index.html

<!doctype html>
<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><meta name="theme-color" content="#174a43"><title>Recursive Play — EnvLoop Research</title><link rel="stylesheet" href="styles.css"><script defer src="app.js"></script></head>
<body><header class="masthead"><a class="brand" href="#top"><span class="brand-mark">▦</span> EnvLoop <span>Research</span></a><nav aria-label="Main"><a href="#play">Try a puzzle</a><a href="#results">Results</a><a href="#examples">Examples</a><a class="paper-link" href="recursive-play-paper.pdf" target="_blank" rel="noopener">Read paper ↗</a></nav></header>
<main id="top"><section class="hero"><div class="hero-copy"><p class="eyebrow">A GENERATED CHALLENGE BANK</p><h1>How far can an agent<br>get without the rules?</h1><p class="lead">Pixels, neutral controls, six changing levels. Explore real puzzle mechanics and the gaps left by eleven evaluated model configurations.</p><a class="primary-link" href="#play">Step into the experiment <span>↓</span></a></div><div class="hero-note"><span class="large-number">15<span>/150</span></span><p>curricula remained unfinished<br>by <strong>every tested configuration</strong>.</p><div class="note-rule"></div><p class="small">At the fixed 512-action / 720-second budget.<br>Two of these tasks yielded zero completed<br>levels across all eleven configurations.</p></div></section>
<section class="stats" aria-label="Benchmark scale"><div><strong>150</strong><span>task candidates</span></div><div><strong>900</strong><span>offered levels</span></div><div><strong>11</strong><span>model configurations</span></div><div><strong>1,650</strong><span>audited outcomes</span></div></section>
<section id="play" class="section play-section"><div class="section-head"><p class="eyebrow">01 / INTERACT</p><h2>A real puzzle, in your browser.</h2><p>The original rules run locally. Explore the numbered controls and watch the pixels change.</p></div><div class="play-layout"><aside class="play-side"><label for="environment">Choose a matched example</label><select id="environment"><option value="a001">a001 / Pivot cargo</option><option value="a005">a005 / Local frame rover</option><option value="d009">d009 / Escrow Cycle</option><option value="e005">e005 / Transitive islands</option></select><p id="case-description"></p><div class="practice-label">LOCAL PRACTICE</div><p class="small">Source-exposed practice. No model is running and no leaderboard score is submitted.</p><button id="new-practice" class="outline-button">Start new practice</button><p class="small">For spatial tasks, current state is on the left and the target is on the right. Other tasks encode their constraints in the image.</p></aside><div class="game-card"><div class="game-top"><span id="game-label">a001 · level 1</span><div><span id="game-actions">0 / 512 actions</span><span id="game-clock">720 s left</span></div></div><div class="canvas-stage"><canvas id="pixels" aria-label="Interactive puzzle pixel observation" role="img"></canvas></div><div id="game-controls" class="controls" aria-label="Numbered puzzle controls"></div><div class="game-bottom"><button id="reset-level">Reset level <span>+1 action</span></button><button id="next-level" hidden>Next level →</button><strong id="game-status" aria-live="polite">Explore the controls.</strong></div><p class="small canvas-help">Keyboard: 1–6; R resets. For coordinate controls, click a cell on the grid.</p></div></div></section>
<section id="results" class="section"><div class="section-head split-head"><div><p class="eyebrow">02 / COMPARE</p><h2>Full wins tell only part of the story.</h2></div><a class="text-link" href="https://huggingface.co/datasets/EnvLoop/Recursive-Play-Bench" target="_blank" rel="noopener">Open the dataset ↗</a></div><p class="section-intro">Each configuration faces the same 150 tasks. A full win clears all six levels; level completion retains partial progress out of 900 offered levels. Equal win counts share a rank.</p><div class="table-wrap"><table><thead><tr><th>Rank</th><th>Configuration</th><th>Full tasks / 150</th><th>Full rate</th><th>Levels / 900</th><th>Level rate</th><th>720 s endings</th></tr></thead><tbody id="leaderboard"></tbody></table></div><details class="method-note"><summary>What does a 720-second ending mean?</summary><p>The six-level curriculum was unfinished when the normal game clock expired. Completed levels still count. The clock includes model-response waiting, tools and gameplay; the ending alone does not establish its cause. Confirmed engineering interruptions were independently reviewed and kept outside the eligible score population.</p></details><div class="chart-switcher" role="tablist" aria-label="Result chart"><button role="tab" aria-selected="true" data-chart="fullwins-vs-levels">Full vs partial</button><button role="tab" aria-selected="false" data-chart="completion-depth-0-to-6">Completion depth</button><button role="tab" aria-selected="false" data-chart="authored-family-heatmap">Task families</button><button role="tab" aria-selected="false" data-chart="normal-budget-endings">Episode endings</button></div><figure class="chart-panel"><img id="result-chart" src="fullwins-vs-levels.svg" alt="Complete tasks and cleared levels for all eleven model configurations"><figcaption id="chart-caption">Exact count columns use 150 tasks and 900 levels as separate denominators.</figcaption></figure></section>
<section id="examples" class="section examples-section"><div class="section-head"><p class="eyebrow">03 / LOOK CLOSER</p><h2>Same task. Different stopping points.</h2><p>Purposively selected matched examples connect scores to unmodified recorded observations.</p></div><div class="example-switcher" role="tablist" aria-label="Recorded example"><button data-example="a001" role="tab" aria-selected="true">a001 · 6 / 5 / 3</button><button data-example="a005" role="tab" aria-selected="false">a005 · ranking reversal</button><button data-example="d009" role="tab" aria-selected="false">d009 · full vs zero</button><button data-example="e005" role="tab" aria-selected="false">e005 · full / partial / zero</button></div><div id="recorded-example"></div><p class="small">Terminal panels may show different attained levels. Times shown are authoritative game times; a saved observation may postdate the terminal timestamp. Different settings, serving routes and dates limit model-only causal attribution.</p></section>
<section class="section pathway"><div><p class="eyebrow">04 / WHAT COMES NEXT</p><h2>From challenges<br>to a testable learning loop.</h2><p>The bank supplies solvable, resettable challenges and verified traces. A future learning experiment can select eligible failures, construct candidate expert data, then test updates on frozen hidden tasks with regression checks.</p><p class="small">The current study reports construction and solving outcomes. It does not report a measured training gain.</p></div><img src="improvement-pathway.svg" alt="Observed task generation, verification and blind solving; proposed candidate learning and held-out evaluation"></section>
<footer><div class="brand">▦ EnvLoop Research</div><a href="mailto:research@envloop.ai">research@envloop.ai</a><a href="recursive-play-paper.pdf" target="_blank" rel="noopener">24-page paper ↗</a><a href="asset-provenance.json">Artifact provenance ↗</a></footer></main></body></html>