An evidence firm
Public trust depends on knowing what actually works.
Prova is designed to make evidence more accessible, more honest, and more decision-useful, starting with the records programs already keep and the claims society already needs to test.
Why it’s been rare
Knowing what works has been expensive, slow, and rationed to a few.
Credible evidence about how well a social program works has long been bound by three constraints. A serious evaluation was expensive enough that funders could commission it only for their largest commitments. It was slow, its findings often arriving after the decisions they were meant to inform. And it required an uncommon combination of methodological, operational, and contextual judgment, held by few, so the work was rationed to where those people could be placed.
The standard for what counts as evidence has not moved; what changes is the cost of reaching it. Frontier AI models can now do much of the reading and synthesis the work depends on, and people still hold the judgment that decides what a finding means.
Model, method, judgment
As frontier AI improves, Prova’s analysis gets stronger; the standard it must meet does not move.
Prova’s engine is being built around three parts that stay distinct. A general language model does the labor-intensive reading and synthesis of a program’s records. The method holds the structure and the standard. People hold the judgment that turns the model’s output into a graded, source-traceable finding. The model is replaceable, so as frontier models improve the reading gets stronger, while the method and the judgment stay fixed.
Frozen general language model
Reasoning substrate. As frontier models improve, Prova gets a stronger reader, extractor, and design-search engine. The gains compound.
Prova's domain intelligence lives here.
- Factors
- Layers
- Load-bearing triads
- Reinforcing loops
- Permanent tensions
- Cascade failure paths
Judgment stays with people.
Whether an effect is real, what it means here, and where the evidence runs out are human calls.
- Certify
- Downgrade
- Refuse
Auditable measurement instrument
Graded causal claim, provenance, and confidence level.
What changed
Frontier AI changes the cost of reading. How far it lowers the total cost is still being established.
Many of the foundational questions and designs of credible evaluation have been established for decades. Methodological work continues. Wider use has been constrained by the cost and time of some evaluations, the expertise they require, and the practical demands of recruitment, measurement, governance, follow-up, analysis, and reporting. Skilled reading is one material part of that work. In bounded evidence-synthesis tasks, frontier AI models can now perform much of that reading and synthesis under human supervision. What remains to be demonstrated is how far that capability lowers the total cost across real program records. Prova is being built to establish the answer.
The bar for what counts as evidence is exactly where it was. People still decide whether a finding holds, what it means in a particular place, and whether to stand behind it or refuse it.
The bet
Prova is being built to lower the cost of reaching each grade in its claim-evidence rubric, while the standard holds.
The judgment stays human.
Whether an effect is real, and what it means here, stays with the people who decide.
Prova’s grades stay fixed
The claim-evidence rubric runs from descriptive to experimental. Design quality and identifying assumptions still determine credibility.
Cost can come down
AI does the labor-intensive reading and synthesis, under supervision. The total cost change for each grade in Prova’s rubric is under test.
Judgment stays human
Whether an effect is real, and what it means here, remains a human responsibility.
- Prova’s four grades stay fixed. Its claim-evidence rubric runs from descriptive to correlational to quasi-experimental to experimental. A design label does not guarantee credibility; execution, measurement, assumptions, and fit to the question still determine what a claim can bear.
- Cost is the bet. AI does much of the labor-intensive reading and synthesis, under supervision. Prova is being built to establish how far the total cost falls without shifting work or risk elsewhere.
- Judgment stays human. Whether an effect is real, and what it means here, remains a human responsibility.
Prova’s rubric places claim-evidence relationships in four grades. The grades describe what a body of evidence can bear. They do not rank programs by worth, and no claim has to reach the top of the rubric. The right grade is the one the decision requires. A well-executed quasi-experiment can be more credible than a poorly implemented experiment.
The risks
Three conditions make the cost change possible, and each carries a risk.
Inexpensive AI-assisted reading can produce a large volume of low-quality evidence. A low price can be mistaken for a lower standard. Routine measurement can slide into surveillance, and metrics under pressure can be gamed. The conditions that lower the cost are the same ones that raise these risks, and both have to be weighed together. Prova is being built to address each one, and it names the same risks in its own work.
Prova is being built to address each one.
The conditions that lower task costs also raise these risks. Both have to be held in view.
lower reading cost
AI
POLLUTION
low-quality evidence at volume
reach more programs
Low cost
DISMISSAL
“too cheap to be credible”
support routine practice
Measurement everywhere
SURVEILLANCE · GOODHART
those measured have no say
Where to start
The starting point depends on the evidence question you bring.
By who you are
- You fund and want to see across what you fund.For funders →
- You run a program and want to know what your evidence can claim.For programs →
- You answer for a public program under scrutiny.For public agencies →
By the question in front of you
- What your current evidence can honestly claim.Engagements →
- Stress-test an evaluation before you commission it.Engagements →
- See across a portfolio you fund.Engagements →
Or see for yourself: a short evaluability read, no charge. Discovery →
Unsure which fits? Start a conversation.