How the Delivery Engine Works

The core problem in niche expertise markets: how does a market with 30 informed participants solve its own liquidity problem? Broadcast destroys signal quality. Manual curation does not scale. The delivery engine solves this with a five-stage pipeline that identifies the specific experts whose verified work makes them relevant, then delivers curated, exclusive opportunities at the click of a button.

📊
Expertise Graph
Capability Map
Matching Engine
Seven Components
🔒
Mandate Gate
Expert Control
📬
Delivery
Invitation
Receipt
Writeback

Everything before the gate is system intelligence. Everything after it is authorized action. The gate is the point.

Stage 1

The Expertise Graph

Every expert's verified proofs produce a domain graph. Not self-reported skills. Not keywords. Capability scores derived from independently verified work through Virgil. Each node is a domain tag with a capability score, recency score, and citation count. Edges connect domains that co-occur in the same proofs, representing methodology intersections.

The graph is the input to everything downstream. Its integrity is what makes matching defensible.

The matching engine scores every expert against every opportunity using seven weighted components:

ComponentWeightWhat It Measures
Domain Fit30%How many of the opportunity's required domains the expert covers
Capability Depth25%Average capability score across matched domains
Methodology Match15%Cross-domain edges. Does the expert's work span the intersection the opportunity needs
Recency10%How recently the expert published in matched domains
Marginal Signal10%Does this expert add information the existing participant set lacks
Urgency5%Time pressure from opportunity deadline
Exclusivity5%How many qualified experts matched. Fewer means each matters more
Marginal signal penalizes redundant matches. Exclusivity rewards scarcity. The system does not broadcast to everyone; it identifies the specific people whose verified work makes them relevant.
Stage 2 + 3

The Expert's Agent

This is the principal-agent problem solved by contract. The expert defines a mandate: their own policy for what their agent can do on their behalf.

Shadow
Agent observes. Logs what it would have done. Expert sees the recommendation after the fact. No action taken.
Recommend
Agent pre-fills the decision. Expert reviews and confirms or overrides. The default starting mode.
Auto
Agent executes within bounds the expert defined. Expert is notified after. Full delegation within the mandate.

The mandate specifies:

Allowed domains Blocked domains Min match score Conviction caps Weekly exposure limits Escalation triggers

If the opportunity hits a novel domain the expert has never published in, it escalates to the human. If the source trust is below the expert's threshold, it escalates. Every evaluation produces an auditable check record showing which rules fired and why.

The agent works for the expert, not the platform. The mandate is the expert's contract with their own agent.
Stage 4

The Invitation

When the matching engine identifies a qualified expert and the mandate layer clears the opportunity, the delivery engine creates an OpportunityInvitation:

🔗
Secure, single-use, time-limited link with 72-hour TTL. One expert, one opportunity, one action.
📋
Personalized match explanation citing the expert's specific verified proofs and why they were selected.
🤖
Agent recommendation showing what the agent would do and why, with full reasoning chain.
👆
One-click action: accept (with conviction slider), decline, or snooze. No forms. No friction.

The expert sees exactly why they were chosen. The selection reasons reference their own verified work. This is not a cold email. It is a curated, evidence-backed introduction to an opportunity the system has determined they are uniquely qualified for.

Stage 5

The Receipt and Writeback

Every action generates a cryptographically hashed receipt with full provenance:

Proof Lineage
Which verified proofs were used in the match
Active Policy
What mandate was active when the agent evaluated
Decision Record
What the agent recommended vs. what the human chose
Conviction
Percentage committed and QUAD exposure

The receipt writes to Campo's Signal Ledger. Accept and auto-execute are strict: if the writeback fails, the action fails rather than silently dropping a commitment. Decline and snooze are tolerant: memory-only is acceptable for low-stakes actions.

The Campo Score updates. The expert's reputation compounds. The flywheel turns.

The system finds opportunities.
The expert controls what gets through.

Every cycle produces a receipt. Every receipt compounds reputation. Every reputation attracts the next opportunity.

← Back to overview See the live economy →