“Not found” is not “does not exist”

An AI agent searches and reports “not found.” The person who delegated the search hears “doesn’t exist.” The first describes a query: its terms, its scope, the index it ran against. The second reaches beyond that evidence. Acting on the difference can mean deleting useful work, overlooking a collision, or telling someone an event never happened.

A positive finding gives you something to inspect. If an agent reports a duplicate at line 240, you can open the file. A bare negative gives you no matching artifact. The same fluent confidence can accompany a sound search or a broken one. That makes “no collisions found,” “never sent” and “doesn’t exist anywhere” high-risk claims to accept without a record of how the search was performed and what it covered.

I’ve watched this fail up close. An agent, asked to verify an event, searched a chat workspace, got zero results, and told the person it was working for that the event never happened. The event was real. One query carried a display name, spaces and all, in a field that expected an ID. The API accepted the malformed filter and answered with empty results instead of an error.

Another query returned so much output that it overflowed to a file. The agent never opened that file before concluding. Its unread contents were unknown. The agent had even run a control query to confirm the tool worked, but that control never exercised the broken clause. It left the relevant failure untested.

The drafts nearly deleted as duplicates

I keep a backlog of unpublished drafts in my knowledge base, and agents verify it in periodic sweeps. An inventory note flagged two files as stale copies of already-published articles and queued them for deletion under my single-source-of-truth rule.

The note was easy to believe. Both files had titles, slugs, December 2025 dates and canonical URLs pointing at my site. They belonged to an AI-assisted development series with published entries before and after those dates. Yet both source files also carried status: draft and an explicit do-not-publish notice. The inventory had treated plausible publication metadata as sufficient evidence and missed the contrary evidence in the files themselves.

A later sweep found neither article in the build output nor in a listing of 454 published posts across my sites. A complete listing answers what was present at that moment; it cannot establish that a page was never publicly reachable.

A December 19, 2025 commit, “Move draft articles to source materials,” removed the two files from the publication source path: 2,841 lines in all. Its message called them unpublished drafts and said they would serve as reference material for twelve new focused articles. Comparing the complete source trees before and after the commit shows those two paths disappearing while a known neighboring article remains. That is evidence of withdrawal from the source tree and of the stated editorial intent.

The first draft also describes an earlier accidental release and identifies the example as “this very article.” That account complicates the withdrawal message’s description: the files were withdrawn as drafts, but withdrawal does not establish that no reader had ever seen them.

The December refresh proposed extracting useful content from both drafts. An August 22, 2026 disposition recorded a refreshed copy of the learning article and recommended splitting the longer guide into three pieces. A catalog correction acknowledged that the shelving had been deliberate. Those editorial records correct my earlier claim that no decision was on record.

The two retained source files contain 12,424 and 3,705 words under LC_ALL=C wc -w, totaling 16,129. The count includes frontmatter, comments and draft material.

The inventory note had carried its own warning that an earlier sweep’s conclusion had not been independently re-verified. That label did not require anyone to reopen the evidence before acting. The useful safeguard was the actual recheck: keep the drafts, inspect their history and decide what to make from them. A warning on a cached claim cannot perform that work.

How a search misses what exists

  • An index may lag behind the source, or omit a content class such as thread replies.
  • The query may cover the wrong repository, account or host. Some systems also present inaccessible items as missing.
  • A malformed filter may be accepted and match nothing instead of raising an error.
  • The visible output may omit a matching item because a listing was truncated or results were written to an unread file.
  • The query may use different vocabulary from the material it needs to find.

Even local tools have scope defaults. The ripgrep guide documents that a recursive directory search normally skips ignored and hidden paths and binary files, and does not follow symbolic links. “I searched the directory” needs those choices made explicit when they affect the claim. A known match in an included file says nothing about a skipped one.

A filter can also be doing exactly what you asked. Removing filters is useful diagnosis; it does not follow that every broad hit with a filtered zero reveals a broken tool.

A zero you can check

The finite-corpus workbook contains four fictional records and runs locally with Python. There is no chat-service connection or model call. You can read the entire corpus before running the query:

IDChannelDateText
M01CH012026-01-02alpha
M02CH022026-01-02alpha
M03CH012026-01-03beta
M04CH012026-01-04alpha

The search uses a case-sensitive literal substring in the text, plus exact channel and date filters. Searching for alpha returns M01, M02 and M04. Adding channel CH01 and January 2 leaves M01. Changing the date to January 3 returns nothing, correctly: that record contains beta.

