<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: BalKrishna Mehta</title>
    <description>The latest articles on DEV Community by BalKrishna Mehta (@balkrishna_mehta_5861b8b9).</description>
    <link>https://dev.to/balkrishna_mehta_5861b8b9</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4173902%2Fc15dc424-b5fe-44a8-bbe2-f965f627f706.jpg</url>
      <title>DEV Community: BalKrishna Mehta</title>
      <link>https://dev.to/balkrishna_mehta_5861b8b9</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/balkrishna_mehta_5861b8b9"/>
    <language>en</language>
    <item>
      <title>The rewritten prompt lost on the lines it had not seen</title>
      <dc:creator>BalKrishna Mehta</dc:creator>
      <pubDate>Sat, 17 Oct 2026 09:22:06 +0000</pubDate>
      <link>https://dev.to/balkrishna_mehta_5861b8b9/the-rewritten-prompt-lost-on-the-lines-it-had-not-seen-5gkm</link>
      <guid>https://dev.to/balkrishna_mehta_5861b8b9/the-rewritten-prompt-lost-on-the-lines-it-had-not-seen-5gkm</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;h1&gt;
  
  
  devchallenge #hf26challenge #gemma
&lt;/h1&gt;

&lt;h2&gt;
  
  
  What I built
&lt;/h2&gt;

&lt;p&gt;A bookkeeper pastes a supplier line such as &lt;code&gt;PARLE G 56G 12PK @ 48.00 INR&lt;/code&gt;. Holdout asks Gemma for six fields: brand, name, size, pack, unit price, and currency.&lt;/p&gt;

&lt;p&gt;The paid part is the second measurement. A new prompt can look better on the lines used to write it. Holdout freezes that prompt, then scores it on 8 lines that were kept out of the fit. Best-of-N and a sequential fix get the same 3 model calls on each of those lines. The report prints the loss when the rewrite is worse.&lt;/p&gt;

&lt;p&gt;The lines in the repository are synthetic. They are labeled that way. I am not describing a shop I visited.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://holdout-zeta.vercel.app" rel="noopener noreferrer"&gt;https://holdout-zeta.vercel.app&lt;/a&gt;&lt;/p&gt;


&lt;div class="crayons-card c-embed text-styles text-styles--secondary"&gt;
    &lt;div class="c-embed__content"&gt;
      &lt;div class="c-embed__body flex items-center justify-between"&gt;
        &lt;a href="https://holdout-zeta.vercel.app" rel="noopener noreferrer" class="c-link fw-bold flex items-center"&gt;
          &lt;span class="mr-2"&gt;holdout-zeta.vercel.app&lt;/span&gt;
          

        &lt;/a&gt;
      &lt;/div&gt;
    &lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The page is the saved run. The model is &lt;code&gt;gemma4:e4b&lt;/code&gt;, on this laptop, through Ollama. The machine has no NVIDIA GPU. The calls went to &lt;code&gt;127.0.0.1&lt;/code&gt;.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Method&lt;/th&gt;
&lt;th&gt;Held-out field accuracy&lt;/th&gt;
&lt;th&gt;Held-out calls&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;best of N&lt;/td&gt;
&lt;td&gt;0.875&lt;/td&gt;
&lt;td&gt;24&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;refine&lt;/td&gt;
&lt;td&gt;0.8958&lt;/td&gt;
&lt;td&gt;24&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;evolved&lt;/td&gt;
&lt;td&gt;0.5417&lt;/td&gt;
&lt;td&gt;24&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The evolved prompt lost to best-of-N. The fit used 17 other calls. Those calls are stored on the page and are not added to the evolved row.&lt;/p&gt;

&lt;p&gt;On &lt;code&gt;KITKAT 37G 24PK @ 720 INR&lt;/code&gt;, best-of-N returned a unit price of 720. The evolved prompt returned 30. Refine returned 720 and filled the product name. The page shows that row, plus Horlicks and Pepsi.&lt;/p&gt;

