</li>\n  <li>For each timestep, compute steerage $\\Omega_t^m = \\nabla_\\theta m \\cdot \\nabla_\\theta L_t$ (Hu 2025)</li>\n  <li>Build a dataset that has the steerage values of all timesteps</li>\n  <li>\n    Find out what high-steerage timesteps have in common:\n    <br>&nbsp;&nbsp;&nbsp;- <i>Forward queries</i>: partition by attribute, compare per-group steerage\n    <br>&nbsp;&nbsp;&nbsp;- <i>Backward queries</i>: filter to high-steerage timesteps\n    <br>&nbsp;&nbsp;&nbsp;- <i>Predictive modeling</i>: fit a model targeting steerage\n  </li>\n</ol>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:30px; right:30px; font-size:11pt; color:#ababab; line-height:1.35\">\n  Hu et al. (2025). <i>A Snapshot of Influence: A Local Data Attribution Framework for Online Reinforcement Learning</i>. NeurIPS.\n</div>\n<div data-anim=\"show@1 +tp-grown@2 +eb-dimmed@1\" style=\"position:absolute; top:200px; left:110px; padding:55px; min-width:603px; background:rgba(255,255,255,0.7); border-radius:8px; z-index:10; color:var(--chai-green); font-size:28pt; display:flex; flex-direction:column; align-items:center\">\n  <div data-anim=\"show@1 hide@2\" style=\"white-space:nowrap\">The problem formulation is novel! 🤯</div>\n  <div class=\"tp\" data-anim=\"show@2\">\n    <div class=\"tp-row\">\n      <span class=\"tp-label\">Existing:</span>\n      <span class=\"tp-text\" data-anim=\"show@2\">Explain one bad action</span>\n      <span class=\"tp-spacer\">\n      </span>\n    </div>\n    <div class=\"tp-arr\">\n      <svg width=\"46\" height=\"90\" viewBox=\"0 0 46 90\">\n        <polygon points=\"14,13 32,13 32,64 46,64 23,90 0,64 14,64\" fill=\"#6d9829\"></polygon>\n      </svg>\n    </div>\n    <div class=\"tp-row\">\n      <span class=\"tp-label\">Ours:</span>\n      <span class=\"tp-text\" data-anim=\"show@2\">Explain a pattern of bad actions</span>\n      <span class=\"tp-spacer\">\n      </span>\n    </div>\n  </div>\n</div><div class=\"blouse-slide-number\">5</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"6\" class=\"overview overview-2\">\n<h1>We’ll talk about:</h1>\n<ul>\n  <li><i>Experience Breakdown</i>: The method that didn’t work 😩</li>\n  <li>Behavior as a new explanation target</li>\n  <li>Explanation targets in XRL</li>\n  <li>HighJax: Our environment for BXRL experiments</li>\n</ul><div class=\"blouse-slide-number\">6</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"7\"><h2>$m(\\pi) : \\Pi \\to \\mathbb{R}$ as a behavior measure: Examples</h2>\n<div style=\"max-width: 500px\">\n  <blockquote>Why does my agent exhibit behavior $m$?</blockquote>\n  <p>Examples:</p>\n  <ul>\n    <li>“Why is my driving agent tailgating other cars?”</li>\n    <li>“Why is my robot limping?”</li>\n    <li>“Why is my datacenter cooling system using too much electricity?”</li>\n  </ul>\n</div>\n<img class=\"dark-include\" src=\"/bxrl-10m/images/chicken.png?v=1786868460\" style=\"position:absolute; right:30px; top:120px; max-width:380px; max-height:380px; border: 1px solid var(--chai-gray)\">\n</img>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:30px; right:30px; font-size:11pt; color:#ababab; line-height:1.35\">\n  Prisco, J. (2021). <i>Why online supermarket Ocado wants to take the human touch out of groceries</i>. CNN.\n</div><div class=\"blouse-slide-number\">7</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"8\"><h2>$m(\\pi) : \\Pi \\to \\mathbb{R}$ as a behavior measure: Benefits</h2>\n<ul>\n  <li>Plot before you explain</li>\n  <li>Focus on recurring problems, not one-offs</li>\n  <li>Avoid overfitting to specific cases</li>\n  <li>Explain multi-action sequences</li>\n</ul><div class=\"blouse-slide-number\">8</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"9\"><h2>$m(\\pi) : \\Pi \\to \\mathbb{R}$ as a behavior measure: Sycophancy</h2>\n<div style=\"position:relative; min-height: 380px\">\n  <div style=\"position:absolute; top:0; left:0; text-align:left; font-size: 22pt; line-height: 1.6\" data-anim=\"hide@1\">\n    <p style=\"margin: 0.4em 0\">$m(\\pi) := \\mathbb{E}_{x \\sim \\mathcal{D}}\\,[$</p>\n    <p style=\"margin: 0.4em 0; padding-left: 2em\">$\\pi_\\theta(\\text{\"You're right to push back on this\"} \\mid x)$</p>\n    <p style=\"margin: 0.4em 0\">$]$</p>\n  </div>\n  <div style=\"position:absolute; top:0; left:0; text-align:left; font-size: 22pt; line-height: 1.6\" data-anim=\"show@1\">\n    <p style=\"margin: 0.4em 0\">$m(\\pi) := \\frac{1}{4} \\mathbb{E}_{x \\sim \\mathcal{D}}\\,[$</p>\n    <p style=\"margin: 0.4em 0; padding-left: 2em\">$\\pi_\\theta(\\text{``You're right to push back on this''} \\mid x) +$</p>\n    <p style=\"margin: 0.4em 0; padding-left: 2em\">$\\pi_\\theta(\\text{``That's an insightful question!''