Experiment learning library

Every completed experiment review gets a row here, linked back to its experiment brief. The library exists for one reason: a test whose result nobody can find gets run again. The second run costs the same as the first and produces the same answer, and nobody realises until someone remembers halfway through.

Null and negative results are filed with exactly the same care as wins. They are the rows most often skipped and the ones that save the most work, because “we tried that and it did nothing” is only useful if it is written down somewhere findable.

The register

The first three rows are examples using Doughnut Labs, a company that sells Doughnut Technology. Delete them and fill in your own.

ID	Hypothesis	Surface	Metric	Result	Status	Decision	Reviewed
E.g. EXP-2026-014	E.g. Naming the price on the ad raises trial quality	E.g. Meta prospecting	E.g. Trial-to-paid rate	E.g. +11%	E.g. Win	E.g. Rolled into the copy brief as a default	E.g. 2026-05-02
E.g. EXP-2026-015	E.g. Social proof above the fold lifts demo bookings	E.g. /lp/book-a-demo	E.g. Demo booked rate	E.g. +1.2%, not significant	E.g. No result	E.g. Reverted. Not worth retesting at current traffic	E.g. 2026-05-20
E.g. EXP-2026-016	E.g. Shorter form increases completions	E.g. /lp/fresh-batch-trial	E.g. Trial started	E.g. +9% starts, minus 14% trial-to-paid	E.g. Loss	E.g. Reverted. Guardrail caught it	E.g. 2026-06-11
[add row]

Status values

Four values, so the register can be filtered rather than read.

Status	Means
Win	Beat the decision rule set before launch. Rolled out
Loss	Moved the wrong way, or a guardrail metric was damaged. Reverted
No result	Did not reach significance, or the difference was inside the noise
Invalid	Something broke: tracking, a mid-flight change, insufficient sample. Not evidence either way

Invalid matters as much as the other three. A test recorded as “no result” when it was actually broken tells the next person the idea doesn’t work, which is not what happened.

Before you design a test, search here

Search on the surface and on the metric, not on the wording of your hypothesis. Someone else will have described the same idea differently. Three outcomes, and all three are useful:

Already tested and won. It should already be a default. If it isn’t, that’s the finding.
Already tested and lost or showed nothing. Either move on, or state what is different now: more traffic, a different audience, a changed product. “We’ll try it again properly” is not a difference.
Not tested. Design it, following run an experiment.

What a row needs to be useful

The register is only as good as the discipline behind the rows.

The ID, matching the pattern in experiment ID, so the test can be traced to the campaigns and events that carried it.
The hypothesis as it was written before launch, not a version rewritten to match the outcome.
The decision rule’s verdict, not an interpretation. Whether it beat the bar is a fact; whether it was worth doing is an opinion, and belongs in the review.
What actually changed as a result. A win with no decision recorded is a test that ran for nothing.

When a result stops being true

Results age. A test won in a different market, at a different price, against a different competitive set may not hold. Mark a row superseded rather than deleting it, and link the test that replaced it, so the history of what was believed and when stays intact.

Run an experiment The full path, ending with a row here.
Experiment review process What produces the row.
Experiment brief The workbook each row links back to.
Experiment ID The ID pattern and the issued-ID log.
Experimentation best practices Why nulls get recorded.