--- title: "Spot double counting in agent receipts" slug: spot-double-counting-in-agent-receipts excerpt: "Read category counts without counting shared items twice. Follow repeated names, compare distinct items with the total, and question unseen overlap." quickAnswer: "Copy the category counts and stated total without changing them. Consider whether one item can belong to several categories. Follow names that recur, then count each distinct item once across the receipt. Compare that count with the total the agent wrote. Where names are missing, ask which items the categories share before deciding how far the total reaches." keyTakeaways: - Adding category counts can count a shared item again. - Repeated names let you follow an item across categories. - A total of distinct items counts each item once, regardless of repeated appearances. - Unnamed overlap needs clarification rather than a guessed count. publishedDate: 2026-10-24 dateModified: 2026-10-24 author: Bobby Huang cluster: A role: post status: published faq: - q: "Does a repeated item make both category counts unusable?" a: "An item can belong in both categories. Each category can include it while a total of distinct items includes it once. Read what the total claims to count." - q: "Can I decide the distinct count from category counts alone?" a: "Only if their relationship is stated. When categories can share unnamed items, their counts alone leave the distinct count unsettled." - q: "What if one item appears in three categories?" a: "Keep that item once in your distinct count. Its third appearance does not add another item, just as its second appearance does not." related: - what-is-a-md-file --- Four shortened passages and three retitled passages sound like seven passages. But what if two passages received both edits? The category counts can describe their own groups while the claimed total counts the same passages again. The fictional reader here is Jane Doe; every AI agent receipt below is invented. All passage names and counts are made up. A receipt written as a [Markdown file](what-is-a-md-file.md) still needs this reading: do the categories contain separate items, or do some names recur? ## Hold the written counts beside the names Jane Doe reads these invented lines: > Shortened passages: 4. Named passages: A, B, C, D. > Retitled passages: 3. Named passages: B, D, E. > Total distinct passages: 7. Copy four, three and seven into your own account before working through the names. Keep the labels with those numbers. Four belongs to shortened passages; three belongs to retitled passages; seven is the agent's stated total of distinct passages. That first copy preserves the claim you are examining. Quietly replacing seven with another number would hide what the receipt originally said. You can write your reading beside it without changing the agent's words. Then consider the relationship between the categories. Shortening a passage does not prevent retitling that same passage. These categories can share items. Their counts do not establish seven separate passages merely because four plus three makes seven. ## Follow B and D through both categories Find B in the shortened group, then locate B again in the retitled group. Do the same for D. Each name refers to a passage already present in the first group. Its appearance in the second group describes another edit to that passage. Follow the remaining names too. A and C appear under shortened passages only. E appears under retitled passages only. Nothing in these lines gives A or C a second appearance, or gives E an earlier one. The written names make the overlap visible: B and D belong to both groups. Their repeated appearances are visible without asking the agent about them. What matters is recognizing their repeated membership when reading the total. This depends on the names identifying the same passages across the receipt. If B instead names a different passage in each category, the names no longer settle the overlap. Ask whether repeated labels refer to the same item rather than deciding from matching letters alone. These examples stay in short quoted lines. A passage could instead sit inside a [Markdown code block](markdown-code-blocks.md), but that visual separation cannot establish whether two names identify one passage. ## Count the passages once across the receipt Read through the names as distinct passages: A, B, C, D, E. That gives five. B contributes one passage even though it belongs to two categories. D contributes one for the same reason. Compare five with the stated seven. The two category counts add to seven appearances, but the named passages account for five distinct items. The receipt calls seven a distinct total, so its total does not fit its own named groups. Keep the category claims intact while stating that difference. Four passages were shortened and three were retitled, according to the invented receipt. Across those groups, five passages are named. Changing the category counts to force their sum down to five would discard the fact that some passages received both edits. An item appearing under a third category still contributes only one distinct item. Suppose an additional invented line reads: > Moved passages: 1. Named passage: B. The distinct names remain A, B, C, D and E. B has another category membership, without becoming another passage. Follow the item through every place it appears instead of treating each appearance after the first as new work on a new item. ## Question only the overlap that stays hidden A different invented receipt for Jane Doe gives less detail: > Shortened passages: 4. Retitled passages: 3. Total distinct passages: 7. Passage names omitted. Here, the categories can share passages, but the receipt does not reveal whether they do. Do not transfer B and D from the earlier example into this one. It is a different receipt, with its own unnamed items. A focused question would be: "Which passages, if any, belong to both categories? Please identify those shared passages so I can read the distinct total." That asks for the missing relationship. Repeating the described work or explaining every edit would go beyond this question. If the agent identifies two shared passages, count each of those once across the categories. If it says the groups have no shared passages, seven distinct passages fits the two counts as stated. If the answer only repeats four and three, the question remains open. Category counts alone do not tell you which items recur. Leave seven as the receipt's claim until you can establish whether it counts separate items. ## Keep appearances apart from distinct items For the named example, Jane Doe can write: "The receipt gives four shortened passages and three retitled passages. B and D occur in both groups. I count five distinct passages, compared with its stated seven." That account keeps the written counts, the shared names and the disagreement together. It also makes clear what you counted: passages across the groups, with repeated names counted once. If you keep this reading in a separate file, [Markdown file naming](markdown-file-naming.md) can help distinguish it from the receipt you received. Preserve the receipt's wording so the question remains attached to the claim that prompted it. Your reading ends when the named items account for the distinct total, or when you have named the overlap still needing clarification. A sum of category counts is enough to describe appearances across categories. Calling that sum distinct requires knowing whether the same items appear again. ## Where Sashmark fits Markdown reading plus text editing define the desktop v1 planned for Sashmark, still in development. Inside the app there is no AI. Your text comes from an agent outside Sashmark; reading it remains your part.