foundation-validation.html

<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
<meta name="description" content="Inspect fifteen retained causal well-event models, group concentration, false alarms and failed checkpoint portability. Validation only; reserved confirmation remains unopened.">
<title>Causal well-model comparison · validation audit</title>
<style>
:root{color-scheme:light;--ink:#17272c;--muted:#52666d;--paper:#f4f5ef;--line:#d9dfd7;--red:#a72b25;--blue:#165c70;--amber:#815719}
*{box-sizing:border-box}body{margin:0;background:var(--paper);color:var(--ink);font:16px/1.55 system-ui,-apple-system,sans-serif}main{max-width:1200px;margin:auto;padding:32px 28px 60px}h1{font:clamp(2.1rem,4vw,3.5rem)/1.12 Georgia,serif;max-width:900px;margin:14px 0}h2{font:1.7rem/1.2 Georgia,serif;margin:34px 0 12px}h3{font-size:1rem;margin:0 0 12px}p{max-width:910px}a{color:var(--blue);text-underline-offset:3px}a:focus-visible,select:focus-visible,summary:focus-visible{outline:3px solid var(--blue);outline-offset:3px}header{border-bottom:1px solid var(--line);padding-bottom:20px}.eyebrow{color:var(--blue);font-size:.76rem;text-transform:uppercase;letter-spacing:.12em;font-weight:700}.muted,.small{color:var(--muted)}.small{font-size:.83rem}.badges{display:flex;gap:8px;flex-wrap:wrap}.badge{padding:5px 10px;border:1px solid #e9bab1;border-radius:4px;background:#f8e4df;color:var(--red);font-size:.82rem;font-weight:650}.badge.neutral{background:#edf0e9;color:var(--ink);border-color:var(--line)}.notice{padding:18px 20px;border-left:4px solid var(--red);background:#fffaf4;margin:22px 0}.controls{display:flex;gap:18px;align-items:end;flex-wrap:wrap;margin:24px 0}label{display:block;font-size:.85rem;font-weight:650;margin:0 0 7px}select{font:inherit;padding:11px;border:1px solid #83969b;border-radius:4px;color:var(--ink);background:white;max-width:100%}#candidate{width:420px}#cohort{width:290px}#group-filter{width:260px}.meta{font-size:.86rem;color:var(--muted);margin:12px 0;overflow-wrap:anywhere}.cards{display:grid;grid-template-columns:repeat(4,1fr);gap:14px}.card,.panel{background:white;border:1px solid var(--line);padding:18px}.value{font-size:2rem;line-height:1.25;font-weight:650;margin:7px 0;letter-spacing:-.035em}.panels{display:grid;grid-template-columns:1fr 1fr;gap:20px;margin-top:20px}.metric{margin:15px 0}.metric-head{display:flex;justify-content:space-between;gap:8px;font-size:.9rem}.track{height:8px;background:#edf0e9;margin:7px 0}.fill{height:100%;background:var(--blue)}.fill.bad{background:var(--red)}.scroll{overflow-x:auto;background:white;border:1px solid var(--line)}table{border-collapse:collapse;width:100%;font-size:.85rem;min-width:910px}th,td{padding:11px 13px;text-align:left;border-bottom:1px solid var(--line);vertical-align:top}th{font-size:.75rem;letter-spacing:.02em;color:var(--muted);background:#edf0e9}caption{text-align:left;padding:12px;font-size:.82rem;color:var(--muted)}tr.selected{background:#e9f0ef}tr.warn{background:#fff3e9}.links{display:flex;gap:12px 22px;flex-wrap:wrap;margin:18px 0}code{font-size:.8rem;overflow-wrap:anywhere}.digest{overflow-wrap:anywhere}.truth{border-top:1px solid var(--line);margin-top:34px;padding-top:20px;font-size:.86rem;color:var(--muted)}details{margin-top:18px}summary{cursor:pointer;font-weight:650}.error{padding:20px;background:#f8e4df;color:var(--red)}[hidden]{display:none!important}@media(max-width:780px){main{padding:24px 18px}.cards{grid-template-columns:1fr 1fr}.panels{grid-template-columns:1fr}.value{font-size:1.65rem}.controls>div{width:100%}#candidate,#cohort,#group-filter{width:100%}.panel{padding:18px}}
</style>
</head>
<body><main>
<header><a href="index.html">← Original confirmation audit</a><div class="eyebrow">Public 3W research · Separately frozen foundation study</div><h1>Which well-event model earns a second look?</h1><p class="muted">Compare fifteen retained methods on the same causal decision grid. Inspect false alarms, missing-data abstention and individual groups before interpreting a headline score.</p><div class="badges"><span class="badge">Validation only</span><span class="badge">Industrial deployment unqualified</span><span class="badge neutral">8 reserved groups unopened</span></div></header>
