how long were these drugs actually tested?

the field we could not fill

The Resolv evidence library records eighteen things about every study: what kind of study it is, how many people were in it, the journal, the year, the DOI, the PubMed ID, who funded it, whether it was preregistered, whether conflicts were disclosed, whether the authors were independent of the product, a trust score, and more.

Trial duration is not one of them. Zero of the 235 records carry a structured duration field.

We could have quietly added one and backfilled it from what we know about each drug. We did not, because the absence is itself the finding — and because a duration inferred from "SSRI trials are usually about 8 weeks" is a fabrication wearing a data field’s clothes.

What we did instead: read the plain-English summary of all 235 records, pull out every horizon stated verbatim, and hand-classify one record at a time. Records that state no horizon are counted in a "not stated" row, not dropped from the denominator and not estimated. Every number below is reproducible by reading the same records.

60 of 72 randomised trials say how long they ran. 50 of those ran three months or less.

Of the 235 records, 72 are individual randomised trials — the design that establishes whether a drug works. Sixty of those state a treatment period or a primary-outcome horizon. Twelve do not.

Of the 60 that do: 50 measured their headline result at three months or sooner. That is 83% of the trials that tell you. Seven of them measured it the same day.

Six of the 72 ran past six months. Three ran past twelve. Those three are: a lithium relapse-prevention trial (up to two years), a 1986 methylene-blue crossover in 31 people with bipolar disorder (two years, 17 completers), and a fish-oil prevention trial with a median of 5.3 years that found no benefit.

One accounting note, published rather than smoothed: the 6-to-12-month row contains three records but only two distinct papers. ANTLER appears twice because it is filed under both fluoxetine and sertraline — the trial randomised people taking either. We count records, and we tell you when a record count and a paper count diverge.

the syntheses almost never tell you either

This is the part we did not expect. Meta-analyses and systematic reviews are supposed to be the layer that contextualises a scattered literature. In this corpus, they overwhelmingly report an effect size with no time attached to it.

Of 93 pooled-evidence records — 72 meta-analyses of randomised trials and 21 systematic reviews of randomised trials — only 10 state how long the underlying trials ran. Split: 8 of the 72 meta-analyses, and 2 of the 21 systematic reviews.

So when a clinician or an article tells you "a meta-analysis of 133 trials shows this drug reduces symptoms", the time horizon behind that sentence is usually not recoverable from the thing being quoted. Two records in this library do it properly and deserve the credit: the Cochrane review of amphetamines for adult ADHD, which reports a mean trial length of 5.3 weeks, and the St John’s wort meta-analysis, which states that every one of its 27 included trials ran only 4 to 12 weeks and that it is therefore unclear whether the finding holds beyond that.

the longest randomised horizon, by drug class

Read this table as "the longest time horizon stated anywhere in this corpus for that class" — across individual trials and pooled evidence together, because several classes have no individual trial record at all. Three rows say nothing stated. Those are not errors; they are the honest state of what this library can tell you.

and how long do people actually take them?

We are careful here, because our library is a library of studies, not of prescriptions — it contains no data on how long any individual stays on a medication, and we are not going to import a statistic from somewhere else to make the contrast look sharper. What it does contain are studies whose own design tells you something about real-world duration, and two FDA labels that state their own limits.

One more, worth flagging because it is a thing we deliberately did not use. A cross-sectional study finding that long-term benzodiazepine use reached 31.4% of users aged 65 to 80 (Olfson 2015, JAMA Psychiatry) was verified during the harvest and then cut by the five-studies-per-medication cap. It is a real, verified number and it is not in our corpus, so it is not in our count — mentioned here so you can go and read it, not counted as if it were ours.

Similarly: a three-year follow-up comparing long-term clonazepam with paroxetine exists, and it was left out of the clonazepam file because its PubMed ID could not be verified in that pass. It would have been the longest benzodiazepine horizon in the corpus. We would rather have a 10-week ceiling that is true of our data than a 3-year one we could not check.

what we will not claim

Four things this page deliberately does not say, because the data cannot carry them.

• We do not claim these drugs are unsafe over the long term. Short randomised evidence plus long observational evidence is a description of an evidence base, not a safety signal. Where this library holds real long-term harm signals — the benzodiazepine cognitive meta-analysis, the antipsychotic rehospitalisation cohort, the MTA height findings — they are recorded with their own caveats and they are observational.
• We do not claim a typical trial length for any drug class. With 5 records per medication, "the median SSRI trial is X weeks" would be a number computed over four trials. That is not a median, it is an anecdote with a decimal point.
• We do not fill the 12 not-stated trials from published papers we did not read in this pass. Their durations are knowable and we could look them up. Doing so mid-count would make the number unreproducible from the records, which is the one property that makes it worth publishing.
• We do not extrapolate from 235 records to psychiatry. This is a curated library, weighted toward the highest-quality evidence for each drug. It is enough to describe itself precisely. It is not a survey of the field.
• 235 is the 18 August 2026 harvest, not the current library size. The library held 251 published resources on 2026-08-19 and keeps growing. Every count on this page is fixed to the harvest so the arithmetic stays reproducible; the live library is free to read in full.