} \\mid x) +$</p>\n    <p style=\"margin: 0.4em 0; padding-left: 2em\">$\\pi_\\theta(\\text{``Great catch, I should have considered that''} \\mid x) +$</p>\n    <p style=\"margin: 0.4em 0; padding-left: 2em\">$\\pi_\\theta(\\text{``And honestly— That's growth ✨''} \\mid x)$</p>\n    <p style=\"margin: 0.4em 0\">$]$</p>\n  </div>\n</div>\n<p style=\"position: absolute; bottom: 25px; left: 30px; right: 30px; font-size: 16pt; color: var(--chai-gray); margin: 0\">$\\mathcal{D}$ = sampled set of user prompts</p><div class=\"blouse-slide-number\">9</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"10\"><h2>$m(\\pi) : \\Pi \\to \\mathbb{R}$ as a behavior measure: Prior art</h2>\n<div style=\"position:absolute; left:30px; top:110px; width:430px; font-size:18pt; line-height:1.5\">\n  <p style=\"margin-bottom:14px\">Used to:</p>\n  <ul style=\"padding-left:24px\">\n    <li style=\"margin-bottom:10px\">Drive exploration via novelty</li>\n    <li style=\"margin-bottom:10px\">Index quality-diversity archives</li>\n    <li style=\"margin-bottom:10px\">Provide gradient axes for QD search</li>\n    <li style=\"margin-bottom:10px\">Auto-learn from data instead of hand-design</li>\n  </ul>\n</div>\n<svg viewBox=\"-130 -60 710 470\" style=\"position:absolute; left:470px; top:112px; width:470px; height:auto\">\n  <circle cx=\"330\" cy=\"190\" r=\"170\" fill=\"rgba(74,111,165,0.07)\" stroke=\"#4a6fa5\" stroke-width=\"2\"></circle>\n  <circle cx=\"210\" cy=\"190\" r=\"110\" fill=\"rgba(218,165,32,0.18)\" stroke=\"#daa520\" stroke-width=\"2\"></circle>\n  <circle cx=\"360\" cy=\"200\" r=\"110\" fill=\"rgba(74,111,165,0.15)\" stroke=\"#4a6fa5\" stroke-width=\"2\"></circle>\n  <circle cx=\"395\" cy=\"215\" r=\"55\" fill=\"rgba(165,74,74,0.18)\" stroke=\"#a54a4a\" stroke-width=\"2\"></circle>\n  <g font-family=\"'Lato',sans-serif\" font-weight=\"700\" font-size=\"32\">\n    <text x=\"330\" y=\"-22\" fill=\"#4a6fa5\" text-anchor=\"middle\">Machine learning</text>\n    <text x=\"96\" y=\"120\" fill=\"#daa520\" text-anchor=\"end\">Evolutionary</text>\n    <text x=\"96\" y=\"156\" fill=\"#daa520\" text-anchor=\"end\">computation</text>\n    <text x=\"362\" y=\"98\" fill=\"#4a6fa5\" text-anchor=\"middle\">RL</text>\n    <text x=\"395\" y=\"152\" fill=\"#a54a4a\" text-anchor=\"middle\">XRL</text>\n  </g>\n  <g font-family=\"'Lato',sans-serif\" font-size=\"26\" fill=\"#222\">\n    <circle cx=\"215\" cy=\"120\" r=\"6\"></circle>\n    <text x=\"229\" y=\"129\" stroke=\"white\" stroke-width=\"5\" paint-order=\"stroke\">Pugh 2016</text>\n    <circle cx=\"215\" cy=\"170\" r=\"6\"></circle>\n    <text x=\"229\" y=\"179\" stroke=\"white\" stroke-width=\"5\" paint-order=\"stroke\">Meyerson 2016</text>\n    <circle cx=\"130\" cy=\"190\" r=\"6\"></circle>\n    <text x=\"116\" y=\"199\" text-anchor=\"end\" stroke=\"white\" stroke-width=\"5\" paint-order=\"stroke\">Lehman 2011</text>\n    <circle cx=\"275\" cy=\"190\" r=\"6\"></circle>\n    <text x=\"261\" y=\"218\" text-anchor=\"end\" stroke=\"white\" stroke-width=\"5\" paint-order=\"stroke\">Conti 2018</text>\n    <circle cx=\"215\" cy=\"265\" r=\"6\"></circle>\n    <text x=\"229\" y=\"274\" stroke=\"white\" stroke-width=\"5\" paint-order=\"stroke\">Fontaine 2021</text>\n    <circle cx=\"395\" cy=\"215\" r=\"7\" fill=\"#a54a4a\"></circle>\n    <text x=\"409\" y=\"224\" stroke=\"white\" stroke-width=\"5\" paint-order=\"stroke\" fill=\"#a54a4a\" font-weight=\"700\">BXRL (us)</text>\n  </g>\n</svg>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:30px; right:30px; font-size:11pt; color:#ababab; line-height:1.35\">\n  <p style=\"margin:0\">Lehman & Stanley (2011). <i>Abandoning Objectives: Evolution Through the Search for Novelty Alone</i>. Evol. Comput.</p>\n  <p style=\"margin:0\">Pugh et al. (2016). <i>Quality Diversity: A New Frontier for Evolutionary Computation</i>. Frontiers in Robotics and AI.</p>\n  <p style=\"margin:0\">Meyerson et al. (2016). <i>Learning Behavior Characterizations for Novelty Search</i>. GECCO.</p>\n  <p style=\"margin:0\">Conti et al. (2018). <i>Improving Exploration in Evolution Strategies for Deep RL</i>. NeurIPS.</p>\n  <p style=\"margin:0\">Fontaine & Nikolaidis (2021). <i>Differentiable Quality Diversity</i>. NeurIPS.</p>\n</div>\n<aside class=\"saturn-venus-note\" data-index=\"9\" class=\"saturn-venus-note\" data-index=\"9\">\n  <ol>\n    <li><b>Drive exploration via novelty</b> (Lehman 2011, Conti 2018). Throw out the fitness function and reward agents for behaving differently from what’s been seen. The behavior measure is the yardstick for “different”. Breaks deceptive optima where every nearby solution looks the same on the fitness landscape.