Evidence, including the misses

How we test
Workbook Inspector.

You shouldn’t have to take our word for it. Here is what we tested, what we found, and what the tool still can’t do.

Hidden content can travel with a spreadsheet that looks ready to share. Our tests check whether the inspector finds it, reports its limits honestly and handles unusual files without sending their contents anywhere.

Current scanner: 0.2.0-poi-fixes.20261104.2 · Full automated suite: . Earlier blind and native tests are identified separately below.

Detection, without seeing the answers

Can it find what it hasn’t seen?

In a blind round, the scanner is frozen before the answer key is released. The workbooks contain planted items, clean controls and decoys: ordinary content that should not be called hidden.

  1. A separate authoring session prepares the workbooks and an answer key.
  2. The key and inputs are sealed with cryptographic hashes. The scanner checks the supplied files without consulting the key.
  3. Results are recorded before the key is released. No scanner changes are allowed during the run.
  4. The key owner scores each planted item and decoy. Fixes come afterwards, and the revealed files become regression tests.

28 October 2026 · B1

20 of 25 items located

One planted workbook and one clean control. 19 of 25 received the expected grade. None of the eight decoys, or the clean control, received an amber Hidden content finding.

Five items were missed: three had a declared gap or known limit; two were silent misses. The follow-up fixed those misses and improved coverage reporting.

Frozen build 0.2.0-stress.20261027.1

28 October 2026 · B2

33 of 33 items located

Eighteen unlabelled files, including planted, clean, decoy and hostile files; four were saved in Excel. 32 of 33 received the expected grade; the remaining item received an alternative grade allowed by the key.

There were no silent misses, but two of 23 decoys and one of five clean files received false amber findings. Those cases were fixed and are permanent regression tests.

Frozen build 0.2.0-excel-saved.20261028.1

We keep the weaker first result visible. These scores describe those files and frozen builds, not a general detection rate. Passing the same files after a fix is a regression result, not another blind result.

Robustness, not a recall score

Broken and unusual files.

Our rule is that an uncompleted check must not look like an all-clear. If a limit or unreadable component stops relevant checks, the report must name the gap or explain why the file was not inspected.

  • Malformed and hostile inputs: damaged archives, unsafe paths, duplicate entries, encrypted files, misleading size headers and markup that must stay inert in reports. Expected outcomes include explicit refusal, partial inspection and readable controls that must complete.
  • Boundary pairs: inputs just below and above the file, expanded-part, aggregate, retained-text and cell budgets, plus combined compression checks. Readable-part findings must survive a partial inspection. 152 scans across Node, Chromium, Firefox and WebKit matched the approved size-boundary expectations in the current validation.

The Apache POI soak test

We ran 835 spreadsheet test files from the Apache POI open-source project, including older and unsupported formats. All were scanned with sensitive search off and on in Node; a 748-file sample was also run in Chromium, Firefox and Playwright WebKit. That made 6,158 scans in the final run.

  • No crashes, hangs or out-of-memory failures were observed in that final run.
  • A separate structural review of the 192 files reported as Checked found no evidence of a false Checked result. This does not prove that all hidden content was detected.
  • All sampled browser reports matched Node on findings, grades, gaps and outcomes; timestamps and duration were excluded from the comparison.
  • Monitoring found no inspection-generated network requests or unapproved content writes. Browser background activity was classified separately; saving a report intentionally writes the requested export.

An earlier pass exposed a browser difference, unclear guidance and noisy findings. We fixed them and reran the corpus. These files have no answer key for concealed content, so this is a robustness result, not a recall score.

Different origins, different browsers

Real-world conditions.

Excel changes files when it saves them
We test script-built workbooks and copies saved by Excel for Windows 16.0.20527.20088 and Excel for Mac 16.114.3. Saved copies have separate expectations where saving recalculates values or removes stored content.
Large workbooks
The inspector allows up to 100 MiB per file and five million stored cells, with additional expanded-size and text budgets. Boundary tests check retained findings and planted items near the end of readable data. What happens beyond a limit →

In the 2 November streaming-build test, a one-million-cell numeric workbook took about five seconds in Chromium on a Linux laptop with an AMD Ryzen 7 7445HS and 16 GB nominal RAM. That is a measurement for that file and machine, not a promised scan time.

Browser parity
The current scanner’s automated suites run in Chromium, Firefox and Playwright WebKit on Linux. WebKit testing is not the same as testing native Safari.
Native Safari on an 8 GB Mac
On 3 November, Safari 27.2 on an Apple M1 Mac mini returned 80 of 80 test scans, all matching the reference Node results. That run used 0.2.0-size-boundary-fixes.20261103.2. A separate live-site scan and export also passed. Later scanner changes have passed WebKit tests, but have not yet had another native Safari run.
Privacy during inspection
The inspector page blocks network connections, and our browser tests monitor scanning and export. Workbook processing stays on the device. Public-page analytics and Pro accounts are separate services; the privacy notice explains them.

Compared with Excel’s Document Inspector.

In our fictional demo, Document Inspector in Windows desktop Excel 16.0.20527.20088, tested 28 October 2026, did not report text hidden by the ;;; number format. After its offered removals and a save, the text was still there. Workbook Inspector reported it.

That is a result for that file and build; behaviour can differ between versions. Excel did report an external-link warning. Document Inspector has useful checks of its own, and this comparison is not a reason to skip them.

Read the native results and download the fictional demo →

What it doesn’t do.

  • File types: the supported input is .xlsx. Older .xls, .xlsb, macro-enabled .xlsm and encrypted/password-protected files are not inspected. The report gives the reason and a next step.
  • Formulas are not recalculated. Macros and external links are never run.
  • Pictures: supported crop and overlap checks can identify concealed content, but cropped JPEG pixels are not analysed and there is no text recognition inside images.
  • Some drawings and legacy objects cannot be fully interpreted. Known limits and file-specific gaps are reported separately.
  • Sensitive search recognises supported patterns such as email addresses, UK mobile numbers and National Insurance numbers, plus selected personal-data headings. It does not recognise every kind of personal data.
  • It is not a certificate. A report records the checks performed; it is not a guarantee that a file is safe to share, a malware scan or a legal compliance assessment.

Gaps describe what was not checked, its location where known, and a manual check. Some unreadable components have an unknown scope; those reports say Only partly checked.

For procurement and information-governance teams

Take the evidence with you.

The one-page summary covers these results and their limits. This HTML page contains the full detail and can be read without a PDF viewer. For supplier assessments, security questionnaires or detailed test records, email info@42analytics.co.uk.

Current validation: 2,858 full-suite tests, 1,307 regression tests and 5,109 browser assertions. Scanner releases rerun these checks; site or account changes have their own checks against the retained scanner freeze.

Tested scanner: 0.2.0-poi-fixes.20261104.2
Last full automated test run:
Evidence reviewed: 5 November 2026