<?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[AutoResearch]]></title><description><![CDATA[Research planning and evidence-checking resources from AutoResearch.]]></description><link>https://autoresearch-me.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Wed, 18 Nov 2026 15:09:37 GMT</lastBuildDate><atom:link href="https://autoresearch-me.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[From a research direction to a testable hypothesis: a five-line template]]></title><description><![CDATA[Published by AutoResearch, the team behind autoresearch.me. This is a product-side article, not an independent review. A research direction is not yet…]]></description><link>https://autoresearch-me.hashnode.dev/from-a-research-direction-to-a-testable-hypothesis-a-five-line-template</link><guid isPermaLink="true">https://autoresearch-me.hashnode.dev/from-a-research-direction-to-a-testable-hypothesis-a-five-line-template</guid><category><![CDATA[Machine Learning]]></category><category><![CDATA[research]]></category><category><![CDATA[Artificial Intelligence]]></category><dc:creator><![CDATA[AutoResearch]]></dc:creator><pubDate>Mon, 16 Nov 2026 01:33:40 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6aef2476d62f923148a0ee55/a1d5f754-2460-4032-a846-4e956ec4492b.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><em>Published by AutoResearch, the team behind <a href="https://autoresearch.me">autoresearch.me</a>. This is a product-side article, not an independent review.</em></p>
<p>A research direction is not yet something you can test. "Use transformers for battery health" or "make fault diagnosis robust to new machines" are good starting points, but neither tells you what to train, what to measure, or what result would change your mind. Most wasted GPU time we see in research projects starts here: the direction is turned straight into code, and the hypothesis is reconstructed afterwards from whatever the runs produced.</p>
<p>Here is a short template that forces the translation step to happen first. It fits in five lines, and each line should be answerable before any training script exists.</p>
<h2>The five lines</h2>
<p><strong>1. Claim.</strong> One sentence, in the form "Under condition C, method M changes outcome Y compared with baseline B."
If you cannot name B, you do not yet have a hypothesis; you have an exploration task. That is fine, but label it as such.</p>
<p><strong>2. Unit of evidence.</strong> What one observation is: a sample, a time window, a recording session, a subject, a machine, a battery cell. This decides how you split the data and how many independent observations you really have. Splitting windows from the same battery cell across train and test usually measures memorisation, not generalisation.</p>
<p><strong>3. Primary metric and direction.</strong> One metric, its unit, how it is aggregated (per window, per cell, macro over classes), and whether higher or lower is better. Secondary metrics are allowed, but they are listed as secondary before the first run.</p>
<p><strong>4. Decision rule.</strong> What result would count as support, what would count as contradiction, and what would be inconclusive. For example: "Support if the mean error across held-out cells is lower than the baseline on at least 4 of 5 seeds; inconclusive if the seed spread overlaps." Writing this first is what stops a post-hoc story.</p>
<p><strong>5. Smallest falsifying experiment.</strong> The cheapest run that could already contradict the claim: a reduced dataset, a single seed, a shorter schedule, or the baseline alone. If the smallest run cannot possibly fail, it is a smoke test, not evidence.</p>
<h2>A worked example (illustrative, no results)</h2>
<p>Direction: <em>sequence models for battery state-of-health estimation.</em></p>
<ol>
<li>Claim: on cells never seen in training, a temporal model M estimates state of health with lower error than a gradient-boosted baseline B using the same hand-crafted cycle features.</li>
<li>Unit of evidence: the battery cell. All cycles from one cell go to exactly one split.</li>
<li>Primary metric: mean absolute error in percentage points of state of health, averaged per cell, then across held-out cells; lower is better.</li>
<li>Decision rule: support if M beats B on held-out cells across 5 seeds with non-overlapping spread; contradiction if B is equal or better; otherwise inconclusive and reported as such.</li>
<li>Smallest falsifying experiment: train B and M on a small subset of cells for one seed each. If M cannot beat B even on the training cells, stop and revisit the features before scaling up.</li>
</ol>
<p>None of these lines contains a result. That is the point: they are written before the result exists, so the paper paragraph that eventually reports the outcome can quote them unchanged.</p>
<h2>Why this matters for automated research</h2>
<p>When literature review, idea generation, implementation, GPU experiments and manuscript drafting are connected in one workflow, the hypothesis stops being a private note and becomes the contract that every later stage is checked against. The experiment code should implement exactly the split from line 2. The analysis should report exactly the metric from line 3. The manuscript should state the decision rule from line 4, including when the outcome was inconclusive.</p>
<p>That is the design idea behind AutoResearch: it connects literature exploration, testable ideas, model implementation, GPU experiments and manuscript drafting, and researchers can inspect the results and artifacts behind the work instead of receiving only a finished text. Current beta research areas include computer vision, time series, natural language processing, mechanical prognostics, battery health and structural health monitoring.</p>
<p>AutoResearch is in private beta. Access requires administrator approval; approved members receive 5,000 starter credits, and available compute, data and research scope affect what can be completed. If you want to try the end-to-end workflow, from an idea to a complete paper, you can apply at <a href="https://autoresearch.me">autoresearch.me</a>.</p>
<h2>Copy-paste version</h2>
<pre><code class="language-text">Claim:            Under C, M changes Y compared with B.