<div class="notice"><strong>The apparent winner is concentrated in one group.</strong><br>The selected logistic discovers 13/24 known onsets within 300 seconds; 11 come from one filename group. Full pretrained TSPulse finds no new onset at its selected alarm budget. Both TCNs fail the separate strict Mac checkpoint replay. None is promoted.</div>
<p class="small">These are production-event indications after labelled onset, not pre-onset forecasting, physical well diagnosis or a controller. Public filename groups do not authenticate physical well identity; public-pretraining contamination remains unresolved.</p>
<div class="notice" id="stream-replay-note"><strong>Complete validation replay now verifies a current-only numerical path.</strong><br>Both unchanged TCNs exactly reproduce all 121,325 eligible scores with native batches and eight copies of only the current input. Ordinary one-reading inference fails on 46,577 / 22,294 scores; frozen alarms do not change. Extra compute and every failure remain. No model is promoted; original Mac failure and all eight unopened reserved groups remain. <a href="tcn-stream-validation.html">Inspect six full-stream profiles, eight groups and retained call costs</a></div>
<div class="notice" id="linux-replay-note"><strong>Linux replay: batch size changes the result.</strong><br>In 96 saved-probe controls, 40 runs reproduce all 128 original scores exactly; 56 fail the unchanged tolerance. Larger tested batches pass with MKLDNN enabled. Batch size 1 fails, so one-reading-at-a-time inference remains unqualified.<br>The original Mac audit stays FAILED; that 128-probe study left complete-stream qualification open. No model is promoted. <a href="https://huggingface.co/sankalpsthakur/well-production-event-causal/blob/f183a37536df3ca979ab36329be5244334c4e7ff/tcn-linux-diagnostic/REPORT.md">Inspect all controls, source and included-compute evidence</a></div>
<noscript><div class="error">Enable JavaScript to inspect saved results. The repository report and original archive remain available at <a href="https://huggingface.co/sankalpsthakur/well-production-event-causal/tree/main/foundation-validation">the research module</a>.</div></noscript>
<div id="load-state" role="status">Loading the frozen result audit…</div>
<div id="audit" hidden>
<div class="controls"><div><label for="candidate">Retained method</label><select id="candidate"></select></div><div><label for="cohort">Eligibility denominator</label><select id="cohort"><option value="common">Common support · all 15 methods</option><option value="native">Native support · selected method</option></select></div></div>
<p id="candidate-meta" class="meta" aria-live="polite"></p><p id="cohort-note" class="small"></p>
<section class="cards" aria-label="Selected validation results">
<div class="card"><div>Episode coverage</div><div class="value" id="coverage"></div><div class="small" id="coverage-detail"></div></div>
<div class="card"><div>New onset within 300s</div><div class="value" id="onset"></div><div class="small" id="onset-detail"></div></div>
<div class="card"><div>False alarms / normal day</div><div class="value" id="false-rate"></div><div class="small" id="false-detail"></div></div>
<div class="card"><div>Normal time in alarm</div><div class="value" id="occupation"></div><div class="small" id="occupation-detail"></div></div>
</section>
<div class="panels"><section class="panel"><h3>What the input supports</h3><div id="eligibility-bars"></div><p id="exposure-note" class="small"></p><p class="small">Unknown labels, warmup and unusable inputs are not correct normal predictions. Full-grid and eligible denominators are retained separately.</p></section><section class="panel"><h3>Does the saved checkpoint reproduce?</h3><p id="reload-result"></p><p id="reload-detail" class="small"></p><p class="small">The included Linux job completed. The separate Mac audit remains <strong>FAILED</strong>: thirteen arms pass; both TCNs fail. A reproducible validation score alone does not establish physical transfer.</p><a id="checkpoint-link">Inspect the unchanged checkpoint</a></section></div>
<h2>All fifteen retained methods</h2><p class="small">Thresholds and the primary method were selected on validation. Both alarm budgets must be ≤1 false alarm per eligible normal day and ≤1% eligible normal time in alarm, with ≥0.5 eligible normal days. Disabled alarms remain visible and do not qualify a useful detector. The same frozen thresholds apply to both displayed cohorts.</p>