&lt;p&gt;The paste box does not send the line to a server. A live run uses Ollama on the same computer. A reader who already has a Gemma server can set &lt;code&gt;HOLDOUT_BASE_URL&lt;/code&gt;, &lt;code&gt;HOLDOUT_API_KEY&lt;/code&gt;, and &lt;code&gt;HOLDOUT_MODEL&lt;/code&gt; in &lt;code&gt;.env&lt;/code&gt;. The key is not committed, and &lt;code&gt;metrics.json&lt;/code&gt; stores the model name, not the server.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://github.com/BalkrishnaMehta/holdout" rel="noopener noreferrer"&gt;https://github.com/BalkrishnaMehta/holdout&lt;/a&gt;&lt;/p&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/BalkrishnaMehta" rel="noopener noreferrer"&gt;
        BalkrishnaMehta
      &lt;/a&gt; / &lt;a href="https://github.com/BalkrishnaMehta/holdout" rel="noopener noreferrer"&gt;
        holdout
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      Held-out score for a local Gemma supplier-line harness
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Holdout&lt;/h1&gt;
&lt;/div&gt;
&lt;p&gt;Holdout fits a local Gemma harness to messy supplier lines, then scores that harness on lines it did not use for fitting. Best-of-N and sequential refine get the same number of model calls on the held-out lines. If the evolved harness loses, the report says so.&lt;/p&gt;
&lt;p&gt;The lines in &lt;code&gt;data/&lt;/code&gt; are synthetic. By default the model is Gemma through Ollama on this computer. A hosted OpenAI-compatible Gemma is optional: put the base URL, model name, and key in &lt;code&gt;.env&lt;/code&gt; (see &lt;code&gt;.env.example&lt;/code&gt;). The report names the model that ran. It does not store the server or the key.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Install&lt;/h2&gt;
&lt;/div&gt;
&lt;div class="highlight highlight-source-powershell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;conda create &lt;span class="pl-k"&gt;-&lt;/span&gt;n holdout python&lt;span class="pl-k"&gt;=&lt;/span&gt;&lt;span class="pl-c1"&gt;3.11&lt;/span&gt; pip &lt;span class="pl-k"&gt;-&lt;/span&gt;y
conda activate holdout
pip install &lt;span class="pl-k"&gt;-&lt;/span&gt;e &lt;span class="pl-s"&gt;&lt;span class="pl-pds"&gt;"&lt;/span&gt;.[dev]&lt;span class="pl-pds"&gt;"&lt;/span&gt;&lt;/span&gt;
ollama pull gemma4:e4b&lt;/pre&gt;

&lt;/div&gt;
&lt;p&gt;&lt;code&gt;gemma3:4b&lt;/code&gt; is used only when E4B cannot return one JSON object and &lt;code&gt;.env&lt;/code&gt; does not name another model.&lt;/p&gt;
&lt;div class="highlight highlight-source-powershell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;copy .env.example .env&lt;/pre&gt;

&lt;/div&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;Run&lt;/h2&gt;
&lt;/div&gt;
&lt;div class="highlight highlight-source-powershell notranslate position-relative overflow-auto js-code-highlight"&gt;
&lt;pre&gt;&lt;span class="pl-smi"&gt;$&lt;span class="pl-c1"&gt;env:&lt;/span&gt;OLLAMA_HOST&lt;/span&gt;&lt;span class="pl-k"&gt;=&lt;/span&gt;&lt;/pre&gt;…
&lt;/div&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/BalkrishnaMehta/holdout" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The license is MIT. &lt;code&gt;data/fit.jsonl&lt;/code&gt; has 16 lines. &lt;code&gt;data/heldout.jsonl&lt;/code&gt; has 8. No fit id appears in the held-out file.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I built it
&lt;/h2&gt;

&lt;p&gt;Wang et al. compare a searched prompt with extra samples at a matched call budget, and they score the result on lines the search did not use. The paper is &lt;a href="https://arxiv.org/html/2607.12227" rel="noopener noreferrer"&gt;arXiv 2607.12227&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Holdout keeps that comparison small enough for a CPU. One rewrite. The fit pass counts missed fields with the labels. One more call asks for four short rules. After the text is generated, the same labels score the fields. They are not copied into the extraction prompt, and they are not used to choose a held-out line. &lt;code&gt;pick_first_valid&lt;/code&gt; keeps the first complete JSON object. It does not look at the label.&lt;/p&gt;