Unit of evidence: one observation = ______ ; splits never share a ______.
Primary metric:   ______ (unit, aggregation, higher/lower is better)
Decision rule:    support if ______ ; contradict if ______ ; else inconclusive.
Smallest test:    the cheapest run that could already contradict the claim.
</code></pre>
<p>Fill it in before writing the training script. If a line stays blank, that blank is the next research task.</p>
]]></content:encoded></item><item><title><![CDATA[Window accuracy is not subject accuracy: write down the denominator]]></title><description><![CDATA[AutoResearch ( provides a separate free, no-account research planner alongside its administrator-approved private-beta research system. Approved members…]]></description><link>https://autoresearch-me.hashnode.dev/window-accuracy-is-not-subject-accuracy-write-down-the-denominator</link><guid isPermaLink="true">https://autoresearch-me.hashnode.dev/window-accuracy-is-not-subject-accuracy-write-down-the-denominator</guid><dc:creator><![CDATA[AutoResearch]]></dc:creator><pubDate>Mon, 09 Nov 2026 01:19:25 GMT</pubDate><content:encoded><![CDATA[<p>AutoResearch (<a href="https://autoresearch.me">https://autoresearch.me</a>) provides a separate free, no-account research planner alongside its administrator-approved private-beta research system. Approved members receive 5,000 starter credits; compute, data and research scope limit what can be completed. The public planner helps prepare an experiment plan: it does not run experiments or generate measured results. This note is prepared from the product side with AI assistance, not as an independent product review.</p>
<p>An evaluation can be computed correctly and still answer a different question from the one in your paper. One easy place to find this mismatch is the denominator: are you estimating accuracy across observations, or average accuracy across subjects?</p>
<h2>A small arithmetic example—not experiment results</h2>
<p>The following counts are invented solely to illustrate aggregation. No model was trained or evaluated to obtain them.</p>
<table>
<thead>
<tr>
<th>Subject</th>
<th align="right">Observations</th>
<th align="right">Correct predictions</th>
<th align="right">Accuracy</th>
</tr>
</thead>
<tbody><tr>
<td>A</td>
<td align="right">100</td>
<td align="right">99</td>
<td align="right">99%</td>
</tr>
<tr>
<td>B</td>
<td align="right">10</td>
<td align="right">0</td>
<td align="right">0%</td>
</tr>
</tbody></table>
<p>Pooling the observations gives <strong>99 / 110 = 90%</strong> accuracy. Giving the two subjects equal weight gives <strong>(99% + 0%) / 2 = 49.5%</strong>.</p>
<p>Neither number is an arithmetic mistake. They describe different weighted summaries. Pooled accuracy weights subjects in proportion to their observation counts; macro subject accuracy gives each subject equal weight. Neither is automatically the right estimate of future performance: that depends on the population, sampling design and claim.</p>
<p>If the intended claim concerns a randomly selected observation from this dataset, the pooled number addresses that descriptive question. If the intended claim concerns the average of these two subjects' accuracies, the macro number addresses that question. Neither, on its own, establishes generalization to unseen subjects or sites.</p>
<h2>Separate three decisions</h2>
<p><strong>1. Split unit.</strong> Decide what must stay together when forming training, validation and test sets. If the intended evaluation is on unseen subjects, a subject must not appear across those sets. Aggregating predictions by subject later does not repair leakage from an inappropriate split.</p>
<p><strong>2. Metric unit.</strong> Decide what one prediction or score represents: a window, example, session, subject or another unit. Record the definition before comparing models.</p>
<p><strong>3. Aggregation weight.</strong> Specify how those scores become the headline result. Equal observation weights and equal subject weights are different choices. Record exclusions and missing predictions too, rather than silently changing the denominator.</p>
<p>Correlated windows from one subject are also not independent subjects. A large window count must not be presented as a large subject sample. Any uncertainty analysis needs assumptions suited to the sampling and dependence structure; a headline percentage alone does not provide that analysis.</p>
