Independent review for AI-written work
AI did the math. We check it.
Give Sternly the analysis ChatGPT, Copilot or Gemini wrote and the data or documents it came from. Every number is recomputed with plain code, every sentence is checked against your sources, and you get back a marked copy and a corrected one.
One check free, no account. Your files are processed and dropped after each check.
Review record · Q3 business review
Not ready to send.
18 numbers recomputed from orders.csv and store_summary.xlsx.
Store net sales for Q3 2026 totaled $784K.
$392,239The file's total row was counted as a store, so everything was added twice.
Across our six stores, the average order value was $83.05.
$75.43A plain average of six store averages. Weighted by orders, it is $75.43.
October net sales are down 70% compared with September.
−69.7%The figure checks out, but October's data stops on the 15th.
West net sales grew 12% from Q1 to Q2.
12.2%
How it works
A strict reviewer that shows its working.
No template to fill in and nothing to install. Sternly reads what you already have and checks it the way a careful analyst would, only every time.
- Drop in the work
The AI's analysis, report or answer, pasted or attached, plus the CSV, Excel files or documents it came from.
- It reads like an analyst
Each column is matched to a meaning, totals and blanks are found, and when a word could mean two things it asks you once.
- Plain code does the math
Every number becomes a recipe run on your files, judged with fair rounding and tested for the traps AI falls into.
- A report card you can hand on
Each line gets a verdict, the right figure and the work behind it, plus a corrected copy to send.
The core rule: AI only reads and matches; plain code does all the math.
Four verdicts
Every claim gets one, in plain words.
Each verdict comes with what your source actually says and what to do about it. The note is about the draft, never about you.
Your data or sources say the same thing, within fair rounding.
Gross margin held steady at 38.5%.
Right in part, or right but misleading: a missing caveat, a partial month, one outlier doing the work.
Average order value in September was $135.
Not in your sources. It may be true, but nothing you gave us shows it. Add a source or cut the claim.
Customer acquisition cost fell to $32.
Your sources say otherwise. We give the right figure and the working, and put the fix in the corrected copy.
We processed 2,664 orders in Q3.
Each verdict is a word, an icon and its own kind of mark, so none of them depends on colour alone.
One reviewer, eight desks
Whatever the AI made, there is a check for it.
Check numbers
An analysis checked against the data it came from: every figure recomputed, every trap named, every line with its work shown.
Check an analysis →Check a spreadsheet
A workbook an AI built, formula by formula: totals that stop short, wrong divisors, broken patterns and typed-over numbers, cell by cell with the fix.
Check a workbook →Review against sources
A report or answer checked sentence by sentence against the documents it came from: grounding, quotes, citations, context and numbers, then handed back revised.
Review a report →Check a deck
A deck checked against its own charts and tables, and against the memo, report or workbook that goes with it: every figure on the slides read every fair way, each disagreement with the right value and where it comes from.
Check a deck →Check references
Every reference, court case and quotation in a review, brief or memo, looked up: sources that do not exist, wrong years and authors, retracted studies and misquotes. A lookup that fails says so.
Check the references →Agent receipts
The agent says the tests pass; its steps say one failed. Every statement in an agent's final message checked against what its tools printed and how they exited. Nothing is re-run.
Check an agent run →Rewrite check
A polish, summary or translation set beside the original, sentence by sentence: changed numbers, dates and names, reversed meanings and dropped facts marked, in English, Spanish and Korean.
Check a rewrite → PreviewForecast Studio
Pick what to predict. Two models scored on rows they never saw, plain-English drivers and every prediction to download. Code trains; AI only narrates.
Open the studio →Your company's words
“Sales” at your company means net sales; “active” means bought in the last 90 days. Answers and notes become a glossary that every check reads first, on top of a shared core of definitions, and nothing changes without a person approving it.
See the ontology →The traps
The mistakes a fluent model makes with a straight face.
A number can be computed correctly and still mislead. Sternly runs each figure the right way and the trap ways, and names the trap when the draft fell in.
- Total row countedThe file's own total added in as one more row.
- Averaging averagesSix store averages averaged, as if every store were the same size.
- Percent vs. points6.3% to 3.0% called “down 3.2%”.
- Wrong baseA share of orders reported as a share of customers.
- Partial periodHalf a month compared with a full one.
- One outlier drives itA single $50,000 order moving the average.
- DuplicatesRepeated rows counted twice.
- Mixed currenciesCAD and USD added as if they were the same.
- Cherry-picked windowA start date chosen to flatter the change.
- Cause from a pattern“Thanks to the new program” when the data only shows timing.
For agents
Your AI agents can check their own numbers.
Sternly's playbooks are MCP tools and an HTTP API, so an agent can check an analysis before it sends it. Every check leaves an anonymized record, never your data values.
→ sternly_check_analysis analysis_path: "q3_business_review.md" data_paths: ["orders.csv", "store_summary.xlsx"] ← 18 numbers checked: 5 correct, 6 wrong, 6 misleading, 1 can't verify ✗ $784K → $392,239 total row counted as data ✗ $83.05 → $75.43 averaging averages ✓ 12% = 12.2% West net sales, Q1 to Q2
Careful with your files
A reviewer you can hand the real numbers to.
Files are processed and dropped. The benchmark keeps an anonymized record of each check, never your data values, and raw files only if you opt in.
Every input is capped, and an attack suite of prompt injection, formula cells, sandbox escapes and invented numbers runs on every release.
Every verdict opens to the recipe that produced it, so anyone on your team can follow the arithmetic without trusting us.
Private preview
Check your first analysis free.
- One check free, no account needed.
- Sign in for two more, and a glossary that remembers your company's words.
- A report card and a corrected copy for every check.
Sign in
Continue with Google Continue with AppleYour files are processed and dropped after each check.