Can you trust the result enough to put your reputation behind it?
Proof structures field evidence so you know three things before anyone draws a chart: which trials can legitimately be compared, what the combined evidence indicates, and how confident you should be. With the inconvenient results still in.
Every research decision risks one of two failures. Only one of them gets talked about.
The false positive
Declaring a product works when it does not
The claim ships. A regulator, a customer or a rival takes it apart, and the trail behind it does not hold. The product survives. The reputation does not.
The false negative
Abandoning a product that actually worked
Noisy sites, incompatible protocols and missing context concealed a real effect. The programme is shelved, and nobody ever finds out the evidence was the problem, not the product.
Both mistakes begin the same way: with comparisons that were never legitimate in the first place.
Pick any two trials. Proof rules before anyone pools.
A dataset can look complete while containing comparisons that are not scientifically equivalent. Proof decides compatibility first, from the method, protocol and context each record carries, and names its reasons.
The gate runs before any chart exists. An average across incompatible trials produces a confident-looking conclusion that is wrong, and no amount of statistics downstream repairs it. Synthetic demo data.
Watch what happens when the inconvenient trials disappear.
Eight compatible trials of one treatment effect. Three found little or nothing. The toggle removes them, the way selective reporting removes them, and the conclusion changes in front of you.
A value arrives from a site: 12.4. Without the instrument, the protocol version, the crop stage, the observer and what else happened that week, it is not data. It is a number.
Your team spends weeks matching sources before analysis can even begin.
Someone corrects a mistake directly in the spreadsheet. Reasonable, quick, and invisible. Two years later nobody can say what changed, who changed it or why.
Without a history, cleaning is indistinguishable from alteration.
The analysis plan said one thing in March. The conference slide says another in November. Between them, the success criteria moved.
Nobody decided to do that. It happened one small, undocumented choice at a time.
What 12.4 looks like when it can defend itself.
Every observation in Proof carries its full context as structure, not as a note somebody may have written. This is the same number, arriving as evidence.
12.4 % soil moisture
Measured
Volumetric soil moisture, 0–15 cm
Method
Capacitance probe · protocol v3.1, locked 12 Feb
Observer
Site technician · attributed, verified
Instrument
Probe #A-114 · calibrated 3 Mar
Where and when
Plot 14B · GS32 · 14 Apr, 09:12
Conditions
11 mm rain in previous 48 h · on the record
Deviation logged: reading taken one day late after field access failed. Recorded at the time, visible in every downstream analysis. Never smoothed over.
What was planned stays on the record next to what was analysed.
Selective reporting rarely looks like fraud. It looks like reasonable choices, made one at a time, invisible by November. Proof keeps the funnel public: what was screened, what was eligible, what was included, and why every exclusion happened.
Every exclusion is named: 38 failed eligibility on method mismatch, 18 excluded with reasons logged and attributed. When the cohort moved from version 3 to version 4, the record shows exactly which trials were admitted, which were removed, and who decided. The analysis cannot quietly change its own history.
Cohorts with rules
Comparisons that survive review
Cohorts are built from compatibility, not convenience. Strict and related-context records are kept distinct, groups under five are withheld, and every figure carries its denominator and method.
Evidence that already exists
Recruit from recorded field seasons
Set crop, practice and region and see how many consented field seasons already match your protocol. In the demo, the median from consent request to approval is eleven days, not two seasons of recruitment.
Corrections with history
Nothing edited in place, ever
Corrections append beside the original with author and reason. Exclusions are logged, not deleted. The negative result stays in the organisational record, which is how the next team avoids repeating it.
Proof never pools trials with incompatible methods or boundaries, never shows a group of fewer than five records, never lets an analysis quietly change its success criteria, never deletes a null result, and never strips attribution from anyone’s work. Every comparison carries its denominator, its method and its limitations.
Proof draws no verdict. It holds the record.
“Does this replace our statistics package?”
No. Proof decides what may legitimately enter the analysis and keeps the provenance; your statisticians keep their tools. The gate runs before the model, which is exactly where the damage is usually done.
“How does trial data get in?”
Honestly: that depends on how your sites record today, and it is exactly what we will walk through on the call, including what is manual for now. What we will not do is promise a connector that does not exist yet.
“How many records are behind the comparisons, honestly?”
Fewer than a database vendor would claim, and we will show you the real counts by crop and region on the call. The method is the product: where a cohort is too thin, the demo withholds it in front of you rather than filling the gap. That behaviour is worth more than a big number that will not survive your reviewers.
Proof is working with a first group of research teams through the 2026 season. There are no testimonials on this page, because we would rather earn real ones than write our own. Early teams help set the compatibility vocabulary their whole sector will inherit.
Bring two trials you have never been sure you could compare.
20 minutes. We will run them through the gate, show you what the combined evidence would say with and without the inconvenient results, and trace one number all the way back to the field.