<h2>A copyable metric–claim contract</h2>
<p>Before running an experiment, fill in this block:</p>
<pre><code class="language-text">Claim to assess:
Target population and intended decision:
Train/validation/test grouping unit:
Prediction or scoring unit:
Primary metric and exact formula:
Aggregation unit and weights:
Missing-output and exclusion rules:
Per-group diagnostics to retain:
Raw prediction file and aggregation script to preserve:
What this evaluation cannot establish:
</code></pre>
<p>For the illustrative table, a descriptive contract could specify: “Report accuracy separately for A and B, plus their equally weighted mean; retain pooled observation accuracy as a secondary summary. These two invented subjects support no empirical generalization claim.”</p>
<p>For an actual study, replace that description with your real design. Do not copy the invented counts into a results section or treat the example as a recommended clinical evaluation protocol.</p>
<h2>Put the contract next to the plan</h2>
<p>The <a href="https://autoresearch.me/en/research-plan/">AutoResearch research planner</a> includes a field for “Primary metric and aggregation rule” and separate fields for data boundaries, baselines, compute budget and evidence. The contract above can be written without creating an account or using the full system.</p>
<p>The useful output is an explicit evaluation definition—not an accuracy promise. Preserve the raw predictions, grouping identifiers and aggregation code so another researcher can reconstruct what each headline number actually counts.</p>
]]></content:encoded></item><item><title><![CDATA[Before a GPU run: a seven-field plan you can audit]]></title><description><![CDATA[AutoResearch at autoresearch.me provides public research-planning resources. Its full research system is an approval-based private beta; approved members…]]></description><link>https://autoresearch-me.hashnode.dev/before-a-gpu-run-a-seven-field-plan-you-can-audit</link><guid isPermaLink="true">https://autoresearch-me.hashnode.dev/before-a-gpu-run-a-seven-field-plan-you-can-audit</guid><dc:creator><![CDATA[AutoResearch]]></dc:creator><pubDate>Sun, 08 Nov 2026 02:01:27 GMT</pubDate><content:encoded><![CDATA[<p>AutoResearch at <a href="https://autoresearch.me">autoresearch.me</a> provides public research-planning resources. Its full research system is an approval-based private beta; approved members receive 5,000 limited starter credits, subject to available compute, data and research scope. This is an AI-assisted guide prepared for AutoResearch, not an independent review.</p>
<p>Before spending compute, record these seven decisions:</p>
<pre><code class="language-text">Question: What falsifiable prediction are you testing?
Data: Which source, version and permissions apply?
Split: Where are the training, validation and test boundaries?
Metric: What definition, units and aggregation will you use?
Comparison: Which baseline and tuning rules make the comparison fair?
Budget: What compute limit and stopping condition apply?
Evidence: Which outputs would support or contradict your claim?
</code></pre>
<p>For time-series or device data, check whether adjacent windows or measurements from the same device cross a boundary that your claim assumes is independent. Fit learned preprocessing on training data, and keep final test observations out of model selection.</p>
<p>After each run, connect the proposed claim to its evidence:</p>
<pre><code class="language-text">Claim: [exact proposed sentence]
Run: [code version, configuration and data split]
Evidence: [metric file and figure source]
Status: [checked / unresolved / unsupported]
Limit: [what this run does not establish]
Next action: [the missing check]
</code></pre>
<p>These are blank templates, not experimental results. A completed run does not establish that a method beats a baseline.</p>
<p>The <a href="https://autoresearch.me/en/research-plan/">free plan builder</a> helps organize these decisions without registration. It does not run experiments or generate measured results. Copy or download your plan before leaving; entries are not saved automatically. The <a href="https://autoresearch.me/en/gpu-experiment-checklist/">GPU checklist</a> supports post-run checks, and the <a href="https://autoresearch.me/en/guides/">guides index</a> collects the public resources.</p>
]]></content:encoded></item></channel></rss>