borderland-rl / proof.html
FatUnc6969's picture
Publish reviewed Borderland recovery research artifact
c5f23a7 verified
Raw History Blame Contribute Delete
8.1 kB
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover">
<meta name="theme-color" content="#101514">
<meta name="color-scheme" content="dark">
<title>Borderland | Archive learning proof</title>
<link rel="stylesheet" href="./assets/style.css">
<link rel="stylesheet" href="./proof.css">
<script src="./proof.js" defer></script>
</head>
<body>
<a class="skip" href="#proof">Skip to evidence</a>
<header class="topbar"><a class="brand" href="./"><span class="brand-mark" aria-hidden="true">B</span><span>BORDERLAND<span class="brand-sub">A CAUSAL CAMPAIGN</span></span></a><nav class="proof-nav" aria-label="Evidence navigation"><a href="./">Campaign and Beam evidence</a><a href="#witnesses">Replay actual witnesses</a></nav><span class="offline">LOCAL RECEIPTS 路 NO INFERENCE</span></header>
<main id="proof">
<section class="proof-intro"><p class="eyebrow">ARCHIVE STUDY 路 HYBRID LEARNING EVIDENCE</p><h1>Five weights learned<br>Archive applicability choices<span class="accent">.</span></h1><p>A reward-trained selector improved under substantial scripted assistance. The frozen selector then transferred to full-start campaigns. This is local policy learning. Beam improvement remains unproved.</p><p class="scope-note">Saved events only. This page makes no model calls and runs no campaign engine.</p></section>
<p id="load-status" role="status" aria-live="polite">Loading hash-verified study projection...</p>
<div id="proof-content" hidden>
<section aria-labelledby="results-title"><div class="proof-heading"><h2 id="results-title">What changed after training</h2><span class="eyebrow">FROZEN BEFORE AND AFTER</span></div><div id="metrics" class="proof-metrics"></div><aside class="assistance"><h3>Assistance is part of every result</h3><p>Only the five-weight Archive selector learns. It chooses all eight applicability records at confirmation and transfer. Scripts control exploration, evidence acquisition, derived facts, comparison proof, filing, bind execution, notebook writing, and every other room. During training, a public conditional teacher fills services the learner does not yet own.</p><p>Full-start transfer preserves this assistance. A fresh scripted controller receives the notebook at Blackout. Erased runs use the existing scripted audits at their actual charge cost. No fully learned campaign solver is established.</p></aside></section>
<section class="proof-section" aria-labelledby="training-title"><div class="proof-heading"><h2 id="training-title">Reward training across three initializations</h2><span id="training-total" class="eyebrow"></span></div><p id="curve-scope" class="muted"></p><div id="learning-chart" class="learning-chart"></div><div id="chart-legend" class="chart-legend"></div><details class="proof-detail"><summary>Inspect checkpoint weights and training blocks</summary><div id="checkpoints"></div></details></section>
<section class="proof-section" aria-labelledby="memory-title"><div class="proof-heading"><h2 id="memory-title">Retained memory saves charge</h2><span class="eyebrow">96 MATCHED CONTRASTS</span></div><div class="memory-grid"><div id="memory-result" class="memory-result"></div><div><p>Both trained arms escape in all 96 trials. Erasure does not prevent escape in this study. The scripted controller recovers facts through audits.</p><p id="memory-detail" class="muted"></p><p class="muted">The 96 pairs share an identical prefix before intervention. They represent 16 public root groups, each with two twins and three initializations. They do not establish full-world memory necessity or a global memoryless bound.</p></div></div><details class="proof-detail"><summary>Inspect all 96 paired charge contrasts</summary><div id="contrast-table" class="proof-table"></div></details></section>
<section id="witnesses" class="proof-section" aria-labelledby="witness-title"><div class="proof-heading"><h2 id="witness-title">Replay an actual campaign witness</h2><span class="eyebrow">SIX PRESELECTED PUBLIC TRACES</span></div><p class="muted">The same public twin pair has an initial retained, trained retained, and trained erased witness. All six use initialization 28001. These witnesses illustrate the cohort and do not replace its 288 trial receipts.</p><div class="witness-choice"><label for="witness-select">Recorded campaign</label><select id="witness-select"></select><p id="witness-outcome"></p></div><ol id="room-rail" class="room-rail" aria-label="Campaign rooms"></ol><div class="replay-grid"><section class="event-panel" aria-labelledby="event-title"><div class="proof-heading"><h3 id="event-title">Recorded event</h3><span id="event-count" class="eyebrow"></span></div><div id="event-metrics" class="event-metrics"></div><p id="event-action" class="event-action"></p><p id="event-feedback" class="event-feedback"></p><p id="event-role" class="muted small"></p><div class="proof-controls"><button id="proof-previous" type="button">Previous</button><button id="proof-play" type="button">Play events</button><button id="proof-next" type="button">Next</button><button id="proof-last" type="button">Final event</button></div><label class="sr-only" for="proof-scrubber">Recorded event number</label><input id="proof-scrubber" type="range" min="0" value="0"><div id="phase-jumps" class="phase-jumps" aria-label="Jump to a room"></div></section><section class="public-state"><h3>Actual public state</h3><p id="notebook-state" class="small muted"></p><pre id="board-state"></pre></section></div><details class="proof-detail"><summary>Inspect the complete public record for this event</summary><p class="muted small">The witness records canonical compare_seal argument order. The submitted input preserves the recorded request order. Both compare the same unordered record pair.</p><pre id="event-public"></pre></details><details class="proof-detail"><summary>Inspect every recorded action and notebook write</summary><div id="event-table" class="proof-table"></div></details><p id="witness-reference" class="reference"></p></section>
<section class="proof-section" aria-labelledby="limits-title"><h2 id="limits-title">What the original ambition still requires</h2><div id="requirements" class="requirements"></div><aside class="beam-note"><h3>Beam evidence remains separate</h3><p>The historical native study has 1,208 HTTP requests across 8 allocated cases, 0 Blackout crossings, and 0 verified escapes. Those failures remain visible in the campaign's Evidence and research view.</p><p>A separate <a href="./probe.html">completed 12-call Beam first-filing probe</a> tests supplied correct, erased, and foil notes. It includes eleven post-run engine witnesses with scripted continuation. That assisted first-filing study leaves this frozen local learning result unchanged. Neither study updates Beam's weights.</p></aside></section>
<section class="proof-section" aria-labelledby="receipts-title"><h2 id="receipts-title">Receipt scope and source commitments</h2><p id="receipt-limit" class="muted"></p><p id="provenance-limit" class="muted"></p><p class="muted">The independent learning audit reports a pass with scope limits. It recomputes recorded updates and probabilities. This viewer's exporter verifies sanitized manifest hashes, cohort structure, paired prefixes, and witness public-state digests. It does not rerun private replay verification.</p><div id="sources" class="source-list"></div><details class="proof-detail"><summary>Inspect all 288 full-start trial summaries</summary><div id="trial-table" class="proof-table"></div></details><p><a href="./proof-receipts.zip" download>Download 173 sanitized learning and transfer receipt files</a></p><p><a href="./proof.json" download="archive-proof.json">Download the full local public-record projection</a></p></section>
</div>
<footer class="page-footer"><span>BORDERLAND 路 RECORDED PUBLIC EVENTS</span><span>LEARNED SELECTION 路 SCRIPTED CAMPAIGN CONTROL</span></footer>
</main>
</body>
</html>