</li>\n    <li><b>Index quality-diversity archives</b> (Pugh 2016, Fontaine 2021). QD wants a collection of high-quality solutions spanning a behavior range, not one best solution. Behavior measures are the axes of that collection. Imagine evolved walkers indexed by leg-length × stride-frequency: each cell holds the best walker exhibiting that combo.</li>\n    <li><b>Provide gradient axes for QD search</b> (Fontaine 2021). When $m$ is smooth in $\\theta$, compute $\\partial m / \\partial \\theta$ and take gradient steps directly along behavior axes. Differentiability turns the behavior measure from something you measure after the fact into a steering wheel.</li>\n    <li><b>Auto-learn from data instead of hand-design</b> (Meyerson 2016). Picking a behavior characterization by hand is brittle. Meyerson trains the BC itself: feed in successful and unsuccessful agents, learn which features best discriminate them. The BC becomes a learnable component, not a hyperparameter.</li>\n  </ol>\n</aside><div class=\"blouse-slide-number\">10</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"11\" class=\"minimal\"><div style=\"position:absolute; inset:0; display:flex; align-items:center; justify-content:center\">\n  <img class=\"dark-include\" src=\"/bxrl-10m/images/def-behavior-measure.png?v=1786868460\" style=\"width:74%; max-height:80%\">\n  </img>\n</div><div class=\"blouse-slide-number\">11</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"12\"><h2>Adapting existing XRL methods for BXRL</h2>\n<p>Each has a modular scalar target. Replace it with $m(\\pi)$:</p>\n<br>\n<ul>\n  <li><b>Data attribution</b> (Hu 2025): which training timesteps raised $m(\\pi)$?</li>\n  <li><b>SVERL-P</b> (Beechey 2023): which observation features contribute to $m(\\pi)$?</li>\n  <li><b>COUNTERPOL</b> (Deshmukh 2023): smallest policy change to reach target $m^*$.</li>\n</ul>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:30px; right:30px; font-size:11pt; color:#ababab; line-height:1.35\">\n  <p style=\"margin:0\">Hu et al. (2025). <i>A Snapshot of Influence: A Local Data Attribution Framework for Online RL</i>. NeurIPS.</p>\n  <p style=\"margin:0\">Beechey et al. (2023). <i>Explaining Reinforcement Learning with Shapley Values</i>. ICML.</p>\n  <p style=\"margin:0\">Deshmukh et al. (2023). <i>Counterfactual Explanation Policies in RL</i>. ICML Workshop.</p>\n</div><div class=\"blouse-slide-number\">12</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"13\" class=\"minimal\"><div style=\"position:absolute; inset:0; display:flex; align-items:center; justify-content:center\">\n  <img class=\"dark-include\" src=\"/bxrl-10m/images/def-bxrl-method.png?v=1786868460\" style=\"width:74%; max-height:80%\">\n  </img>\n</div><div class=\"blouse-slide-number\">13</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"14\" class=\"overview overview-3\">\n<h1>We’ll talk about:</h1>\n<ul>\n  <li><i>Experience Breakdown</i>: The method that didn’t work 😩</li>\n  <li>Behavior as a new explanation target</li>\n  <li>Explanation targets in XRL</li>\n  <li>HighJax: Our environment for BXRL experiments</li>\n</ul><div class=\"blouse-slide-number\">14</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"15\"><p>Position papers warn that XAI lacks clear problem definitions:</p>\n<br>\n<ul>\n  <li>Miller et al. (2017). <b>XAI: Beware of Inmates Running the Asylum</b></li>\n  <li>Lipton (2018). <b>The Mythos of Model Interpretability</b></li>\n  <li>Freiesleben & König (2023). <b>Dear XAI Community, We Need to Talk!</b></li>\n  <li>Haufe et al. (2024). <b>XAI Needs Formal Notions of Explanation Correctness</b></li>\n  <li style=\"font-size: 95%\">Gyevnar & Towers (2025). <b>Objective Metrics for Human-Subjects Evaluation in XRL</b></li>\n</ul><div class=\"blouse-slide-number\">15</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"16\" class=\"minimal\"><div style=\"display:flex; align-items:center; justify-content:center; flex-direction:column; height:540px; width:340px; position:absolute; right:40px; top:0\">\n  <img src=\"/bxrl-10m/images/Hempel.jpg?v=1786868460\" style=\"max-height:360px; border-radius:8px; box-shadow:0 4px 12px rgba(0,0,0,0.15)\">\n  </img>\n  <p style=\"margin-top:14px; text-align:center; font-size:14pt; line-height:1.35; color:var(--chai-gray)\">\n    Hempel (1965). <i>Aspects of Scientific Explanation</i>. Free Press.\n  </p>\n</div>\n<div style=\"position:absolute; left:90px; top:80px; width:500px\">\n  <p style=\"font-size:22pt; color:var(--chai-gray); line-height:1.5\">\n    <b style=\"color:var(--chai-green)\">Explanandum / explanation target</b>\n    <br>\n    <span style=\"font-size:18pt\">“Why is the sky blue?”