<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[Xplormity Labs]]></title><description><![CDATA[Xplormity Labs]]></description><link>https://xplormitylabs.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Mon, 16 Nov 2026 20:42:52 GMT</lastBuildDate><atom:link href="https://xplormitylabs.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[A Search Console handoff needs an observation column]]></title><description><![CDATA[A traffic report is easier to review when the next person can tell which sentences came from the data and which came from an interpretation. That…]]></description><link>https://xplormitylabs.hashnode.dev/a-search-console-handoff-needs-an-observation-column</link><guid isPermaLink="true">https://xplormitylabs.hashnode.dev/a-search-console-handoff-needs-an-observation-column</guid><dc:creator><![CDATA[Xplormity]]></dc:creator><pubDate>Sat, 14 Nov 2026 11:45:51 GMT</pubDate><content:encoded><![CDATA[<p>A traffic report is easier to review when the next person can tell which sentences came from the data and which came from an interpretation.</p>
<p>That distinction sounds small. It changes what a teammate can actually verify.</p>
<p>Consider this fictional handoff:</p>
<blockquote>
<p>The documentation redesign hurt search traffic. We should undo it.</p>
</blockquote>
<p>It contains a cause and a proposed action, but no reproducible observation. The reviewer cannot tell which property, pages, dates or search type the sentence describes. A plausible explanation has arrived before its supporting record.</p>
<p>A more useful handoff keeps four fields separate.</p>
<h2>1. Observation</h2>
<p>Record what the result actually contains. State the exact Search Console property, reporting dates, search type and filters. Preserve the requested dimensions and the units of any metric.</p>
<p>For a fictional example, write: “The returned rows show fewer clicks for the selected documentation URLs in the second date range.” Do not upgrade that sentence to “the redesign caused the loss.” The first describes the result. The second requires more evidence.</p>
<p>Keep the original response separately from a presentation table. If the table groups pages or rounds values, say so. A reader should be able to follow the transformation rather than trust a screenshot.</p>
<h2>2. Unknown</h2>
<p>List what the result cannot settle. Were the two periods comparable? Did the selected pages change? Is the result a query-level breakdown or a property total? Was a filter applied before the export?</p>
<p>An absent query row is not proof that nobody searched for that phrase. Search Console has privacy and reporting limitations, so a missing row needs different wording from an observed zero.</p>
<p>This field prevents a report from becoming more confident as it moves between tools.</p>
<h2>3. Hypothesis</h2>
<p>Give explanations their own label. A content change, seasonal demand, reporting delay or a different URL distribution may deserve investigation. None becomes a finding simply because it appears beside a metric.</p>
<p>A good hypothesis also names the evidence that would weaken it. If the decline happened before the deployment, the deployment is a poor explanation for its start. If unchanged pages show a similar pattern, a page-specific explanation needs another check.</p>
<p>Those are proposed checks, not claims about an actual site.</p>
<h2>4. Next check</h2>
<p>Choose a bounded action another person can repeat. Compare the same exact URL set and reporting settings. Inspect the specific affected URLs separately when the question concerns indexing. Record the deployment timestamp before comparing it with a reporting window.</p>
<p>Do not turn a performance export into an indexing verdict. They answer different questions. A published page, an accessible page and an indexed page are also different observations.</p>
<h2>Redact before sharing</h2>
<p>Public examples should use clearly fictional data. An internal handoff may include private property names, unpublished paths, query details or customer information. Review those fields before pasting them into a public issue or a shared AI conversation. Never include credentials or authentication files.</p>
<p>The final handoff can be short: observation, unknown, hypothesis, next check. Its value is that a teammate can disagree with the interpretation without losing the evidence.</p>
<p>Prepared with AI assistance. This is a reporting-method example, not a measured traffic study or an authenticated product demonstration.</p>
]]></content:encoded></item><item><title><![CDATA[Seven speech models, one H100, no clear winner on news broadcasts]]></title><description><![CDATA[The Open ASR Leaderboard has a clear #1 this month: a 30B mixture-of-experts model. Below it sit models a tenth its size. We wanted to know if that order…]]></description><link>https://xplormitylabs.hashnode.dev/seven-speech-models-one-h100-no-clear-winner-on-news-broadcasts</link><guid isPermaLink="true">https://xplormitylabs.hashnode.dev/seven-speech-models-one-h100-no-clear-winner-on-news-broadcasts</guid><category><![CDATA[Speech Recognition]]></category><category><![CDATA[machine learning models]]></category><dc:creator><![CDATA[Xplormity]]></dc:creator><pubDate>Fri, 13 Nov 2026 08:20:59 GMT</pubDate><content:encoded><![CDATA[<p>The Open ASR Leaderboard has a clear #1 this month: a 30B mixture-of-experts model. Below it sit models a tenth its size. We wanted to know if that order holds on real news.</p>
<p>We ran seven speech models on 31 broadcasts (2.6 hours): studio anchors, field reporters, press conferences and crowd noise. One rented H100 80GB, 72 minutes, roughly $3 to $5.</p>
<h2>Results</h2>
<p>Word error rate, lower is better:</p>
<ul>
<li>Parakeet Unified EN 0.6B: 7.34</li>
<li>Qwen3-ASR 1.7B: 7.58</li>
<li>Whisper large-v3-turbo: 7.64</li>
<li>Parakeet TDT 0.6B v3: 7.97</li>
<li>Nemotron-3.5-Transcribe 30B-A4B: 8.00</li>
<li>Granite Speech 5.0 2B: 8.02</li>
<li>freetranscribe parakeet-ft-v1: 8.28</li>
</ul>
<h2>What we found</h2>
<p>The leaderboard order does not hold on news audio. All seven models land within 0.94 WER points of each other, and a paired test (Holm-corrected, smallest p = 0.65) cannot tell any two apart.</p>
<p>What does differ is speed: 23x between the fastest and the slowest.</p>
<p>So for news broadcasts, pick on speed and cost, not on leaderboard rank. A 0.6B model tied the 30B one.</p>
<h2>About this study</h2>
<p>This is stage 2 of NewsShorts-Bench, which turns long TV news bulletins into vertical shorts without cutting out the context that changes the meaning. Every number traces to raw results.</p>
<p>Full study with confidence intervals: <a href="https://xplormity.com/research/asr-news-broadcasts-2026-10">https://xplormity.com/research/asr-news-broadcasts-2026-10</a></p>
]]></content:encoded></item></channel></rss>