Validator Nodes

Validators are being built as the Grid’s independent audit layer. Today’s preview checks specific text job contracts and submits attributable signed evidence. It does not prove a remote machine runs a particular model, establish general answer quality, or grant workers more routing trust.

At a Glance

Validator Data Flow

The architecture is staged. Production Core issues a sealed text assignment: the node sees its opaque ID, capability, and commitment before execution, then verifies the disclosed worker, challenge, nonce, and result afterward. Core hard-targets the worker and verifies submitted signatures and evidence binding before aggregating non-economic quorum state. Missing assignments fail closed; the public preview does not fall back to random production probes. Current preview evidence does not move money, change routing, or slash stake; media reference lanes and Base roots remain disabled.

Distinct registered nodes are not automatically independent operators. Sealing limits advance disclosure; it does not make public templates unrecognizable to workers. A script can pass simple echo checks, so these checks are not quality proofs. Reference-worker comparisons, media issuance, blind quality audits, and Base commitments are future lanes with separate rollout gates.

Raw prompts, raw outputs, challenge seeds, answer keys, and golden media hashes stay off-chain. Any future Base integration should carry compact commitments and reviewed economic state, not private workload content.

Why It Needs a Separate Role

The original network treated workers as both the compute layer and the trust layer. That works when workers compete on the same model and you can compare outputs across them, but it breaks for:

• Long-context generation where outputs are non-deterministic
• Image generation where bit-exact comparison is meaningless
• Specialized models where only one or two workers are running them at any moment

A separate validator role lets the network measure delivered capability without pretending every model label or quantization choice is directly provable. If a worker can solve the text task, follow the image or video contract, and match a certified deterministic workflow within tolerance, the network can reward that usefulness. If it repeatedly fails objective checks, routing reputation should drop first. Slashing is reserved for later, objective fraud cases with quorum and a dispute path.

Validation Lanes

Validators span text, image, and video, but each lane needs different evidence.

• Proof of usefulness — text reasoning, JSON/schema output, code tests, tool-use chains, image prompt constraints, and video duration or motion checks.
• Proof of fidelity — deterministic workflows with workflow hashes, model hashes, explicit seeds, dimensions, and pHash/SSIM/LPIPS-style comparison.
• Proof of honesty — max-token compliance, explicit seed compliance, dimension and duration compliance, usable outputs, signed receipts, and honest failure reporting.

For text, the Grid should care about measured capability more than whether a worker runs a specific quantization. For deterministic media workflows, the Grid can be stricter because fixed parameters and reference outputs make fidelity measurable.

Base Integration

The Grid uses Base for public, auditable economic state. Validator evidence should not put every probe or raw response on-chain. The practical on-chain surface is:

• validator registry and stake
• worker bond and future slash events
• epoch attestation roots
• deterministic workflow certificate roots
• validator reward roots

What You’ll Need

• A dedicated signing identity, created locally during setup; no funded wallet or personal private key is needed for the preview
• A reliable host with stable bandwidth
• CPU resources for V0 text canaries and signed evidence
• Optional GPU resources for future reference or re-execution lanes

Final stake size, reward schedule, and slashing conditions will be published with the staking and reward contracts.

Current Operator Preview

The evidence-only operator preview is live against production Core. The exact unsigned release is v0.1.0-preview.13. Its installer, four platform archives, checksum manifest, SPDX SBOM, and build provenance are published as one immutable payload. Linux x64/ARM64 users may also pull ghcr.io/aipowergrid/validator:v0.1.0-preview.13; prereleases do not publish latest.

The preview bundles the bounded image and video decoder dependencies on every supported platform. This makes the public artifact media-capable, but Core still issues no media assignments and media evidence has no economic effect.

Core 0029 contains the atomic budget and settlement terminal for future compensated quality audits, but it is deployed dark. No audit scheduler, private corpus, challenge selection, or quality scoring integration is live, and current validator assignments remain unpaid and economically inert.

curl -fsSLO https://github.com/AIPowerGrid/grid-validator/releases/download/v0.1.0-preview.13/install-validator.sh
gh attestation verify install-validator.sh --repo AIPowerGrid/grid-validator
bash install-validator.sh
cd ~/.aipg-validator
aipg-validator self-test
aipg-validator enroll
aipg-validator check --no-probe
aipg-validator run

Windows: extract the Windows x64 ZIP, then double-click aipg-validator.exe. Choose 8 to open the local operator app, then Set up node, confirm, and Start validator. Existing operators skip setup and keep their configuration. The preview is unsigned; verify its checksum and GitHub provenance before running.

macOS/Linux desktops: aipg-validator app opens the same local controls. The commands above remain the headless-server path. The app shows registration, acknowledged heartbeats, assignment activity, accepted evidence, and redacted diagnostics. Registration alone is not completed validation; waiting during assignment cooldown is normal. Native offline packaging checks pass on all four platforms, but complete live Windows qualification is still outstanding.

Stop stops the process the app started, preserving its signing identity and recovery journal. Closing only the browser tab leaves it running; Exit app also closes the local server. Local stop does not revoke a key or suspend its registration. Never share the private localhost URL or forward its port.

Automatic setup creates an empty local signing identity, authenticates a dedicated node account with the Grid, and saves a validator-only API key. It never asks you to paste a private key. Google and GitHub login are optional, and no funds or stake are required. Back up the private configuration file; do not post it or its contents in Discord or issues.

This does not pair the node with an existing Google or wallet account. Authenticated existing-account pairing is separate work. Existing operators should keep their current configuration when upgrading, not enroll a new identity. A successful registration alone is not a validation result: confirm a fresh heartbeat and accepted signed evidence after running.

The preview is deliberately limited:

• V0 evidence is useful for dashboards and engineering feedback.
• V0 does not prove exact model weights.
• V0 does not slash workers or pay validators.
• Production Core enables targeted text probing only through a Grid-issued assignment. Arbitrary worker targeting is not supported.
• Three first-party nodes prove the protocol and deployment path, but they share one operator and hypervisor. Independent operation is not yet proven.
• Validator rewards, stake enforcement, routing influence, slashing, and media assignments remain disabled.

AI Power Grid is recruiting 5-10 independently controlled operators for a 72-hour, no-reward qualification. One person or organization counts once even if it runs several nodes. See the preview cohort runbook and volunteer on the tracked cohort issue.

Future Economic Path

The later validator economy is straightforward in shape, but it should not be enabled until the evidence path is hard to game:

1. Grid issues short-lived assignments with a nonce and target worker.
2. Validators run the assigned probe and sign an evidence hash.
3. Grid aggregates attestations into scorecards and quorum outcomes.
4. Routing uses reputation cautiously before money is attached.
5. Accepted attestations could earn rewards after the economic and dispute gates ship.
6. Objective fraud may become slashable only after dispute tooling exists.

Join the Preview

Start with the tracked cohort issue so enrollment and the independent-operator gate remain visible:

Volunteer to run a validator

For community support, join the AIPG Discord:

discord.gg/W9D8j6HCtC

If you want to dig into the broader network design, see Proof of Intelligence, Architecture Overview, and Autonomous Network.