</span>\n  </p>\n  <svg style=\"margin:5px 0 -30px 156px\" width=\"40\" height=\"140\" viewBox=\"0 0 40 140\">\n    <line x1=\"20\" y1=\"0\" x2=\"20\" y2=\"122\" stroke=\"var(--chai-green)\" stroke-width=\"5\"></line>\n    <polygon points=\"6,122 20,140 34,122\" fill=\"var(--chai-green)\"></polygon>\n  </svg>\n  <p style=\"font-size:22pt; color:var(--chai-gray); line-height:1.5\">\n    <b style=\"color:var(--chai-green)\">Explanans / explanation source</b>\n    <br>\n    <span style=\"font-size:18pt\">“Sunlight scatters in atmosphere, blue has shorter wavelength so it scatters the most.”</span>\n  </p>\n</div>\n<aside class=\"saturn-venus-note\" data-index=\"15\" class=\"saturn-venus-note\" data-index=\"15\">\n  <p>Problem domain, solution domain.</p>\n  <p>mixed feelings, condenscending, like management consultant, there’s thing called PROBLEM, and SOLUTION SOLVES the PROBLEM, here’s my INVOICE</p>\n</aside><div class=\"blouse-slide-number\">16</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"17\">\n<h2>Example: Reward decomposition (Juozapaitis 2019)</h2>\n<div class=\"ex-split\">\n  <div>\n    <p><b>Explanation target:</b> “Why did the agent choose $a_1$ over $a_2$?”</p>\n    <p><b>Explanation source:</b> Per-reward-type Q-values $Q_c(s, a)$ &mdash; one per reward component (velocity, fuel, crash...)</p>\n    <p><b>Output:</b> Bar chart of per-type advantages &mdash; e.g. “<i>fire-main</i> beats <i>no-op</i> because +ground-contact outweighs −landing-pad-distance.”</p>\n  </div>\n  <div>\n    <img class=\"dark-include\" src=\"/bxrl-10m/images/juozapaitis2019-rdx.png?v=1786868460\" style=\"height: 368px; width: auto;\">\n    </img>\n  </div>\n</div>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:30px; right:30px; font-size:11pt; color:#ababab; line-height:1.35\">\n  Juozapaitis et al. (2019). <i>Explainable Reinforcement Learning via Reward Decomposition</i>. IJCAI/ECAI Workshop on XAI.\n</div><div class=\"blouse-slide-number\">17</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"18\">\n<h2>Example: Causal chains (Madumal 2020)</h2>\n<div class=\"ex-split\">\n  <div>\n    <p><b>Explanation target:</b> “Why did the agent choose $a_1$ over $a_2$?”</p>\n    <p><b>Explanation source:</b> Human builds DAG of environment events, agent learns weights between them during training.</p>\n    <p><b>Output:</b> A causal chain: “I built a supply depot rather than barracks because more supply → more units → more destroyed buildings (the goal).”</p>\n  </div>\n  <div>\n    <img class=\"dark-include\" src=\"/bxrl-10m/images/madumal2020-causal.png?v=1786868460\">\n    </img>\n  </div>\n</div>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:500px; right:30px; font-size:11pt; color:#ababab; line-height:1.35\">\n  Madumal et al. (2020). <i>Explainable Reinforcement Learning Through a Causal Lens</i>. AAAI.\n</div><div class=\"blouse-slide-number\">18</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"19\"><h2>Explanation targets in XRL</h2>\n<img class=\"dark-include\" src=\"/bxrl-10m/images/explanation-targets-table.png?v=1786868460\" style=\"width:100%\">\n</img>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:30px; right:30px; font-size:11pt; color:#ababab; line-height:1.35\">\n  <p style=\"margin:0\">Vouros (2022). <i>Explainable Deep Reinforcement Learning: State of the Art and Challenges</i>. ACM Computing Surveys.</p>\n  <p style=\"margin:0\">Milani et al. (2024). <i>Explainable Reinforcement Learning: A Survey and Comparative Review</i>. ACM Computing Surveys.</p>\n  <p style=\"margin:0\">Saulières et al. (2025). <i>A Survey of Explainable Reinforcement Learning: Targets, Methods and Needs</i>. arXiv.</p>\n</div><div class=\"blouse-slide-number\">19</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"20\" class=\"overview overview-4\">\n<h1>We’ll talk about:</h1>\n<ul>\n  <li><i>Experience Breakdown</i>: The method that didn’t work 😩</li>\n  <li>Behavior as a new explanation target</li>\n  <li>Explanation targets in XRL</li>\n  <li>HighJax: Our environment for BXRL experiments</li>\n</ul><div class=\"blouse-slide-number\">20</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"21\" class=\"minimal\"><h2>HighwayEnv (Leurent 2018)</h2>\n<p>github.com/Farama-Foundation/HighwayEnv</p>\n<div style=\"position: absolute; top: 0; left: 0; width: 100%; height: 100%\">\n  <img src=\"/bxrl-10m/images/highway.gif?v=1786868460\" style=\"width: 100%; height: 100%; object-fit: contain\">\n  </img>\n</div>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:30px; right:30px; font-size:11pt; color:#ababab; line-height:1.35\">\n  Leurent (2018). <i>An Environment for Autonomous Driving Decision-Making</i>. GitHub.