&lt;p&gt;The rewrite from this run talked about people's names, titles such as Mr. and Dr., and currency written as LaTeX. It had been asked for rules about name, currency, and brand, which were the fields missed most often on the fit lines. Held-out field accuracy went from 0.875 to 0.5417. I left that loss in &lt;code&gt;data/metrics.json&lt;/code&gt;. I did not edit the rules afterward to produce a win.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;gemma3:4b&lt;/code&gt; is installed for the case where &lt;code&gt;gemma4:e4b&lt;/code&gt; cannot return one JSON object. This run returned JSON from E4B, so the fallback stayed unused.&lt;/p&gt;

&lt;p&gt;Two DEV posts sat next to this design. &lt;a href="https://dev.to/jangwook_kim_e31e7291ad98/ollama-structured-outputs-in-practice-getting-type-safe-json-from-local-llms-with-pydantic-m38"&gt;Ollama Structured Outputs in Practice&lt;/a&gt; shows a typed JSON parse from a local model. &lt;a href="https://dev.to/robertodelacamara/variance-testing-flipped-my-ollama-benchmark-ranking-3pod"&gt;Variance testing flipped my Ollama benchmark ranking&lt;/a&gt; shows a single pass changing which local model looks best. Holdout still takes the first JSON object in the reply and scores the six fields by exact match. A schema helper would not answer the held-out question.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why open innovation matters
&lt;/h2&gt;

&lt;p&gt;Gemma 4 E4B is an open-weight model. This laptop has 16 GB of RAM and an Intel CPU. The 89 calls behind the table, 17 on the fit and 72 on the held-out lines, stayed on that machine.&lt;/p&gt;

&lt;p&gt;A closed extraction API would bill each of those calls. A bookkeeper running the comparison on a stack of supplier lines would pay that bill every evening. A local model makes the check cheap enough to run on lines that should not leave the shop.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize category
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Best Use of Gemma.&lt;/strong&gt; The saved model name is &lt;code&gt;gemma4:e4b&lt;/code&gt;, through Ollama. This post enters that category only.&lt;/p&gt;

&lt;h2&gt;
  
  
  First user
&lt;/h2&gt;

&lt;p&gt;No one else has used Holdout. I have not been paid. This section stays empty until someone who is not me runs it on lines I did not write.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
    </item>
    <item>
      <title>Desire paths are pull requests the city never merged</title>
      <dc:creator>BalKrishna Mehta</dc:creator>
      <pubDate>Sun, 11 Oct 2026 19:47:55 +0000</pubDate>
      <link>https://dev.to/balkrishna_mehta_5861b8b9/desire-paths-are-pull-requests-the-city-never-merged-doh</link>
      <guid>https://dev.to/balkrishna_mehta_5861b8b9/desire-paths-are-pull-requests-the-city-never-merged-doh</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week1-2026-10-05"&gt;Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

&lt;h1&gt;
  
  
  devchallenge #hf26challenge #gemma #tabpfn
&lt;/h1&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;In Utrecht a desire path is an &lt;em&gt;olifantenpaadje&lt;/em&gt;: grass worn off because people cut the corner the pavement took. The official walk is already on the map. The shortcut is the line the map never accepted.&lt;/p&gt;

&lt;p&gt;Desire Lines looks for those cuts inside one box, 5.09°E, 52.075°N to 5.155°E, 52.125°N. Geometry comes from a Geofabrik OpenStreetMap extract. The pixels come from PDOK, the Dutch 25 cm orthophoto. A local Gemma model looks at each aerial chip and answers yes, no, or unclear. The pipeline writes a map, a GPX file, a printable field card, and mapper tasks. I have not walked the route. Nothing is written to OpenStreetMap.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://desire-lines.vercel.app" rel="noopener noreferrer"&gt;https://desire-lines.vercel.app&lt;/a&gt;&lt;/p&gt;


&lt;div class="crayons-card c-embed text-styles text-styles--secondary"&gt;
    &lt;div class="c-embed__content"&gt;
      &lt;div class="c-embed__body flex items-center justify-between"&gt;
        &lt;a href="https://desire-lines.vercel.app" rel="noopener noreferrer" class="c-link fw-bold flex items-center"&gt;
          &lt;span class="mr-2"&gt;desire-lines.vercel.app&lt;/span&gt;
          

