Reproduce the lab’s own numbers first, to the decimal

There is one honest way to introduce analysis software to a laboratory: reproduce the numbers they already have, on their own data, before asking them to believe anything else.

That is how we built ours. A lab sent a real instrument export and the workbook they process it with. Our analysis matched every result, not approximately, not within a rounding tolerance, but to the decimal place they report.

That exercise turned out to be less about our software than about what a spreadsheet quietly hides.

Why exact matching is the right bar

“Close enough” is not a scientific standard, and a difference of 0.02 in a reported value is not a rounding curiosity. It is a question nobody wants raised at review. If new software produces slightly different numbers from the workbook a lab has used for three years, exactly two explanations exist: the software is wrong, or the workbook has been wrong. Both are worth knowing, and neither can be settled by a feature list.

Insisting on an exact match forces the question into the open at the only point where it is cheap to answer: before anything depends on it.

What the exercise exposes

Reproducing someone’s calculation means reproducing every decision baked into their sheet. That is where it gets interesting, because those decisions are usually undocumented:

• What the result is normalised against, and how multiple reference values are combined. Arithmetic or geometric mean changes the answer, and the choice is rarely written down.
• Which sample is the baseline, and whether it is a single measurement, a mean of replicates, or a plate-level or batch-level control.
• How replicates are aggregated, and at which step. Averaging early and averaging late do not produce the same number.
• What happens to outliers. Most workbooks have exclusions applied by hand, with no record of which measurements were dropped or why.
• Where rounding happens. Rounding intermediate values rather than only the final result is the single most common source of small, unexplainable differences.
• Which assumptions the formula carries, and whether anyone has checked that they hold for this assay. Plenty of workbooks apply a standard model to data it was never verified against.

None of these are errors. They are choices, and a competent lab has made them deliberately. The problem is that the choices live in cell formulas rather than in a method, so nobody outside the workbook, including the lab six months from now, can see them.

The part a spreadsheet cannot do

Once the numbers match, the difference between the two systems is not the arithmetic. It is everything around it.

A workbook cannot tell you which version of itself produced a result. It cannot record that a measurement was excluded on a QC flag rather than a hunch. It cannot bind the number to the sample it came from, the operator who ran it, or the instrument export that fed it. It cannot show that history has not been edited since.

That is the actual upgrade: not better maths, but a result that carries its own evidence. The arithmetic has to be identical first, because identical arithmetic is what earns the right to change anything else.

What we would suggest, even if you never use our software

Whatever you use, do this once:

1. Write down the method. Normalisation, baseline definition, replicate aggregation, exclusion policy, rounding rule. One page. Version it.
2. Re-run an old dataset against the written method. If the numbers differ from what you published, you have found undocumented drift, and better now than in a correction.
3. Record exclusions where the data lives, not in a comment on a cell or a message thread.
4. Check the assumptions your formula depends on at least once per assay, and note the result in the method.

None of that requires buying anything. It only requires deciding that the method is a document rather than a habit.

The offer

Send us one raw instrument export and the workbook you process it with. We will put your numbers next to ours, tell you plainly where and why anything differs, and you keep the comparison whether or not you ever become a customer. Fifteen minutes, and worst case you get an independent check of a calculation your figures depend on.

---

How the Lattice founding cohort works · Why provenance is a recordkeeping problem