For a positive control, search for beta with the same CH01 and January 3 clauses. It must return M03. A positive control is a known matching item sent through the mechanism being tested. Here it exercises the same channel and date restrictions; it does not establish that an arbitrary alternative phrase would retrieve the right document.

The expected IDs and the hash identifying the corpus bytes were fixed before the query and checker were implemented. The query emits a receipt with its terms, corpus hash, record count, returned IDs and completeness. A separate checker compares that receipt against the literal expectations. For the empty result, it also requires the matching control receipt. Changing the corpus while keeping four records does not pass.

From the extracted folder, run:

python3 run_cases.py --out results
python3 -m unittest -v test_workbook
python3 fault_controls.py --out fault-results

The first command retains the receipts for the nine cases. They include the broad search, the matching filter, the legitimate zero and its control, an invalid channel, a deliberately broken filter, capped output, a different corpus, and unreadable input. The other commands check the implementation and deliberately introduce eight faults in copies. The tests must reject both missing results and a checker that incorrectly rejects the legitimate zero.

One result is deliberately incomplete: alpha with --limit 1 displays only M01 and sets complete: false. The local program has read the file, but the receipt has withheld matches. The checker labels it incomplete_no_absence. Rerunning without the cap produces the complete set. An invalid channel and actual invalid UTF-8 or JSON produce error status and exit 2; they do not become successful empty results.

For the valid filtered zero, the checked conclusion is precise: among these four records, no text contains the literal alpha with channel CH01 and date January 3, 2026. The corpus is complete for that question, the expected result is inspectable, and the same-clause control finds M03. The result establishes nothing about other records, another account or a historical deployment.

A complete, validated query over a known finite corpus can establish a bounded absence. Reading the primary source directly is the remedy when the search path cannot establish coverage, or when the claim requires interpretation that the query did not test.

Five rules for the agent and the reviewer

I use the synthesis-grounding-discipline skill, a reusable instruction module in my public collection, to carry this verification discipline into agent work. The finite-corpus example here makes the method more precise: completeness and the tested predicate determine what an empty result can support.

  1. Scope the claim to the instrument. “This query, over this corpus, returned zero” tells a reviewer what to inspect. Name the source, terms and restrictions. Reserve a stronger absence claim for evidence that establishes complete coverage of the stated question.

  2. Establish the bounds from the primary source. Read the relevant channel, directory or log over an explicit interval, or run a verified exhaustive query against the complete named set. “No matching record in this channel between March 3 and March 10” carries its limits with it. For repository history, name the revision range and paths: Git’s log documentation describes reachable commits and further filtering, not a record of every deployed page.

  3. Run a positive control through the suspect clauses. A known record in another channel cannot test whether this channel filter works. Use a control of the relevant kind that should satisfy the restrictions. A passing control tests that path; it does not prove that the selected corpus contains everything you meant to investigate.

  4. Compare filtered and bare queries. When the bare query finds something excluded by the filtered one, inspect whether the record should satisfy the filter. In the workbook, M01 correctly disappears when the requested date changes to January 3. A control that should pass the filter and disappears is different evidence.

  5. Follow truncated output to its source. “Showing 50 of 1,244,” an ellipsis or an overflow-file path means more material needs inspection. Open that file or obtain the remaining output before making a claim about what it lacks. Until then, its contents are unknown.

Build the evidence into the workflow

A false “not found” can look entirely plausible. Resolving to pay closer attention does not give the reviewer evidence that the search covered the right material. I want that evidence built into the agent’s standing instructions and my acceptance criteria. Before acting on “not found,” I ask:

  • What instrument ran, with which terms and restrictions?
  • What complete source or bounded interval did it cover?
  • What known match exercised the same mechanism and clauses?

If the agent cannot show the instrument, the bounds and a relevant control, the finding goes back for verification. I also need to know whether the evidence answers the decision in front of me. A file’s absence from a source tree can answer a source-tree question. It cannot establish a complete history of public availability, and it cannot decide whether preserved draft material is worth developing.

Two useful drafts returned to editorial consideration because the duplicate claim was challenged against the sources.