        &lt;/a&gt;
      &lt;/div&gt;
    &lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;The map is central Utrecht. It draws 40 candidates. Click a line and a card opens with the length, the seconds that detour would save, the wear number, and a verdict.&lt;/p&gt;

&lt;p&gt;The field tab links a GPX file, an A6 card, and mapper tasks. The card lists eight stops. The boxes for worn, faint, gone, and paved are empty.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://github.com/BalkrishnaMehta/desire-lines/" rel="noopener noreferrer"&gt;https://github.com/BalkrishnaMehta/desire-lines/&lt;/a&gt;&lt;/p&gt;


&lt;div class="ltag-github-readme-tag"&gt;
  &lt;div class="readme-overview"&gt;
    &lt;h2&gt;
      &lt;img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"&gt;
      &lt;a href="https://github.com/BalkrishnaMehta" rel="noopener noreferrer"&gt;
        BalkrishnaMehta
      &lt;/a&gt; / &lt;a href="https://github.com/BalkrishnaMehta/desire-lines" rel="noopener noreferrer"&gt;
        desire-lines
      &lt;/a&gt;
    &lt;/h2&gt;
    &lt;h3&gt;
      
    &lt;/h3&gt;
  &lt;/div&gt;
  &lt;div class="ltag-github-body"&gt;
    
&lt;div id="readme" class="md"&gt;&lt;div class="markdown-heading"&gt;
&lt;h1 class="heading-element"&gt;Desire Lines&lt;/h1&gt;
&lt;/div&gt;
&lt;blockquote&gt;
&lt;p&gt;Pedestrian pull requests the city never merged.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;A desire path is a shortcut worn into grass by repeated footsteps. Desire Lines finds
places where the official pedestrian network takes a detour, checks open aerial imagery
for a continuous loss of vegetation, ranks candidates (TabPFN when its license gate
allows, otherwise sklearn and a transparent heuristic), and optionally asks a local
Gemma vision model for a conservative second opinion.&lt;/p&gt;
&lt;p&gt;This project was built for the
&lt;a href="https://dev.to/challenges/hacktoberfest-week1-2026-10-05" rel="nofollow"&gt;Hacktoberfest Open-Source AI Challenge: Touch Grass&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;No field walk has been done. The site can export a GPX and a printable card as an
invitation for a human mapper. The pipeline never writes to OpenStreetMap.&lt;/p&gt;
&lt;div class="markdown-heading"&gt;
&lt;h2 class="heading-element"&gt;What it produces&lt;/h2&gt;
&lt;/div&gt;
&lt;ul&gt;
&lt;li&gt;a map where each candidate looks like a pull request;&lt;/li&gt;
&lt;li&gt;a spatially cross-validated comparison of TabPFN (license-gated) and conventional baselines;&lt;/li&gt;
&lt;li&gt;optional GPX / A6 card / MapRoulette GeoJSON after &lt;code&gt;desire-lines walk&lt;/code&gt; — not a completed…&lt;/li&gt;
&lt;/ul&gt;&lt;/div&gt;
  &lt;/div&gt;
  &lt;div class="gh-btn-container"&gt;&lt;a class="gh-btn" href="https://github.com/BalkrishnaMehta/desire-lines" rel="noopener noreferrer"&gt;View on GitHub&lt;/a&gt;&lt;/div&gt;
&lt;/div&gt;


&lt;p&gt;Source code is MIT. OpenStreetMap data remains ODbL. Aerial imagery is Beeldmateriaal.nl / PDOK, CC BY 4.0. TabPFN weights stay under Prior Labs' license. Gemma is Google's open-weight model. The local install was &lt;code&gt;ollama pull gemma3:4b&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;Mapped &lt;code&gt;informal=yes&lt;/code&gt; ways are removed from the walk graph before any detour is measured. A pair becomes a candidate when the straight line is 20–400 m, at least 80% of it lies in mapped green space, it crosses no mapped barrier, the official walk is at least 1.25× longer, and the line saves at least 15 m.&lt;/p&gt;

&lt;p&gt;This run found 81 candidates. Wear, scoring, and the map use the top 40.&lt;/p&gt;

