Document & data assurance for digital agencies

The numbers in your client report,
checked before your client sees them.

We recompute every assertion a document derives from its own figures — percentages, shares, rates, totals, averages, growth — and return a report showing what held, what failed with the arithmetic shown, and what we could not check and why. Deterministic arithmetic, not a language model's opinion.

23published agency documents verified
9arithmetic errors in those documents
0false accusations raised in them

The checks

Six things we recompute from your document's own numbers

Not a spellchecker and not a chatbot. Each check below is arithmetic that either holds on your figures or does not — which is why we can put the working in front of you and stand behind every line of it.

01

Growth and change

Any percentage change recomputed from the two figures the document itself states, so “up 34.5%” is tested against its own £118,400 and £152,900. Superlatives are held to the same standard: “doubled” has to be twice.

02

Shares and totals

Parts of a whole that must sum to 100%. A row advertising seven channels is expected to account for all of them.

03

Column and row sums

Every table total recomputed from the cells above it, to the penny, with any difference stated plainly.

04

Averages and per-unit figures

Cost per conversion, revenue per session, average order value — recomputed from the underlying totals, and flagged when a simple average has quietly over-weighted one group.

05

The same figure in two places

Summary against detail, prose against table, caption against chart. Where a document states one measure twice with two different values, we say so.

06

Figures that imply a base

A share, a rate or a multiple that can only be true if an unstated base exists — and the base the document implies is not the one it uses.

Worked example

What a finding looks like

Two findings of the kind we return, shown in full. They come from a fictional test report we wrote ourselves to prove the engine works — not a client's document, and not anyone else's published page. Real findings are returned the same way: stated value, recomputed value, and the working between them.

Must fix Table 2 · Sessions column · the printed total does not equal its own cells
Printed
234,050
Recomputed
232,050
152,900 + 41,300 + 28,700 + 9,150 = 232,050 the table prints 234,050 — difference +2,000
Must fix Executive summary · cost per conversion · a simple average where a weighted one is needed
Printed
£127.42
Recomputed
£133.71
total revenue 486,300 / total conversions 3,637 = 133.71 the printed 127.42 is the unweighted mean of the two per-site averages: (141.03 + 113.80) / 2 = 127.415 — the printed figure to the penny, and the wrong measure, because it over-weights the smaller site
Held And when a document is right, we say that too — with the count

A clean report is not an empty one. It states how many assertions were recomputed and passed, so you can see the document was actually examined rather than skimmed. Checks that cannot be settled from the document alone are grouped by kind, counted, and given an example — never quietly dropped.

Check it yourself The test report and its source data are published with this page

Download both files, recompute every line in a spreadsheet, and the figures above come out the same — or we are wrong. This is the standard the work is held to: the test report · the source data.

report_planted.md sha256 202cfa268345867dabc7d9acbf3b33be204937af4564ecc4f846ffb7a612d23b source_northwind.csv sha256 10ad15dcc9cc67856018bfbe73dde6982f728dfdd9e11c4aee35a197e953aac6

The limits

What we do not check — stated before you buy

Most of what goes wrong between an agency and a disappointed client is not arithmetic. We are precise about our edge, because a verification report that quietly implied more than it did would be the worst thing we could sell you.

In scope

  • Arithmetic the document performs on its own figures
  • Tables, charts, captions, footnotes and body text, treated as one document
  • The same measure appearing more than once, wherever it appears
  • Percentages, points, shares, rates, multiples, averages and totals
  • Anything settled by arithmetic — written down with its working

Out of scope

  • Whether a claim is true in the world — we test internal arithmetic, not external facts
  • Spelling, grammar, tone, house style, or how persuasive the document is
  • Strategy, channel choice, recommendations, forecasts or predictions
  • The quality of the analysis, or whether the campaign worked
  • Data methodology — whether a sample is large enough, or a survey was sound
  • Design, copy, SEO performance or the security posture of your systems — no opinion is offered, and none is charged for
  • Numbers that cannot be read from the file — scans and images are declined, never guessed at

How it works

One document, one working day

Send the document

Email the finished document — PDF, .docx or a link. Nothing else: no client data beyond the document itself, no account access, no installation, no call.

If a PDF cannot be read faithfully — a scan, a password, or tables that only partly reconstruct — we say so, ask for the Word file, and charge nothing. We would rather lose the fee than issue a report we cannot stand behind.

We verify it

Every assertion the document derives from its own figures is recomputed. Findings are checked again before they leave us. A check that cannot be settled from the document alone is grouped with its kind, counted, and given a worked example.

The report comes back

Within one working day — documents arriving after 15:00 London time, or several at once, are handled in arrival order the next working day. Branded as your own, with no mention of us.

Coverage

The document verified, the moment of verification in UTC, how many assertions were recomputed from its own numbers, and how many failed.

Every finding, with its working

Severity — must fix, should fix, or a point worth a look — the location, the figure the document prints, the figure its own numbers produce, and the sentence or cell as published.

The checks we declined

Grouped by kind, with a count and one worked example of each, and the statement that the full list is available on request.

Provenance

The file or link verified, the fetch timestamp, and the sha256 of the exact bytes the findings came from — so there is never an argument about which version was checked.

Price

Fixed fee, published, no per-word meter

Prices exclude VAT. A document means up to 30 pages, up to 25 tables or up to 15,000 words — whichever comes first. Beyond that we quote before starting, and a document we decline on size does not use up your allowance.

£450 per month · up to 4 chargeable documents
  • That is £112.50 a document
  • One working day, arrival order after 15:00 London
  • White-label report, your brand on it
  • Additional documents at £90 each
  • Cancel any month; no minimum term
£90 single report · first month · no commitment
  • The whole report, one document
  • The same checks as the retainer
  • See the working before you commit
  • No setup fee · no minimum term

What the report is, and is not. It verifies the arithmetic a document performs on its own numbers, prepared from the document as supplied on the date shown. It is not a guarantee that every error will be found — a figure the document does not derive cannot be checked, and those are listed in the report. If we miss something inside the scope of the engagement, we correct or re-run the report at no charge.

Think a finding is wrong? Send us the cell. We re-derive it from the document as you supplied it and answer the same working day. If the arithmetic does not reproduce, the finding is withdrawn, the report is re-issued — and if it was the only must-fix finding, the document is free.

A document where we find nothing that must be fixed is free, and does not use up one of the four. You are paying for problems found, not for volume — and we do not flag anything to justify a fee.