GuideHollow_
Search ⌘ K

/ evidence_library

Don't just trust the example. Test it.

Original teaching data, expected answers and failure cases from our AI workflow guides. Each exercise makes it clear what was invented, what you can calculate yourself, and what our published records do not prove.

Start with evidence, not a score

Most AI tutorials show a successful prompt but leave you unable to check the result. Our examples instead preserve small inputs and explicit pass/fail conditions. You can download the fixtures, change one detail and see whether the recommended method catches it. These are not production benchmarks or verified real-world customer data.

Choose one exercise below. The numbers are intentionally small so a reader can reproduce them without an AI account. We include incorrect or ambiguous conditions because a procedure that works only on perfect inputs is not much of a procedure.

1. A spreadsheet average that looks right but is wrong

Question: did a proposed average selling price account for how many units sold at each price? Our original fictional eight-row dataset has 27 units and revenue of 530. The revenue-weighted unit price is 530 ÷ 27 = 19.629629…, displayed as 19.63. The simple average of the eight listed unit prices is 153.25, a very different measurement.

Download the eight-row sales CSV · Run the independent browser calculator · Read the complete check.

  1. Paste the CSV into the workbench and check that Web has 7 units and revenue 280, while Store has 20 units and revenue 250.
  2. Find line L-106: its price is 999, but it sold zero units. Its contribution to sales revenue must be zero.
  3. Find line L-107: it sold five units at an actual price of zero. Blank price and zero price are not interchangeable.
  4. Delete the price on L-107. A reliable preflight check rejects the missing input rather than silently treating it as free.

Scope: the calculator independently checks the original arithmetic in JavaScript. It does not reproduce a commercial accounting system, a live merchant sale or any AI model's answer.

2. A prompt score that must not masquerade as a benchmark

Question: does a classification rule focus on the requested action rather than on a misleading word? The original eight-case teaching fixture has a target answer for each fictional support message. Two editor-written prediction columns have five and seven matches out of eight, respectively. They were not generated by ChatGPT or another AI service.

Download the routing CSV · Read the evaluation method.

  1. Compare the expected label with each illustrative prediction column: the counts should be 5/8 and 7/8.
  2. Inspect case T05: the message mentions invoices, but its requested action is restoring login access.
  3. Inspect T07: two requests have no clear priority. Under this specific teaching policy the expected label is OTHER, yet both illustrative predictions choose ACCESS. Do not hide this failure.
  4. To measure a real model, run both prompts yourself, save every raw response, record the model and date, and replace the invented example columns.

Scope: the counts demonstrate transparent scoring, not that one prompt improves model accuracy by 25 percentage points in actual use.

3. A spreadsheet filter you can verify cell by cell

Question: when asked to total Paid orders in the East region, does a formula exclude West orders and unpaid East orders? Our four-row CSV contains East/Paid amounts of 120 and 30, so 150 is the expected sum. A West/Paid 80 and an East/Pending 50 must not be included.

Download the four-row CSV · Inspect our recorded Calc results · Follow the spreadsheet guide.

  1. Import the four rows into a disposable sheet and check the data types before using the formula.
  2. Try =SUMIFS(C2:C5,A2:A5,"East",B2:B5,"Paid") in a compatible locale. Compare its result with 150.
  3. Change the West amount. The East/Paid result must not change. Then change the East/Pending status to Paid; the result should rise to 200.

Scope: the original tutorial includes a dated publisher-recorded LibreOfficeDev run. We can check the underlying sample calculation independently, but the historic app session was not rerun during the November 1 editorial audit. This is not a claim that Microsoft Excel was tested.

4. A document summary must preserve the source's uncertainty

Question: can a summary distinguish a refundable deposit from a total cost, and a tentative booking from a confirmed reservation? Our fictional venue policy states a capacity of 32 seated people, a request lead time of at least seven calendar days, and a refundable deposit of 50 credits. It does not state the rental fee or confirm that a room was reserved.

Read the original fictional policy · Inspect the archived request and response · Read the source-check method.

Mark any sentence claiming that the total price is 50 credits, that a request can be confirmed tomorrow, or that the room is always available on weekends as unsupported or contradicted. These checks matter more than whether the resulting summary sounds polished.

Scope: the archived AI answer is a publisher-provided transcript; this page does not independently authenticate the historical interface session.

5. A source citation can still support the wrong claim

Question: does blocking a page in robots.txt guarantee removal from Google Search? No. Crawling restrictions and index exclusion are different mechanisms. The original claim ledger marks that guarantee as contradicted and distinguishes it from the requirement that Googlebot can read a page's noindex instruction.

Inspect the source-check ledger · Read the claim verification guide · Read Google's robots.txt documentation · Read Google's noindex documentation.

This is a check against published primary guidance, not a live experiment in how quickly Google will add or remove a URL from its index.

What we checked, and what remains unproven

Our editorial policy distinguishes checked calculations, source comparisons, fabricated examples and recorded model responses. The 31-article audit register describes the scope of our November 1 review and its unresolved limitations. The related test file independently recalculates example figures and verifies that published articles have structured source references.

We have not independently repeated every historic ChatGPT run, tested every spreadsheet product version, or proven a universal AI performance claim. The purpose of these exercises is to help readers detect mistakes themselves, not to make a small example look like large-scale research.

Found a contradiction, broken download or missing condition? Report it to GuideHollow, including the relevant example and the result you obtained.