\n</div><div class=\"blouse-slide-number\">21</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"22\" class=\"minimal\" data-bg=\"#1e1e1e\"><div style=\"position: absolute; top: 0; left: 0; width: 100%; height: 100%; z-index: 0\">\n  <video src=\"/bxrl-10m/videos/outrun-barebones.mp4?v=1786868460\" muted controls loop style=\"width: 100%; height: 100%; object-fit: contain\" data-blouse-autoplay=\"1\" preload=\"auto\">\n  </video>\n</div>\n<div style=\"position: relative; z-index: 1\">\n  <h2>HighJax: our port of HighwayEnv to JAX</h2>\n  <p>github.com/HumanCompatibleAI/HighJax</p>\n</div>\n<div class=\"cite\" style=\"position:absolute; bottom:14px; left:30px; right:30px; font-size:11pt; color:rgba(255,255,255,0.75); line-height:1.35\">\n  Bradbury et al. (2018). <i>JAX: composable transformations of Python+NumPy programs</i>. GitHub.\n</div>\n<aside class=\"saturn-venus-note\" data-index=\"21\" class=\"saturn-venus-note\" data-index=\"21\">\n  <p>This is one of the contributions of this project.</p>\n  <p>It’s a lot more efficient than the existing implementing because all of the environment rules are defined in terms of matrix operations rather than imperative code. This means that rollouts can be done on the GPU, and they can be batched. This is very helpful in RL, because the computational bottleneck is sequential.</p>\n</aside><div class=\"blouse-slide-number\">22</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"23\" class=\"minimal\" data-bg=\"black\"><div style=\"position: absolute; top: 0; left: 0; width: 100%; height: 100%; z-index: 0\">\n  <video src=\"/bxrl-10m/videos/outrun-dark.mp4?v=1786868460\" muted controls loop style=\"width: 100%; height: 100%; object-fit: contain\" data-blouse-autoplay=\"1\" preload=\"auto\">\n  </video>\n</div>\n<div style=\"position: relative; z-index: 1\">\n  <h2>HighJax: our port of HighwayEnv to JAX</h2>\n  <p>github.com/HumanCompatibleAI/HighJax</p>\n</div><div class=\"blouse-slide-number\">23</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"24\" class=\"minimal\" data-bg=\"black\"><div style=\"position: absolute; top: 0; left: 0; width: 100%; height: 100%; z-index: 0; background: transparent !important\">\n  <video src=\"/bxrl-10m/videos/outrun-tui-playback.mp4?v=1786868460\" muted controls loop style=\"width: 100%; height: 100%; object-fit: contain\" data-blouse-autoplay=\"1\" preload=\"auto\">\n  </video>\n</div><div class=\"blouse-slide-number\">24</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"25\" class=\"minimal\" data-bg=\"black\"><div style=\"position: absolute; inset: 0; display: flex; align-items: center; justify-content: center\">\n  <img src=\"/bxrl-10m/images/octane-behavior-dialog.svg?v=1786868460\" style=\"width: 100%; height: 100%; object-fit: contain\">\n  </img>\n</div><div class=\"blouse-slide-number\">25</div></section>\n</foreignObject>\n</svg>\n<svg data-mercury-svg viewBox=\"0 0 960 540\" preserveAspectRatio=\"xMidYMid meet\">\n<foreignObject width=\"960\" height=\"540\">\n<section id=\"26\" class=\"closing\"><p>humancompatible.ai</p>\n<p>Ram Rachum</p>\n<p>ram.rachum@berkeley.edu</p><div class=\"blouse-slide-number\">26</div></section>\n</foreignObject>\n</svg>\n</div>","css":"div.mercury>svg>foreignObject>section:where(section){--chai-green:#6d9829;--chai-green-dark:#5e8722;--chai-green-light:#7da933;--chai-gray:#595959;--text-muted:#888;--accent-bar-width:19px;container-type:size}div.mercury>svg>foreignObject>section:where(section) *{margin:0;padding:0}div.mercury>svg>foreignObject>section:where(section){width:960px;height:540px;color:var(--chai-gray);border-left:var(--accent-bar-width) solid var(--chai-green);background:#fff;padding:56px 41px 50px 22px;font-family:Lato,Segoe UI,Roboto,Helvetica Neue,Arial,sans-serif;font-size:18pt;position:relative}@font-face{font-family:Cascadia Mono;font-style:normal;font-weight:400;font-display:swap;src:url(data:font/woff2;base64,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)format(\"woff2\")}div.mercury>svg>foreignObject>section:where(section) .blouse-slide-number{color:#ababab;z-index:2;opacity:1;white-space:pre;text-align:right;min-width:5ch;font-family:Cascadia Mono,Consolas,Courier New,monospace;font-size:.55em;line-height:1;position:absolute;bottom:5px;right:50px}div.mercury>svg>foreignObject>section:where(section).lead .blouse-slide-number,div.mercury>svg>foreignObject>section:where(section).closing .blouse-slide-number{display:none}div.mercury>svg>foreignObject>section:where(section).invert .blouse-slide-number{color:#ffffffb3}div.mercury>svg>foreignObject>section:where(section) h1{color:var(--chai-green);border:none;margin-top:0;margin-bottom:.735em;font-size:28pt;font-weight:400;line-height:1!important}div.mercury>svg>foreignObject>section:where(section) h2{color:var(--chai-green);border:none;margin-bottom:.8em;font-size:28pt;font-weight:400}div.mercury>svg>foreignObject>section:where(section) h3{color:var(--chai-green-dark);font-size:1.1em;font-weight:600}div.mercury>svg>foreignObject>section:where(section) p{color:inherit;margin-top:18px;line-height:1.37}div.mercury>svg>foreignObject>section:where(section) ul,div.mercury>svg>foreignObject>section:where(section) ol{color:var(--chai-gray);margin-top:18px;line-height:1.37}div.mercury>svg>foreignObject>section:where(section) li{margin-top:11px}div.mercury>svg>foreignObject>section:where(section) ul li::marker,div.mercury>svg>foreignObject>section:where(section) ol li::marker{color:var(--chai-gray)}div.mercury>svg>foreignObject>section:where(section) ul,div.mercury>svg>foreignObject>section:where(section) ol{padding-left:1.5em}div.mercury>svg>foreignObject>section:where(section) a{color:var(--chai-green);text-decoration:underline}div.mercury>svg>foreignObject>section:where(section) a:hover{color:var(--chai-green-dark)}div.mercury>svg>foreignObject>section:where(section) code{color:var(--chai-gray);background:#f5f5f5;border-radius:3px;padding:2px 6px;font-size:.9em}div.mercury>svg>foreignObject>section:where(section) pre{border-left:4px solid var(--chai-green);background:#f8f8f8;border-radius:4px;padding:1em}div.mercury>svg>foreignObject>section:where(section) pre code{background:0 0;padding:0}div.mercury>svg>foreignObject>section:where(section) table{border-collapse:collapse;margin:.5em auto}div.mercury>svg>foreignObject>section:where(section) table th,div.mercury>svg>foreignObject>section:where(section) table td{text-align:center;background:#fafaf5;border:1px solid #333;padding:10px 20px}div.mercury>svg>foreignObject>section:where(section) table th{font-style:italic;font-weight:400}div.mercury>svg>foreignObject>section:where(section) blockquote{border-left:4px solid var(--chai-green);color:#666;background:#f9f9f9;margin:1em 0;padding:.5em 1em;font-style:italic}div.mercury>svg>foreignObject>section:where(section) blockquote>p:first-child{margin-top:0}div.mercury>svg>foreignObject>section:where(section):before{content:\"\";background:url(/themes/chai/small-chai-logo.png) 50%/contain no-repeat;width:52px;height:52px;position:absolute;bottom:0;right:0}div.mercury>svg>foreignObject>section:where(section).lead:before,div.mercury>svg>foreignObject>section:where(section).section-header:before,div.mercury>svg>foreignObject>section:where(section).split-heading:before{display:none}div.mercury>svg>foreignObject>section:where(section) .chai-logo-small{width:36px;height:36px;display:none;position:absolute;bottom:15px;right:20px}div.mercury>svg>foreignObject>section:where(section) footer{color:var(--chai-gray);font-size:.6em;position:absolute;bottom:20px;left:50px;right:80px}div.mercury>svg>foreignObject>section:where(section) footer em,div.mercury>svg>foreignObject>section:where(section) footer i{font-style:italic}div.mercury>svg>foreignObject>section:where(section).lead{text-align:center;background:linear-gradient(to bottom, var(--chai-green) 70.6%, white 70.6%);border-left:none;flex-direction:column;justify-content:flex-start;align-items:center;padding:70px 15px 36%;display:flex}div.mercury>svg>foreignObject>section:where(section).lead h1{color:#fff;margin:0 0 80px;font-size:36pt;font-weight:700;line-height:1.2}div.mercury>svg>foreignObject>section:where(section).lead h2{color:#fff;margin:0 0 .1em;font-size:28pt;font-weight:400}div.mercury>svg>foreignObject>section:where(section).lead p{color:#ffffffe6;margin:.05em 0;line-height:1.2;font-size:28pt!important}div.mercury>svg>foreignObject>section:where(section).lead a{color:#ffffffe6;text-decoration:none}div.mercury>svg>foreignObject>section:where(section).lead:before{content:\"\";background:url(/themes/chai/big-chai-logo.png) 0/contain no-repeat;width:344px;height:149px;display:block;position:absolute;bottom:5px;left:5px}div.mercury>svg>foreignObject>section:where(section).lead .lead-footer{color:var(--chai-green);text-align:right;font-size:.55em;line-height:1.4;position:absolute;bottom:20px;right:40px}div.mercury>svg>foreignObject>section:where(section).lead .lead-footer a{color:var(--chai-green);text-decoration:underline}div.mercury>svg>foreignObject>section:where(section).lead:after{display:none}div.mercury>svg>foreignObject>section:where(section).lead .lead-footer-right{position:absolute;bottom:15px;right:20px}div.mercury>svg>foreignObject>section:where(section).lead .lead-footer-right img{height:132px}div.mercury>svg>foreignObject>section:where(section).statement{align-items:center;padding:47px 0 47px 32px;display:flex}div.mercury>svg>foreignObject>section:where(section).statement