# Multivac

There are more questions worth answering than resources to answer them. Even when
one question is inexpensive, billions of people asking questions can place enormous
demands on energy, compute, and human attention. Some questions overlap; others need
different tools, expertise, or access. Multivac explores how to connect those questions
with the right resources, while letting people choose what they support.

The name comes from the computer in Isaac Asimov's *The Last Question*, one of this
project's inspirations. The ambition is a shared platform for allocating scarce
resources across questions that need answering.

Research projects increasingly use different systems for coordinating agents,
experiments, computation, and human judgment. Contributors face a separate problem:
finding worthwhile work, deciding which system to support, and understanding what their
resources accomplished.

Multivac connects these sides. Independent projects expose an interface for
obtaining work and returning results. Contributors discover projects, offer resources
under their own constraints, and receive the project's account of what happened.
The platform can match across projects without taking over how they conduct research.

This includes CPU/GPU compute, work through contributors' existing AI tools, human
expertise, and analysis where data or software must stay local. A project can ask a broad
research question; a capable assistant can investigate and interpret it using the access
its owner permits. Only permitted findings are returned. This is particularly useful
when resources cannot simply be moved into one centralized research system.

Several projects may pursue the same objective through different orchestration methods.
They keep separate queues, research state, capacity limits, and acceptance rules. The
platform records allocation decisions, resource use, and outcomes across these efforts.
That creates a practical place to learn which arrangements help, while respecting the
fact that different projects' receipts do not necessarily measure the same thing.

An endpoint can represent a continuing research program, an organization, or a tool
for investigating a class of problems. The broader design includes people submitting
problems and feedback, standing offers of resources, and discovery of datasets that
can be analyzed where they are held. These are directions for development; the
current pilot demonstrates contributions to project-issued work.

## What works today

The invited pilot connects remote contributors to two Trefethen projects: an
independently hosted arena using EinsteinArena software with a custom Trefethen
integration, and a centrally coordinated research project. The arena is separate
from the official EinsteinArena service; contributions are not routed to
einsteinarena.com. Real investigations, certificate
reproductions, immutable research-file delivery, project receipts, cancellation, and
recovery have been exercised. Contributor and researcher tooling builds on MCP. A2A
supports optional delegated verification in the full integration.

The platform is publicly browsable; participation and project hosting initially require
approval. The small client and a project starter can be downloaded from the site without
access to the private platform repository. Donors retain their model accounts and local
execution. Projects retain orchestration and interpretation.

The protein-interaction and CryoET entries are proposed research tools. People would
submit their own problems; the tools would recruit donated resources to investigate
them, return findings, and use permitted results and feedback to improve their methods
over time. People could contribute problems, resources, feedback, or a combination.
Each tool would own its problem intake, research coordination, and learning process.

Both biology tools and Minecraft parkour TAS remain placeholders being developed with
their researchers. They have no live problem intake, resource contributions, or research
activity. GPU donation and a biological local-data workflow are not demonstrated yet.
Test fixtures are kept separate from research activity; no simulated discoveries,
participants, or performance gains are presented as real.

## What the pilot establishes

The immediate question is whether these connections work and are useful enough for
people to participate. Existing tests establish working contribution paths, not causal
scientific value or a performance advantage. A separately resourced controlled study can
use recorded policy versions, offers, resource classes, execution evidence, and native
outcomes to evaluate those questions later. Missing usage and costs remain unknown.

The allocation mechanism is replaceable. An expert agent can inspect activity and manage
versioned policy settings while the service continues routing eligible offers. It need
not be the mechanism invoked for each grant. Research projects remain independent of
that platform-level supervision.

Start with the [participation and project guide](getting-started.md).