<div class="scroll"><table><caption id="all-caption"></caption><thead><tr><th>Method</th><th>Episode coverage</th><th>New onset ≤300s</th><th>False alarms/day</th><th>Normal time in alarm</th><th>Selected alarm</th><th>Mac replay</th></tr></thead><tbody id="all-models"></tbody></table></div>
<h2>Where the outcome comes from</h2><p class="small">Eight validation filename groups; rows with no eligible normal exposure have no false-alarm rate. They are not zero-risk evidence.</p><div class="scroll"><table><caption id="group-caption"></caption><thead><tr><th>Filename group</th><th>Coverage</th><th>New onset ≤300s</th><th>Carried-in coverage</th><th>False alarms/day</th><th>Normal time in alarm</th><th>Normal eligibility</th></tr></thead><tbody id="groups"></tbody></table></div>
<div class="panels"><section class="panel"><h3>Group uncertainty</h3><div id="uncertainty"></div><p class="small">1,000 bootstrap replicates resample filename groups, not authenticated independent physical wells. Undefined replicates are explicit. These intervals describe selected validation data.</p></section><section class="panel"><h3>Recorded study boundaries</h3><p>357 training recordings / 16 groups; 33 validation recordings / 8 groups; 113 reserved recordings / 8 groups.</p><p>Eleven compact fits, full pretrained TSPulse with zero pretrained fits, and three conventional references. Causal 512-second history, a 64-second reconstruction tail and decisions every 30 seconds.</p><p class="small">No reserved measurements or predictions were opened. Scalers are training-only; alarm calibration is validation-only. This is an original causal adaptation, not an exact offline-paper replication.</p></section></div>
<h2>Inspect every labelled episode</h2><div class="controls"><div><label for="group-filter">Episode group</label><select id="group-filter"></select></div><div id="episode-count" class="small" aria-live="polite"></div></div>
<div class="scroll"><table><caption>Rows preserve source event IDs, recording identity, input support, carried-in alarms and onset delay. No eligible input means abstention.</caption><thead><tr><th>Source recording</th><th>Event</th><th>Eligible decisions</th><th>Covered</th><th>Carried in</th><th>First coverage delay</th><th>New onset ≤300s</th></tr></thead><tbody id="episodes"></tbody></table></div>
<details><summary>What the 48 TCN diagnostic controls established</summary><p>Two saved checkpoints × six batch sizes × two deterministic settings × two repetitions. Every condition repeats exactly on the observed Mac host. Batch size changes scores; the deterministic flag changes none. Every configuration still fails the unchanged original tolerance.</p><p class="small">Batch 256 duplicates the 128 saved probes; it does not reconstruct the full original validation batches. No configuration was adopted. The exact kernel/architecture contribution remains unisolated. No new fits, retuning or raw/reserved access occurred.</p></details>
<h2>Inspect and reproduce the evidence</h2><div class="links"><a id="report-link">Full result audit</a><a id="archive-link">Original 70-member output archive</a><a id="rebuild-link">No-fitting replay guide</a><a id="protocol-link">Frozen protocol</a><a id="diagnostic-link">TCN diagnostic and failures</a><a id="data-link" href="foundation-validation.json" download>Saved presentation data</a></div>
<p class="small digest">Original output SHA256: <code id="archive-digest"></code><br>Selection SHA256: <code id="selection-digest"></code><br>Artifact content revision: <code id="content-revision"></code></p>
</div>
<footer class="truth">Saved research results; no live telemetry, model inference, plant connection or command path. Physical source/clock/units, independent calibration, target-device deadlines and recovery, site acceptance, field adoption and qualified expert growth remain unverified. MIT authored source; Apache-2.0 upstream TSPulse material retains its notices. Existing 3W source rights and attribution remain in the repository card.</footer>
</main><script src="foundation-validation.js" defer></script></body></html>