p{color:var(--chai-green);text-align:left;max-width:890px;margin:0;font-size:38pt;font-weight:400;line-height:1.21}div.mercury>svg>foreignObject>section:where(section).minimal{border-left:none;padding:0}div.mercury>svg>foreignObject>section:where(section).statement-alt{text-align:center;justify-content:center;align-items:center;padding:47px 120px;display:flex}div.mercury>svg>foreignObject>section:where(section).statement-alt p{color:var(--chai-green);max-width:800px;margin:0;font-size:30pt;font-weight:400;line-height:1.3}div.mercury>svg>foreignObject>section:where(section).section-header{text-align:center;background:linear-gradient(to bottom, var(--chai-green) 70.5%, white 70.5%);border-left:none;flex-direction:column;justify-content:flex-start;align-items:center;padding:130px 15px 36%;display:flex}div.mercury>svg>foreignObject>section:where(section).section-header h1{color:#fff;font-size:36pt;font-weight:400}div.mercury>svg>foreignObject>section:where(section).section-header h2{color:#fffc;font-size:28pt;font-weight:300}div.mercury>svg>foreignObject>section:where(section).invert{background:var(--chai-green);color:#fff;border-left:none}div.mercury>svg>foreignObject>section:where(section).invert h1,div.mercury>svg>foreignObject>section:where(section).invert h2,div.mercury>svg>foreignObject>section:where(section).invert h3{color:#fff}div.mercury>svg>foreignObject>section:where(section).invert p,div.mercury>svg>foreignObject>section:where(section).invert ul,div.mercury>svg>foreignObject>section:where(section).invert ol{color:#ffffffe6}div.mercury>svg>foreignObject>section:where(section).invert a{color:#fff;text-decoration:underline}div.mercury>svg>foreignObject>section:where(section).invert code{color:#fff;background:#ffffff26}div.mercury>svg>foreignObject>section:where(section).invert blockquote{color:#ffffffd9;background:#ffffff1a;border-left-color:#fff}div.mercury>svg>foreignObject>section:where(section).columns{grid-template-columns:1fr 1fr;gap:40px;display:grid}div.mercury>svg>foreignObject>section:where(section).columns-60-40{grid-template-columns:60% 40%;gap:40px;display:grid}div.mercury>svg>foreignObject>section:where(section).columns-40-60{grid-template-columns:40% 60%;gap:40px;display:grid}div.mercury>svg>foreignObject>section:where(section) img{border-radius:4px;max-width:100%;height:auto}div.mercury>svg>foreignObject>section:where(section) img[alt~=center]{margin:0 auto;display:block}div.mercury>svg>foreignObject>section:where(section) img[alt~=shadow]{box-shadow:0 4px 12px #00000026}div.mercury>svg>foreignObject>section:where(section) .highlight{border-left:4px solid var(--chai-green);background:#6d98291a;border-radius:0 4px 4px 0;margin:1em 0;padding:1em}div.mercury>svg>foreignObject>section:where(section) .warning{background:#fff3cd;border-left:4px solid #ffc107;border-radius:0 4px 4px 0;margin:1em 0;padding:1em}div.mercury>svg>foreignObject>section:where(section).split-heading{background:#fff;border-left:none;grid-template-columns:50% 50%;padding:0;display:grid}div.mercury>svg>foreignObject>section:where(section).split-heading h1{text-align:center;color:var(--chai-green);justify-content:center;align-self:center;align-items:center;margin:0;padding:20px 40px 60px;font-size:42pt;font-weight:400;display:flex}div.mercury>svg>foreignObject>section:where(section).split-heading ol,div.mercury>svg>foreignObject>section:where(section).split-heading ul{height:100%;color:var(--chai-gray);background:#efefef;margin:0;padding:34% 40px 40px 80px;font-size:20pt;line-height:1}div.mercury>svg>foreignObject>section:where(section).split-heading li{margin-top:38px;padding-left:1em;display:list-item}div.mercury>svg>foreignObject>section:where(section).split-heading li:first-child{margin-top:0}div.mercury>svg>foreignObject>section:where(section).split-heading:before{display:none}div.mercury>svg>foreignObject>section:where(section).closing{background:linear-gradient(to bottom, white 73.2%, var(--chai-green) 73.2%);border-left:none;flex-direction:column;justify-content:flex-start;align-items:center;padding:0;display:flex}div.mercury>svg>foreignObject>section:where(section).closing:before{content:\"\";background:url(/themes/chai/big-chai-logo.png) 50%/contain no-repeat;width:49.7%;height:38.3%;display:block;position:absolute;top:16.4%;left:25.1%}div.mercury>svg>foreignObject>section:where(section).closing p:first-of-type{color:var(--chai-gray);text-align:center;font-size:18pt;position:absolute;top:62.4%}div.mercury>svg>foreignObject>section:where(section).closing p:nth-of-type(2){text-align:center;color:#fff;margin:0;font-size:28pt;position:absolute;top:80%;left:0;right:0}div.mercury>svg>foreignObject>section:where(section).closing