https://www.jevreduce.com/use-cases/ Recent content in Use cases on jevreduce Hugo en Mon, 21 Sep 2026 00:00:00 +0000 https://www.jevreduce.com/use-cases/event-forecasting/ Mon, 21 Sep 2026 00:00:00 +0000 https://www.jevreduce.com/use-cases/event-forecasting/ <h2 id="why-this-shape">Why this shape</h2> <ul> <li><strong>The bundle is the state.</strong> Every item read together, once.</li> <li><strong>The criterion is the question.</strong> Verbatim, date included.</li> <li><strong>The price stays out of the request.</strong> Joined for the reviewer only.</li> <li><strong>Disagreement is the queue.</strong> p(yes) next to price, sorted.</li> </ul> <h2 id="how-it-gets-scored">How it gets scored</h2> <ol> <li>Run weekly on fresh snapshots.</li> <li>Add resolutions as markets settle. Join on <code>market_id</code>.</li> <li>Bucket by p(yes). Read the resolved share per bucket.</li> <li>Replay with another model or wording. Compare the tables.</li> </ol> <p>Nothing here trades.</p> https://www.jevreduce.com/use-cases/model-grading/ Mon, 21 Sep 2026 00:00:00 +0000 https://www.jevreduce.com/use-cases/model-grading/ <h2 id="why-this-shape">Why this shape</h2> <ul> <li><strong>Grade alone, prefer together.</strong> Independent grades, then A against B in one view.</li> <li><strong>Typed, not parsed.</strong> A probability, a rubric position, a failure name, a preference split.</li> <li><strong>Swap the judge.</strong> One field in the manifest. Same snapshot. Two columns.</li> <li><strong>Disagreements are the labelling queue.</strong> Both judges’ raw responses are in the trace.</li> </ul> https://www.jevreduce.com/use-cases/interview-scoring/ Mon, 21 Sep 2026 00:00:00 +0000 https://www.jevreduce.com/use-cases/interview-scoring/ <h2 id="why-this-shape">Why this shape</h2> <ul> <li><strong>One rubric, frozen.</strong> Every candidate scored against the same words.</li> <li><strong>Ordinal scores.</strong> 2.7 means between two levels, and you can see how sure.</li> <li><strong>The evidence turn.</strong> Read one answer, not the transcript.</li> <li><strong>Must-haves only above the bar.</strong> The other 81 keep their scores.</li> </ul> <h2 id="what-it-replaces">What it replaces</h2> <ul> <li>Six people remembering what was said.</li> <li>A 3 that means a 4 to someone else.</li> <li>Not the interviewer. People decide.</li> </ul> https://www.jevreduce.com/use-cases/churn-reasons/ Mon, 21 Sep 2026 00:00:00 +0000 https://www.jevreduce.com/use-cases/churn-reasons/ <h2 id="why-this-shape">Why this shape</h2> <ul> <li><strong>Unit: the customer, not the ticket.</strong> Two angry tickets are one relationship.</li> <li><strong>One request, three questions.</strong> 1,284 tickets, 412 requests. One question at a time: 1,236.</li> <li><strong>Filter before you read more.</strong> History is read for 38 customers, not 412.</li> <li><strong>Distributions, not labels.</strong> Close calls look like close calls.</li> <li><strong>The receipt.</strong> Why is Northstar on the list? Two clicks, not a re-run.</li> </ul> <h2 id="when-the-questions-change">When the questions change</h2> <ol> <li>Edit a question. A new version is saved.</li> <li>Old runs keep the old version.</li> <li>Replay to apply the new wording to old data. New run, parent kept.</li> </ol> https://www.jevreduce.com/use-cases/ticket-triage/ Mon, 21 Sep 2026 00:00:00 +0000 https://www.jevreduce.com/use-cases/ticket-triage/ <h2 id="why-this-shape">Why this shape</h2> <ul> <li><strong>Four decisions per ticket, one request.</strong> Category, priority, engineering, sensitive data.</li> <li><strong>Your priority scale.</strong> P3 to P0, in your words, inside the state.</li> <li><strong>Incidents are a snapshot.</strong> Matched against what was open at 09:00. The trace shows which.</li> <li><strong>Routing in your code.</strong> Decision traced here, action audited there.</li> </ul> https://www.jevreduce.com/use-cases/deals-likely-to-close/ Mon, 21 Sep 2026 00:00:00 +0000 https://www.jevreduce.com/use-cases/deals-likely-to-close/ <h2 id="why-this-shape">Why this shape</h2> <ul> <li><strong>Per call first.</strong> Next step, objection, stage. One call is the natural unit.</li> <li><strong>Per deal second.</strong> All its calls plus the record. A true dependency, so a second stage.</li> <li><strong>Between them, no model calls.</strong> Group, then join.</li> </ul> <h2 id="make-the-probability-earn-its-threshold">Make the probability earn its threshold</h2> <ol> <li>Run the same definition on last quarter’s snapshots.</li> <li>Join closed-won and closed-lost on <code>opportunity_id</code>.</li> <li>Bucket by p(close). Count won and lost per bucket.</li> <li>Read the table. That is what 0.8 means for your team.</li> </ol> <table> <thead> <tr> <th>p(close)</th> <th>deals</th> <th>won</th> <th>share</th> </tr> </thead> <tbody> <tr> <td>0.90 and up</td> <td>your number</td> <td>your number</td> <td>your number</td> </tr> <tr> <td>0.70 to 0.89</td> <td>your number</td> <td>your number</td> <td>your number</td> </tr> <tr> <td>0.50 to 0.69</td> <td>your number</td> <td>your number</td> <td>your number</td> </tr> <tr> <td>below 0.50</td> <td>your number</td> <td>your number</td> <td>your number</td> </tr> </tbody> </table>