Public marketplace - vetted experts, pay per query

The public marketplace is HumanChain’s open surface. Vetted human experts across hundreds of domains answer your agents’ questions on a pay-per-query basis. Expertise grading decides who gets each question.

Use this when:

• You don’t have an in-house expert for the domain.
• You want to bootstrap fast without onboarding a chain of your own.
• Your agent serves a long tail of niche questions and you need coverage breadth more than you need internal control.
• You’re a developer experimenting with HumanChain and just want it to work without standing up infrastructure.

How it differs from a nominated chain

	Nominated chain	Public marketplace
Who answers	Humans you invited to your chain	Vetted public experts
Domain coverage	Whatever your team covers	Long tail, hundreds of domains
Per-query cost	Approx. zero (covered by seat fee)	Per-query, market rate
Question privacy	Stays inside your chain	Sent to public expert under NDA
Latency	Depends on your chain’s responsiveness	Depends on marketplace depth for that domain
Setup time	Days (invite + onboard humans)	Zero - works on day one
Best for	Internal-knowledge questions	Niche / cross-domain / one-off questions

Many teams use both - nominated chain for the domains they own, public marketplace as fallback (or first choice) for the long tail. The Nominated chains page covers the allow_public_fallback toggle that wires the two together.

How expert grading works

Expertise grading is HumanChain’s relative measure of demonstrated answer quality: an expert’s standing in a domain is a percentile among peers, never a published absolute score. It reflects how they did on calibration, how their answers hold up to peer review, and whether they answer in that domain regularly and recently. The grading model itself is proprietary.

A higher relative standing means higher routing priority for the matched domain_tags. The marketplace is a meritocracy: experts who answer well rise in their percentile standing; those who answer poorly or evasively fall.

Expert verification

Public marketplace experts go through a verification flow before they can answer paid queries:

1. Identity check. Government-ID verification via a third-party KYC provider. Stops impersonation.
2. Domain claim. Expert lists the domains they want to answer in.
3. Calibration battery. A short timed test for each claimed domain - questions with known answers, designed to detect bluffing without disclosing the right answers in advance.
4. Probation window. First 20 paid queries are flagged for meta-review by other experts. Bad probationary answers trigger review and possible removal.
5. expertise grading starts accumulating. After probation, the expert is in the active routing pool.

Not every signup makes it through. Verified experts are a small fraction of registrations; that’s the point.

Cost mechanics for marketplace queries

Each consult() call that routes to the public marketplace is billed at:

• 1 credit per response (so target_responses=3 costs 3 credits)
• Credits cost the marketplace rate for the matched domain - niche domains (specialised tax, medical, regulatory) cost more credits per query than general domains (software, data analysis)
• LLM fallback is metered at a flat 1 credit (regardless of domain)

The full cost table per domain is on the Pricing page and in the dashboard.

Note

On the free tier (25 consults per month), all credits are free - including marketplace queries. Past the monthly allowance, you need a paid plan to continue using marketplace routing.

What experts see

When the Hub routes a query to a public expert:

• They see the question text and any context_brief.
• They see the domain_tags matched.
• They see which sector the question is from (e.g. “fintech agent”), but not your company name, your agent’s name, your user’s identity, or your API key.
• They cannot see prior queries from your account.

This is sufficient for them to answer well without disclosing your business to them. If your question text contains identifying information (a customer name, a contract counterparty, an internal project codename), that’s your responsibility to redact - the Hub doesn’t auto-redact question text.

The standard expert NDA covers what they see; it forbids them from republishing, sharing, or using your question for any purpose other than answering it.

When to override the routing

Two override knobs are available on consult(), both optional and both unused by most integrations:

• exclude_responder_ids: list[str] - explicitly exclude experts who’ve previously given you bad answers. Useful when you want to prune your local “do not route to this person” list.
• prefer_responder_ids: list[str] - bias toward specific experts you’ve had good results with. Expertise grading still does the primary ranking; this is a thumb on the scale.

In v0.15 these accept a maximum of 5 IDs each. In v0.16+ they’ll support saved “preferred expert lists” you can manage from the dashboard.

Marketplace etiquette (for your agents)

A few things to bear in mind when your agent fires queries at the public marketplace:

• Don’t pepper. Don’t fire 50 queries on the same question with slight variations to “average out” the response. Expertise grading notices, experts get annoyed, and you pay 50× more than you needed to. Use target_responses=3, approval_method="consensus" instead.
• Provide context. A bare question with no context_brief gets a generic answer. A well-framed question with relevant context gets a specific, useful answer. Spend a sentence on framing.
• Respect the LLM fallback line. If your wait_seconds is short and you keep hitting LLM fallback, accept it - you didn’t actually want a human if 30 seconds was your budget. Either lengthen wait_seconds or stop expecting humans.

The single-voice LLM fallback

When the marketplace can’t route fast enough - no online expert in the domain, or wait_seconds ran out - the LLM fallback fires. Currently that’s Groq’s Llama 3.3 70B, hosted on Groq’s fast inference infrastructure. The response carries provenance="llm_fallback" and a fallback_reason field.

The LLM fallback is not five voices in committee. That’s a v0.16 roadmap item - for v0.15 we ship one voice, well-tuned, fast, and free during the trial window.

When to not use the marketplace

The marketplace is overkill (and wrong) for:

• Questions whose right answer depends on your internal state. (“What’s our SOW IP clause?” - a public lawyer doesn’t know your SOW.) Use a nominated chain for these.
• Time-critical inline questions where 60+ seconds is too long. The marketplace optimises for quality, not latency. If you need sub-5-second latency, you want the LLM fallback (or a different tool entirely).
• Tasks the LLM does well on already. Pure information lookups, code generation, format translation - pay the LLM, not a human. See Triggers for the calibration.

Next step

Now you’ve covered the patterns - when to call, what to do with the response, and where the humans come from. Operational details next:

Pricing → Security & privacy →