p:nth-of-type(3){text-align:center;color:#fff;margin:0;font-size:18pt;position:absolute;top:90%;left:0;right:0}div.mercury>svg>foreignObject>section:where(section).closing a{color:#fff;text-decoration:none}[data-anim].anim-dim{opacity:.3;transition:opacity .3s}[data-anim].anim-strike{color:var(--chai-gray);text-decoration:line-through;transition:color .3s,text-decoration .3s}[data-anim].anim-highlight{color:var(--chai-green);font-weight:700;transition:color .3s,font-weight .3s}body.apron-slide{background:#fff}body.apron-slide #blouse-apron-overlay{background:linear-gradient(to right, #6d9829 calc(var(--apron-left) + 19px * var(--apron-scale) + 1px), transparent 0)}body.apron-minimal #blouse-apron-overlay,body.apron-lead #blouse-apron-overlay,body.apron-section-header #blouse-apron-overlay,body.apron-closing #blouse-apron-overlay,body.apron-invert #blouse-apron-overlay,body.apron-split-heading #blouse-apron-overlay{background:0 0}body.apron-minimal{background:#fff}body.apron-lead,body.apron-section-header{background:linear-gradient(to bottom, #6d9829 calc(var(--apron-top) + .706 * (var(--apron-bottom) - var(--apron-top))), white 0)}body.apron-closing{background:linear-gradient(to bottom, white calc(var(--apron-top) + .732 * (var(--apron-bottom) - var(--apron-top))), #6d9829 0)}body.apron-invert{background:#6d9829}div.mercury>svg>foreignObject>section:where(section){box-sizing:border-box;overflow:hidden}h1::part(auto-scaling),h2::part(auto-scaling),h3::part(auto-scaling),h4::part(auto-scaling),h5::part(auto-scaling),h6::part(auto-scaling){max-height:563px}pre::part(auto-scaling){max-height:529px}@page{size:960px 540px;margin:0}@media print{body{margin:0}}\n\n\n  .overview-1 li:nth-child(-n+1):not(:nth-child(1)) { text-decoration: line-through; opacity: 0.5; color: var(--chai-gray); }\n  .overview-1 li:nth-child(1) { color: var(--chai-green); font-weight: 700; }\n\n\n\n\n  section:has(.fc) { --question-bg:#edf2ff; --question-pre:#4a6fa5; --answer-bg:#edfff2; --answer-pre:#3d8b5e; --bad-answer-bg:#ffeded; --bad-answer-pre:#a54a4a; --strategy-bg:#fff8ed; --strategy-pre:#a5804a; }\n  .fc .bx { position:absolute; padding:5px 8px !important; border-radius:6px; font-size:18pt; line-height:1.3; border:1.5px solid rgba(0,0,0,0.06); box-shadow:0 1px 3px rgba(0,0,0,0.04); }\n  .fc .question { background:var(--question-bg); } .fc .question b { color:var(--question-pre); }\n  .fc .answer { background:var(--answer-bg); } .fc .answer b { color:var(--answer-pre); }\n  .fc .bad-answer { background:var(--bad-answer-bg); } .fc .bad-answer b { color:var(--bad-answer-pre); }\n  .fc .strategy { background:var(--strategy-bg); } .fc .strategy b { color:var(--strategy-pre); }\n  .fc .lbl { position:absolute; font-size:12pt; color:#666; font-style:italic; white-space:nowrap; }\n\n\n\n  .tp { width: 100%; display: flex; flex-direction: column; }\n  .tp-grown { width: 634px; }\n  .tp-row { display: flex; align-items: baseline; }\n  .tp-label { width: 110px; flex: none; text-align: right; padding-right: 22px; box-sizing: border-box; white-space: nowrap; position: relative; left: -30px; color: #000; font-size: 17pt; font-weight: 700; }\n  .tp-text { flex: 1; text-align: center; font-size: 22pt; font-weight: 700; color: #6d9829; white-space: nowrap; line-height: 1.35; }\n  .tp-spacer { width: 110px; flex: none; }\n  .tp-arr { display: flex; justify-content: center; margin: 16px 0; }\n\n\n  section:has(.eb-dimmed) > ol > li:nth-child(n+2) { opacity: 0.28; transition: opacity 0.3s; }\n  section:has(.eb-dimmed) > ol > li:first-child { background: rgba(255, 255, 0, 0.2); transition: background 0.3s; }\n\n\n\n  .overview-2 li:nth-child(-n+2):not(:nth-child(2)) { text-decoration: line-through; opacity: 0.5; color: var(--chai-gray); }\n  .overview-2 li:nth-child(2) { color: var(--chai-green); font-weight: 700; }\n\n\n\n\n\n\n\n\n\n\n  .overview-3 li:nth-child(-n+3):not(:nth-child(3)) { text-decoration: line-through; opacity: 0.5; color: var(--chai-gray); }\n  .overview-3 li:nth-child(3) { color: var(--chai-green); font-weight: 700; }\n\n\n\n\n\n  .ex-split { display:flex; gap:20px; align-items:flex-start; }\n  .ex-split > div:first-child { flex:1; }\n  .ex-split > div:last-child { flex:0 0 55%; }\n  .ex-split img { width:100%; border-radius:4px; }\n\n\n\n  .ex-split { display:flex; gap:20px; align-items:flex-start; }\n  .ex-split > div:first-child { flex:1; }\n  .ex-split > div:last-child { flex:0 0 50%; }\n  .ex-split img { width:100%; border-radius:4px; }\n\n\n\n\n  .overview-4 li:nth-child(-n+4):not(:nth-child(4)) { text-decoration: line-through; opacity: 0.5; color: var(--chai-gray); }\n  .overview-4 li:nth-child(4) { color: var(--chai-green); font-weight: 700; }\n\n\n\n\n\n\n\n\n"}