<?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[HK Analytical]]></title><description><![CDATA[HK Analytical]]></description><link>https://hkanalytical.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Fri, 20 Nov 2026 16:23:29 GMT</lastBuildDate><atom:link href="https://hkanalytical.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[What ± and k = 2 mean, explained for engineers]]></title><description><![CDATA[A lab report says 9.6 ± 0.4 mg (k = 2). If you work with error bars, latency percentiles or SLOs, you already have most of the mental model. You also have…]]></description><link>https://hkanalytical.hashnode.dev/what-and-k-2-mean-explained-for-engineers</link><guid isPermaLink="true">https://hkanalytical.hashnode.dev/what-and-k-2-mean-explained-for-engineers</guid><category><![CDATA[statistics]]></category><dc:creator><![CDATA[HK Analytical]]></dc:creator><pubDate>Wed, 18 Nov 2026 06:02:06 GMT</pubDate><content:encoded><![CDATA[<p>A lab report says <code>9.6 ± 0.4 mg (k = 2)</code>. If you work with error bars, latency percentiles or SLOs, you already have most of the mental model. You also have some habits that give wrong answers here.</p>
<p>The worked example is peptide testing, because several labs have measured the same batch and the numbers are public.</p>
<h2>A result is an interval</h2>
<p>The Guide to the Expression of Uncertainty in Measurement (GUM, JCGM 100:2008) treats a result as complete only when it carries a statement of uncertainty. So <code>9.6 ± 0.4 mg (k = 2)</code> does not mean "9.6, more or less". It means the lab expects the true content to lie between 9.2 and 10.0 mg, with a level of confidence of about 95%.</p>
<h2>Standard uncertainty and expanded uncertainty</h2>
<p>There are two quantities, and reports often mix them up.</p>
<ul>
<li><strong>Standard uncertainty, u.</strong> One standard deviation for the result. It is not only the scatter of replicates. It combines every known source: repeatability, weighing, dilution, and the value assigned to the reference standard.</li>
<li><strong>Expanded uncertainty, U.</strong> <code>U = k × u</code>, where k is the coverage factor. U is the number after the ±.</li>
</ul>
<p>For independent components in the same unit, the usual combination is the root sum of squares. The numbers below are invented for illustration. They are not test data.</p>
<pre><code class="language-python">import math

# Illustration only: invented values, not measurements.
components_mg = {
    "repeatability": 0.12,       # scatter between replicate preparations
    "reference_standard": 0.15,  # uncertainty of the standard's assigned content
    "weighing_dilution": 0.05,
}

