Expert judgment your agent can ask for.

Your agent knows facts — not which trade-off a particular person would make. Avva publishes that judgment. One link into Claude, Cursor, Gemini, or ChatGPT, and your agent consults a named expert before it commits.

Nothing here says the second answer is more correct — it says the expert’s record arrives before the answer instead of after it. That is what the protocol does: your agent consults the model before it commits to a call, not once you have already pushed back.

What happens, in four steps

Publishing takes an afternoon. Connecting takes one link. Everything after that is your agent’s habit, not yours.

An expert publishes their record

In Studio, their own assistant reads the decisions they already made and turns them into a model. Titles first — nothing private leaves.

Your agent connects

One link into Claude, Cursor, Gemini, or ChatGPT. No account and no key to start.

It reads before it commits

Facing a decision in that domain, your agent pulls the expert’s record first: their calls, the forces behind them, where they would stop.

It answers under their rules

The verdict is your agent’s own, made under the expert’s judgment — and it says where the model was silent.

Inside a published model

A model is not the output of a questionnaire. It is a small graph, fitted from calls the expert already made. It holds those calls, the forces each one turned on, and the order those forces resolve in when they collide. Nobody fills in a ranking. The math reads it off the decisions themselves.

A rules file converges: the fortieth decision and the four-hundredth both land on roughly the same handful of criteria. A case base does not — one decision in is one case out, so it is the half that keeps getting better as the expert keeps working.

How the order gets fitted

Three steps, and the expert is never asked to rank anything — every number here is read off decisions they already made.

1 · Every case is evidence

One decision reveals several contests

Each case records which forces settled it and which lost. That split alone is a set of head-to-head contests — one decision gives four here — and nobody ranked anything.

2 · Where the order comes from

Forces get a resolution order

The order comes from the method tournament statistics uses to rank players from their match results, run on those contests. It answers the one question that changes what an AI does: when two of this person’s forces collide, which wins.

3 · Matching is asymmetric

Which case counts as nearest

A case built on forces you do not have is about something else, so it is penalized in full. Extra pressures on your side cost little — the shared forces still hold.

Where the math stops

The rest runs on hard minimums. Every published claim carries the count of approved decisions behind it; a hard stop rule needs at least two; one decision becomes at most one case. Coverage labels (Light / Solid / Broad) say how much of the record a model rests on — not how accurate it is.

Whose answer it is

Avva runs no model of its own. Your agent does the reasoning, and the verdict it reaches is its own, made under the expert’s published judgment. The packet ends by asking what the model was silent on, so silence is never read as approval.

What your agent actually receives

Avva doesn’t invent an answer for you. It returns the expert’s verification protocol — the six parts listed here — and your own AI applies it to your context, on your tokens, in your tools.

Raw decision stories stay out of storage. Published models keep reusable judgment only. Details live in the Privacy Center.

Their cases: The calls they made, nearest to your situation first.
What to weigh: The forces in play, and which wins when two collide.
When to stop: The lines they would not cross — checked before anything is scored.
What changes the answer: The conditions under which their call flips.
What to ask first: Their questions, before an approach gets locked in.
Risks outside the model: What the expert was silent on — named as your agent’s own.

Two ways in

Avva is for professionals who get paid for judgment calls. Ask for someone’s, or put your own on the record.

For buyers

If you want an expert’s judgment

A standing reviewer inside Claude, Cursor, Gemini, or ChatGPT — not another chatbot to talk to.

Pick an expert. Read the calls they have made, the reasons they gave, and where they would tell you to stop.
Copy their link. One URL from their page, pasted into your AI’s settings. No account, no key, and it works right away.
Make it standing. Save the note from their page into the file your AI already reads, and it consults the expert without being asked.

Browse experts

For experts

If you publish your decisions

Not a resume, not a clone. The calls you already made, turned into rules your clients’ AIs can ask for.

Start from the assistant you already use. One link into the AI that has your work history. It searches, then shows you titles first — you cut anything private before any text reaches us.
Let it build the model. It reads what you approved, works out the rules behind your calls, and checks its own draft against them.
Skim, confirm, sign in to own it. You read the finished rules, not the raw stories. Sign in only at publish; unpublish any time.

Publish your decisions

What you end up with

A public page and an MCP endpoint you own. Hand the link to a client and their AI works under your rules, by name.
Or a fully private model — no page, no listing, only your keys open it. Go public when you trust the result.
You can unpublish or delete any time, and republish onto the same page as your judgment changes.
A count of how often your model was actually consulted — in your dashboard, nowhere else.
Your source material is never stored. Keep the decisions file your assistant saved — it is named for the model, like avva-decisions-b2b-pricing.json, so a second model cannot write over the first: it is the rebuild copy, and there is no published-model file to download.
Only the finished model publishes, with client names, employer names, and source quotes stripped out.

SoonEarn from it. When paid access opens, and it stays your call whether your model is ever paid.