Veralia
For teams running an AI agent that resolves support cases
An agent can hallucinate a resolution as easily as an answer, and some platforms count a customer going silent as a success. Every number you have comes from someone with a stake in how it looks. That is part of why so many companies are spending heavily on AI and cannot say what they got for it.
Veralia is an independent measurement firm. We lock the test in writing before we look, score your real cases against evidence the agent cannot touch, and hand your engineers the failing cases with the exact rule each one broke. You learn what your agent actually does. Your team gets what it needs to tune it.
By their own published definitions, these can be the same event.
The gap
None of this requires anyone to be acting in bad faith.
In research, evidence is ranked: the higher the tier, the harder it is to fake. Agent metrics have the same structure. Your headline figure is drawn from the bottom of the pyramid. We score at the top.
Three things follow.
One platform counts a resolution after seventy-two hours of silence, no customer confirmation required. Another counts an assumed resolution when the customer leaves without asking for more help. A third lets its own system classify the outcome. The platforms say so themselves: their own documentation warns these figures can include customers who dropped off or were unhappy.
Resolution rate, automated resolution, automation rate, deflection, containment, and the list keeps going. Nine or more names for what is presented as one thing makes comparison impossible.
A rate without a denominator cannot be checked.
What we score against instead
Not the agent's opinion of how it did. We score against evidence that exists outside it: a resolution code, a refund issued, a shipment sent, a human stepping in.
Sooner or later someone asks you to prove the agent is working. Answering that question is the only thing we do.
Founder
Veralia is run by Aidan Rubio, a published scientist at Stanford Medicine. He started in research at sixteen as an NIH research intern and has held research roles across five labs at Stanford and the University of Pennsylvania.
The discipline here is the one that keeps published science honest: decide what counts as a pass before you look, write the sample down so anyone can redraw it, report the result you got.
We are measurement specialists rather than agent builders, and that is deliberate. The person scoring the test should not be the one selling the fix. Your engineers apply every change; we do not advise on implementation.
Independence
Revenue we have foreclosed, in writing.
Independence here is a design constraint, not a slogan. We narrowed the business to measurement alone and turned down every revenue stream that could lean on a result, so the conditions stay scientific. The list is written down so you can hold us to it.
The one interest we do have
Design partners get a reduced rate in exchange for a named case study and a reference call. Case studies are worth more to us when the result is good. We would rather you hear that from us now than notice it later.
Method
This is the part you should argue with.
Everything that could be bent to flatter a result is written down and sealed with a digital fingerprint before any case is read. Change one character of the plan afterward and the fingerprint no longer matches. Clinical trials run on the same rule, called pre-registration: the test is fixed before the results exist. Changes are allowed; silent changes are not: every amendment is numbered, dated, and printed next to the original.
Before we see anything
Which cases count is defined in writing first. Exclusions are declared and counted, never quietly dropped.
The sample is drawn by recorded rules and a recorded seed. Anyone can re-run the draw and get the same cases.
What counts as a pass is fixed per metric before a single case is read. Nothing is added after we see data.
How scoring works
Four standing checks on every case: did the claimed resolution match the recorded outcome, did it hand off when its own policy required, was every factual claim supported by the record, did it stay inside its declared limits.
Every case is scored more than once, independently. Disagreements are resolved against the written rule, and the disagreement rate is published.
Proof the fixes worked
Part of the sample is sealed from the start and never used to tune anything. After your engineers fix, we re-test on a fresh sample under the same locked plan, and report before, after, and holdout side by side, even where they disagree.
Every number can be taken apart
Each scored case carries case id, sample group, seed, scorer ids, timestamp, and rule version. Dispute any figure and we walk you back to the exact cases behind it.
The engagement
Fixed scope, fixed price, fixed end date.
One engagement, three moves. Scope, price, and end date are fixed at signature.
You see the full protocol before we run it.
Historical data, never production.
We re-test and report whether it worked.
What you walk away with
A scored, provenance-stamped reliability record: the test, the results, and the proof in one document you can put in front of anyone. What failed is reported as clearly as what passed.
Qualification
Thirty minutes. A qualification call, not a pitch.
Two people on the first call: the one who answers for the quality number, and the one who can pull the data.
At least 300 qualifying cases.
Not just transcripts, and never the agent grading itself.
One of the two data modes, owner named.
If all five clear, you get a written scope and draft protocol before anything is scored. If any fail, we say which and what would have to change. Not yet is a good outcome.
Limits
Before you consider this
The hostile ones are here on purpose.
You do, first. The record is delivered to you, not published. It reports the misses as plainly as the passes; nothing gets softened or moved to an appendix. Every figure traces back to specific scored cases, so a number you dispute can be taken apart rather than argued about.
It is the right question, and "we are independent" is not an answer to it. The answer is the method, not the money. There is nothing we can sell you afterwards that depends on how the result comes out. The price is fixed before the work starts, so it cannot move with what we find. The rules and the sample are locked before any case is read, changes are allowed but never silent, and the one interest we do have is printed above rather than left for you to find.
Not for a client. Veralia is new, and we would rather say so than dress it up. There is no case study and no result to show you. What exists is the method above, written down in full and open to attack, and the separate record below. If you need a completed engagement before you will consider one, we are not the right firm yet, and that is a reasonable position. Design partner slots are open for exactly this reason.
Then one question matters: what sets your recorded outcome? If the agent grades itself, there is nothing independent to score against, and that is a disqualifier here rather than something we work around. It is also worth checking what your headline number actually counts, because cases with provable outcomes are a smaller set than inquiries. The same goes for the dashboard your vendor gives you: it reports the vendor's definition.
We cannot. We do not implement, advise on implementation, or touch any system. Fix-it code, agent development, production access and ongoing monitoring are all outside the engagement, permanently.
Scoring is done independently more than once, blind to the agent's own confidence and to any vendor-supplied quality label. Disagreements are logged, resolved against the written rule, and the disagreement rate is published in the record. Ask on the call how that is staffed for your engagement and we will tell you plainly.
No production access, ever. We work from historical case data. You choose how it is handled before any of it moves, and one of the two modes never moves the data at all. We do not hold raw personal data, and that is an operating rule rather than a preference.
A fixed price, quoted in writing before any work starts, agreed together with the scope and the end date. There is no figure on this page. Ask and you will have one.
Our own record
Provenance, not accuracy.
An automated instrument we operate makes forecasts on public questions and scores them against what actually happened. Before we look at any result, the true outcome is pinned down from independent sources. We grade ourselves against the world, never the other way around.
Every resolution in the record, one mark each
As of July 28, 2026: 196 of 196 wire-gradable resolutions matched pre-registered truth cards pinned from independent sources before any result was read. Multiple related markets often resolve off a single fact; the 196 matches rest on roughly 45 independent ground-truth facts.
Read that for what it is. It proves we lock in the answer before we look and keep our misses on the books. It does not prove our forecasts were good, and we claim no such thing. Discipline is what we sell.
The full record is not published at this address yet. We would rather tell you that than point at a link that does not resolve.
Contact
The design partner rate is exchanged for a named case study and a reference call.
Tell us a little and we take it from there. Every reply comes from Aidan directly.
If the timing is wrong, say so. Not now is a complete answer.
aidan@veralialabs.com