u = math.sqrt(sum(v ** 2 for v in components_mg.values()))
k = 2
U = k * u

print(f"u = {u:.2f} mg, U = {U:.2f} mg")   # u = 0.20 mg, U = 0.40 mg
</code></pre>
<p>More replicates make the repeatability term smaller. They do nothing to the reference standard term. If the content assigned to the standard is 5% too high, every result calculated from it reads about 5% high, and no number of injections shows it.</p>
<h2>Why k = 2, and where "about 95%" comes from</h2>
<p>For a normal distribution, ±1 standard deviation covers about 68% and ±2 covers about 95%. The GUM states that k = 2 gives an interval with a level of confidence of about 95% when the distribution is close to normal. That condition is an assumption. If u comes from only a few replicates, 2 is too small for 95%, and a larger factor from the t-distribution is the correct one.</p>
<p>Two practical results:</p>
<ol>
<li><strong>A ± with no k is incomplete.</strong> It can be one standard deviation, which covers much less. The Eurachem/CITAC guide recommends <code>x ± U</code> with the coverage factor named, and advises against ± for a standard uncertainty.</li>
<li><strong>Digits must agree with the uncertainty.</strong> The GUM says to give an uncertainty to at most two significant digits and to round the result to agree. A result of 10.457 mg with an uncertainty of 0.4 mg is printed as 10.5 ± 0.4 mg.</li>
</ol>
<h2>Repeatability is not agreement between labs</h2>
<p>A small ± from one lab tells you how much that lab's method scatters. It does not tell you what a second lab will report.</p>
<p>A published case shows the size of the difference. In May 2026, Peptide Critic compared reports from five labs on one retatrutide kit labelled 10 mg. Every purity result fell between 99.74% and 99.93%, which is agreement within about 0.2 points. The mg results ran from 10.18 to 12.80 mg, a spread of about 26%. Each lab's own results were within about 4% of each other.</p>
<p>So each method looked repeatable, and the methods disagreed. This is one published comparison. It does not prove that any lab is wrong.</p>
<p>The reason is structural. Purity is a ratio of peak areas inside one run, so it needs no standard of known content. Content in mg is an absolute amount. It depends on the reference standard and on the basis of the mg: free base, the salt, or the whole powder. A lab's ± covers the scatter inside its own method. It does not cover a different basis or a different standard at another lab. The guide on <a href="https://hkanalytical.com/guides/why-labs-disagree">why labs disagree</a> lists the causes with sources.</p>
<p>The systems analogy: two servers with stable clocks that are not synchronised. Each has low jitter. The offset between them is a different quantity, and you cannot estimate it from the timestamps of one server.</p>
<h2>Comparing a result with a limit or a label</h2>
<p>The uncertainty matters most when the interval straddles the line. This illustration comes from the <a href="https://hkanalytical.com/guides/peptide-test-accuracy">peptide test accuracy guide</a>. It is not real data. The specification is 90 to 110% of a 10 mg label, and the result is 9.1 ± 0.4 mg (k = 2).</p>
<p>ILAC G8 describes two decision rules. With simple acceptance, a result inside the limit passes. With guarded acceptance, the limit moves inward by a guard band, often equal to U.</p>
<pre><code class="language-python">def verdict(x, U, lower_limit, rule="simple"):
    guard = U if rule == "guarded" else 0.0
    return "pass" if x &gt;= lower_limit + guard else "fail"

verdict(9.1, 0.4, 9.0)              # 'pass'
verdict(9.1, 0.4, 9.0, "guarded")   # 'fail': needs 9.4 mg
</code></pre>
<p>The same measurement gives two verdicts. The interval of 8.7 to 9.5 mg straddles the limit, and a bare "PASS" hides that. ISO/IEC 17025 asks a lab that states pass or fail to apply a documented decision rule that takes uncertainty into account. A guard band is the same idea as an alert threshold that fires before the SLO is breached.</p>
<p>Two more comparisons use the same logic:</p>
<ul>
<li><strong>Against a label claim.</strong> For a result of 4.8 ± 0.3 mg (k = 2) and a 5 mg label, the claim is inside 4.5 to 5.1 mg. The result does not show an underfill.</li>
<li><strong>Against a second lab.</strong> Two labs that each work within ±10% can report 9.0 and 11.0 mg for a vial that holds 10 mg. A gap smaller than the two margins combined is not evidence that either lab made an error.</li>
</ul>
<h2>Where the analogy stops</h2>
<ul>
<li><strong>A confidence interval from sampling becomes smaller with n.</strong> Measurement uncertainty includes components that come from certificates, specifications and judgement. More data does not make them smaller.</li>
<li><strong>"About 95%" is not a guarantee of coverage.</strong> It is only as good as the uncertainty budget and the normality assumption. An effect that is not in the budget is not in the ±.</li>
<li><strong>With an SLO, you own the definition of the metric.</strong> Here, two reports can define the measured quantity differently. The first source of uncertainty in the GUM's list is an incomplete definition of what is measured. "10 mg" of what?</li>
</ul>
<h2>What to look for on a report</h2>
<p>Look for the result with its ± and its k, the number of preparations, injections and vials behind it, the reference standard, the basis of the mg, and the decision rule behind each verdict.</p>
<p>HK Analytical is a peptide testing lab. The full guide, with sources, is <a href="https://hkanalytical.com/guides/peptide-test-accuracy">Peptide test accuracy: what ± and k = 2 on a COA tell you</a>.</p>
]]></content:encoded></item></channel></rss>