ventuno --world relay

Relay

next world

VIDEO SYNDICATION · MULTI-SIDED NETWORK · 2 CODEBASES

A managed video syndication network: content owners supply video, publisher destinations distribute it, advertisers buy the inventory, and the platform allocates revenue to every party at the terms in force when each event happened.

Relay is about money and decisions across a multi-party marketplace.

hear when it lands see the live product →

The world is built from a product we run: relay.ventunotech.com.

release

Next quarter after FAST World's general release

best fit

Both evaluation and RL training

planned potential

About 15–25 frontier-grade plus 40–60 training tasks

Task counts are planning ranges, not commitments.

// two codebases, one offline image

A platform and an ad server that must agree on who is owed what.

the network

Platform

Python · FastAPI

Partner admissions and agreements, rights and content management, media processing, recommendation and ranking, analytics, and operator, partner and advertiser consoles.

decides the ads

Ad server

TypeScript

Client-side ad decisions with targeting, brand safety, frequency caps, pacing and billable-event pricing (CPM, CPC, CPV).

The two share versioned contracts; together they carry about 1,100 automated tests.

// capability areas — what the agent has to get right

Revenue allocationSplitting each billable event between network, publisher and content owner at the agreement in force at event time

Period close and correctionClosed periods are immutable; late or corrected data becomes an auditable true-up, never a rewrite

Ad decisioningReservations, pacing, frequency caps and brand safety across the platform and the ad server

Rights and eligibilityContent rights, territories and monetisation restrictions applied before ranking and at delivery

Trust boundariesSigned requests, tenant isolation and attribution across services

Multi-party consolesWhat each party may see and change

// example archetypes

The same families as FAST World, in a marketplace setting: a single-service root cause in the platform or the ad server; a cross-service feature against a written contract, where both services must agree on who is owed what; and environment-dependent tasks, where the deciding fact is in the deployment's live data or configuration.

// fit

Both evaluation and RL training: money correctness between parties is easy to grade exactly and hard to get right, and the ad server adds a second service with its own state.

// where it stands — an honest ledger

◐ProductThe product's services and ad server are built and tested; partner statements and payouts are next on the product roadmap.

◐World imageA prototype world exists on an earlier harness. It moves to the FAST World harness (certified image, Harbor-format tasks, sealed gates) before release, which we estimate at 2–3 weeks of world engineering.

○TasksIts first tasks will be rebuilt to the current standard. Planned potential: about 15–25 frontier-grade plus 40–60 training tasks.

○ReleaseNext quarter after FAST World's general release.

Planned task counts are planning ranges, not commitments. The world is built and verified with the same methodology as FAST World.

// shared methodology

Every world is built and verified the same way.

The world is built from a real product we run, packaged as a certified, offline image.
Every task has a hidden milestone grader, a write-boundary policy, and a gate (reference solution 3/3, do-nothing 0) before use.
Every reported model failure is verified by replaying it in a fresh, network-isolated task container, and every pass is reviewed for grader holes.
Product updates never reveal task answers.

Build agents that can handle the real thing.

Working with foundation-model labs, coding-agent companies, computer-use teams and post-training researchers.

discuss a world FAST eval