PROVA.

Serious evaluation, coming within reach.

Machine intelligence is starting to lower the cost of evidence, so questions that only well-funded programs can afford to answer today may come within reach of most programs.

Within reach

Questions that have always been too expensive to ask may be coming within reach.

Since the beginning of social science research, evidence has always been expensive, because reading has always been expensive. Language models now read interviews, case notes, and reports in many languages, and they help draft the design of a study. A program that chooses participants by lottery, or one with an eligibility cutoff or a staggered start, may already hold a fair comparison in its own records.

91%

In a 2025 study within six systematic reviews, data extraction assisted by a language model and checked by people was correct 91 percent of the time, against 89 percent for people working alone, and it saved a median of 41 minutes per study.1

The study used Claude models released between November 2023 and June 2024. On independent tests run by Artificial Analysis in November 2026, the leading models answered about 96 percent of a set of graduate-level science questions on which PhD experts score 65 percent, and about 88 percent of questions that require reasoning across documents of up to 100,000 tokens.2

The Prova Method

The Prova Method sets out how serious evaluation is done.

The Prova Method brings together the long-established designs of impact evaluation, such as lotteries, eligibility cutoffs, and staggered starts. It is organized in 11 layers that hold 149 factors a study has to get right, from study design and measurement to ethics and the use of the findings. It shows how a choice in one place changes the rest: changing the comparison can change who must be recruited, what must be measured, and what the result can show.

The program’s question, its records and documents, the Prova Method, and a frontier language model feed the work. The people who decide stand over it. The result is work that others can check.

Records read at scale

Everything a program has written down can now be read.

A team with years of interviews can have all of them read, in the languages in which they were given. Case notes, intake forms, and reports can be read in full, and a claim in a report can be checked against far more of the published research.

Who decides

A named person answers for every finding.

AI drafts designs, analyses, and reports quickly. Could a study that has always taken years take months?

Ravi NarainFounder of Prova
Knowing whether your work helps people should not depend on how much money you have. If a machine can bring that answer within reach, we should let it, and then hold the answer to the highest standard we have.

The method lives in the scaffold, outside the weights.

The Prova Method gives a frozen model 82 specified tasks, and each one has its own test.

The Prova Method is a compiled prior: an authored specification of impact evaluation, held over a frozen frontier model. It has 11 layers and 149 factors, drawn from roughly fifty bodies of research. Ravi Narain authored it, with frontier models as the build tool. The scaffold carries the task-specific intelligence, and the model fills sub-tasks.

82 of 149

Of the 149 factors, 82 give the model real work. The split between model and people follows the levels-of-automation framework of Parasuraman, Sheridan and Wickens.

  • The model aloneData pipelines, routine quality checks, power calculations.
  • The model drafts, a person approvesThe study design, the pre-analysis plan, the evidence output.
  • People feed the modelField context, quality annotations, protocol deviation notes.
  • People onlyRelationships, ethical judgment, community engagement.

Why any frontier model can be swapped in

  1. Every task has its own test.Each of the 82 tasks is specified with its inputs, its measurement indicators and its failure modes. A new model is held to the same checks at the same points.
  2. The split is re-tested.Which tasks the model takes is set by measured error, override rates and stakes, and it is reassessed as models improve.
  3. What it learns stays outside the weights.A calibrated library of effect sizes, intra-cluster correlations and attrition rates sharpens each next study and survives a model swap.

What it opens up

Because the method stays fixed, one scaffold can shape a model into very different products. On this site it already takes four forms: the Library’s collections, 47 free methods documents, the Lab and the Course. One person, working with frontier models, built all four. The scaffold makes it possible for one person to build work on this scale.

The same method also points toward directions Prova is exploring:

  • Why honest evidence is rare, and which rules might make it the ordinary outcome rather than an act of courage.
  • Answering a single question quickly, where the records and the scarcity already exist.
  • Measuring how quickly an institution learns to move money toward what works.
  • Making it safe to find out that something did not work.
  • Judging evidence apart from whoever paid for it.

Questions

How can we trust this?

Do not trust it on our word. Read our free method notes, articles, and standards, and judge the work yourself. Everything else is held to the same bar.

What happens when the AI gets something wrong, and has it?

It has, and it will again. That is why a person reads everything in full before release, and why every document tells you how to report an error.

If an AI can produce this, what is Prova adding?

The decisions about which problems deserve a document, what a leader needs at each step, and what the evidence can honestly support, together with the Prova Method, which sets out how the work is done. A model does not make those choices, and it does not answer for them. We do.

Why not just ask a chatbot?

You can, and for many questions you should. A chatbot will answer whatever you ask, but no one stands behind its answer.

Which AI models does Prova use?

Prova uses frontier language models and changes them as better ones appear. Everything the work must follow is written into the Prova Method, so a change of model leaves the standards the same.

What does this mean for people who work in evaluation?

Prova is built to help leaders understand evaluation: whether they need one and at what level, which questions to ask, and how to look for answers already in their own data. When more leaders see that this work is within their reach and their budget, we expect them to ask for more of it. Our hope is that this creates many more jobs in measurement, evaluation, and learning, for evaluators, researchers, and the people who do this work inside organizations.

Ask us anything about how we work.

Write to us, and a person will answer.

Notes

  1. AI-assisted data extraction. Gerald Gartlehner and colleagues, “Artificial Intelligence–Assisted Data Extraction With a Large Language Model: A Study Within Reviews,” Annals of Internal Medicine 178, no. 12 (2025): 1763–1771, DOI: 10.7326/ANNALS-25-00739. The study compared 9,341 data elements from 63 studies within six ongoing systematic reviews. Claude models made the first extraction, and a person checked it. Accuracy was 91.0 percent for the AI-assisted process and 89.0 percent for people working alone, with a median time saving of 41 minutes per study. Accuracy was measured for each data element, including whether the study reported it at all. The authors name the subjective judgment involved in classifying errors, and the difficulty of tracking time consistently, as limitations. The study covers one task, data extraction. The wider statement that language models can now take on part of the reading, source tracing, and checking, with a person checking the result, is Prova’s own assessment.
  2. Today’s models. Artificial Analysis, “GPQA Diamond” and “Artificial Analysis Long Context Reasoning (AA-LCR),” accessed November 6, 2026. The top GPQA Diamond score was 96.3 percent, against 65 percent for PhD experts. The top AA-LCR score was about 88 percent. Artificial Analysis does not score the models used in the 2025 study, so these figures describe today’s models and do not measure the change directly.