&lt;p&gt;For each line the app computes Excess Green, &lt;code&gt;2G - R - B&lt;/code&gt;, inside a 1.5 m corridor and compares it with two parallel control strips. &lt;code&gt;wear&lt;/code&gt; is the greenness drop, in standard deviations of those strips. &lt;code&gt;continuity&lt;/code&gt; is the share of corridor pixels under the local vegetation threshold. Shade, dry grass, trees, and the age of the photo can move that number.&lt;/p&gt;

&lt;p&gt;A line is a positive when its Hausdorff distance to a mapped &lt;code&gt;informal=yes&lt;/code&gt; way is at most 12 m and the angles agree within 20°. Unmapped desire paths are labeled negative. Folds are 1 km blocks, five of them. Candidates in the same block stay in the same fold.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;ROC-AUC&lt;/th&gt;
&lt;th&gt;PR-AUC&lt;/th&gt;
&lt;th&gt;Precision at 10&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;logistic regression&lt;/td&gt;
&lt;td&gt;0.9771&lt;/td&gt;
&lt;td&gt;0.8762&lt;/td&gt;
&lt;td&gt;0.5&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;histogram gradient boosting&lt;/td&gt;
&lt;td&gt;0.5&lt;/td&gt;
&lt;td&gt;0.125&lt;/td&gt;
&lt;td&gt;0.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;detour + wear heuristic&lt;/td&gt;
&lt;td&gt;0.4029&lt;/td&gt;
&lt;td&gt;0.1176&lt;/td&gt;
&lt;td&gt;0.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;TabPFN&lt;/td&gt;
&lt;td&gt;no token&lt;/td&gt;
&lt;td&gt;n/a&lt;/td&gt;
&lt;td&gt;n/a&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Logistic regression is the measured model. Samples: 40. Mapped positives: 5. TabPFN returned no score. This machine had no &lt;code&gt;TABPFN_TOKEN&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;utrecht-080&lt;/code&gt; ranks first. It is a mapped informal way, 27.403 m, wear -0.392, continuity 0.204, score 0.9927. The corridor is greener than the grass beside it.&lt;/p&gt;

&lt;p&gt;The five positives are stored with a filled-in detour. The code writes detour ratio 1.8, metres saved as 0.8 × length, and seconds as that distance at 1.3 m/s. The 0.9771 ROC-AUC is agreement with those labels.&lt;/p&gt;

&lt;p&gt;Gemma 3 4B reviewed the top 8 chips through Ollama at &lt;code&gt;http://127.0.0.1:11434&lt;/code&gt;. The laptop is a CPU-only Windows machine with no NVIDIA GPU. The prompt asks whether a narrow worn track is visible under the red line. All 8 verdicts were &lt;code&gt;unclear&lt;/code&gt;. The other 32 cards use a fallback that says yes only when wear is at least 1 and continuity is at least 0.55.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;PDOK publishes the national 25 cm orthophoto under CC BY 4.0, layer &lt;code&gt;Actueel_ortho25&lt;/code&gt;. That license allows a published measurement from the pixels, so the corridor score and the chips can sit in the repository.&lt;/p&gt;

&lt;p&gt;The chips are aerial frames of Utrecht parks. Gemma ran on the laptop. The request went to &lt;code&gt;127.0.0.1&lt;/code&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize categories
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Best Use of Gemma.&lt;/strong&gt; Local Ollama, model &lt;code&gt;gemma3:4b&lt;/code&gt;, verdicts &lt;code&gt;yes&lt;/code&gt;, &lt;code&gt;no&lt;/code&gt;, or &lt;code&gt;unclear&lt;/code&gt;. This run saved 8 Gemma verdicts, all &lt;code&gt;unclear&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Best Use of TabPFN.&lt;/strong&gt; The scorer calls &lt;code&gt;TabPFNClassifier&lt;/code&gt; on the same spatial folds as logistic regression and histogram gradient boosting. This run had no token and no TabPFN ROC-AUC.&lt;/p&gt;

&lt;h2&gt;
  
  
  Honest limits
&lt;/h2&gt;

&lt;p&gt;I have not walked the eight stops. The worn, faint, gone, and paved boxes are unticked.&lt;/p&gt;

&lt;p&gt;The task file says to survey a line before mapping it. No stage calls an OpenStreetMap write API.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>gemma</category>
      <category>tabpfn</category>
    </item>
  </channel>
</rss>
