Whitepaper · v1 · June 2026

Basis: the claim verification layer for the agent economy

basis.watch · web-native document · sections 10–12 describe research, not shipped product

Abstract

The AI agent economy on Base is growing fast, and its public metrics are hard to compare. Agent counts, repo counts, processed volume, and GDP figures are mostly self-reported, sybil-sensitive, or framed with the most favorable denominator available. Basis is a public verification registry that separates claimed activity from verified activity: it labels project-made claims against public evidence, scores claim reliability in the open, adjusts populations for sybil risk with published heuristics, anchors every report's hashes in an append-only onchain registry, and serves the same evidence to humans, developers, and autonomous agents. This document describes the system as shipped, the methodology that constrains it, and — explicitly separated — a researched economic model that has not launched and may never launch.

1 · The problem: claims without verification

In Basis's first audit, the pattern was consistent: claims that are onchain-checkable or anchored in primary documents survive scrutiny, while claims about populations and volumes almost never do. Of the 41 claims examined across 9 prominent Base AI projects, 15 verified, 11 partially verified, 11 remained unverified, and 4 were disputed — contradicted by primary evidence, in several cases by the claimant's own primary sources.

This is not an accusation of bad faith. Counter-style metrics inflate by default when creation is free and verification is absent. But the consequences are real: agents route payments to services based on unverified reputations, builders choose dependencies on self-reported traction, and capital prices narratives instead of evidence. Every market matures the same way — by growing an independent verification layer. The agent economy does not have one. Basis is built to be it.

2 · Market context: Base AI, x402, MCP

The structural conditions for a verification layer arrived in 2025–2026. x402 made machine-to-machine payment a protocol primitive; agentic transactions on Base grew past 100 million cumulative by Q1 2026 (Chainalysis), while independent analysis attributed roughly half of peak activity to incentive farming — both facts matter, and holding both at once is precisely the job of a verification layer. MCP became the standard surface through which agents consume tools and data. Agent identity standards (ERC-8004) and commerce standards (ERC-8183) are deployed or drafted but young, with registration counts that say nothing about live capability.

The result is an economy where the rails are real, the activity is partly real, and the headline numbers are systematically unreliable — documented in detail, with sources preserved, in the first Basis audit.

3 · The claim verification layer

Basis treats a claim — not a project, not a token — as its unit of work. A claim is a falsifiable public statement: "42% of supply burned," "33,000 agents," "$500M processed." Each claim is recorded with the claim text as circulated, the claimant, typed and dated sources, prose evidence, a status label, and a confidence grade describing the robustness of the evidence process itself. Records are stable, addressable, copyable as JSON, and re-checked on a published cadence. The registry is public by construction: no accounts, no paywalled labels, no private reports.

4 · Evidence taxonomy

Five statuses — verified, partially verified, unverified, disputed, stale — with strict definitions and a source-ranking discipline: onchain evidence over primary documents over independent secondary reporting over self-reported figures. Two principles do unusual amounts of work. First, unverified does not mean false: it means public evidence does not yet support the claim, and the correction process exists precisely to receive that evidence. Second, labels critique claims, not people: a disputed label is a statement about evidence conflict, never about intent. The full taxonomy is published in the methodology.

5 · Claim Reliability Index

The CRI compresses one question into one number: when this project makes a public, checkable claim, how well does it hold up? Only claims made by the project itself are scored — third-party findings and Basis context never move a score. Status and confidence weights, the disputed penalty, and grade bounds are published constants; changing them is a versioned methodology event with a public diff, never a quiet edit. The CRI is explicitly not a product-quality score, not a safety rating, and not an investment signal.

6 · Sybil-adjusted metrics

Counter-style metrics are treated as sybil-sensitive by default whenever creation is free. Basis computes four published component heuristics — young-wallet share, funding-graph concentration, retention collapse, and timing regularity — calibrated against the documented October 2025 x402 farming wave, and reports adjusted figures alongside raw ones, never instead of them. Heuristics are floors, not proofs; the methodology says so and the limitation is part of the published record.

7 · Onchain anchoring and correction history

Every report and its underlying dataset are canonicalized (recursively sorted keys, UTF-8) and hashed with SHA-256; hashes are anchored in BasisReportRegistry, an append-only registry contract where corrections create new versions and revocation marks — never deletes — prior ones, with the reason onchain. Anyone can re-derive the hashes from published files. Corrections follow a public process: evidence in, graded by the same rules as everything else, acknowledgment within two business days, versioned outcome in a public changelog. The correction process is not an escape hatch; it is the product's core credibility mechanism.

8 · Distribution: API, MCP, x402

The registry is machine-native. A free API serves claims, projects, and reports without accounts or keys. An MCP server lets any agent run trust checks inside its own loop — before a payment, an integration, or a citation. A premium route speaks x402: HTTP 402 challenges are live today, with USDC settlement enabling when the production facilitator is configured. Agents consume exactly the evidence humans read, under exactly the same disclaimers.

9 · Economic model (current, shipped)

Basis runs on USDC: per-call premium API pricing, verification and diligence engagements, and API licensing as coverage grows. Two structural rules protect the product: evidence standards are identical for customers and non-customers, and commercial relationships are disclosed wherever they could matter. The open registry is the public good; paid work funds it; nothing about a label is for sale at any price.

10 · The planned token — not live

The design question is narrow: open demand for verification work to anyone — including autonomous agents — without corrupting the work. The only design that survived independent tokenomics, game-theory, legal, and integrity review is deliberately boring: a fixed-supply work-request token. Burning it files a public audit request (a question, not an accusation) into a public queue with capped, logarithmic priority weighting. A hard-floored majority of capacity remains editorial. Every funded record carries permanent disclosure. Corrections stay free, forever. The scoring pipeline contains no knowledge of payments — enforced in code, in public. And the system is designed to be indifferent to token price: if the token went to zero, the registry degrades to exactly what exists today.

A distribution path under evaluation — not announced, not a partnership — is the Virtuals Protocol agent ecosystem on Base: staged entry as an agent profile and ACP service provider first (verification lookups and diligence intake against the existing registry), with any tokenization decision gated behind the public position-change process, counsel review, and a standing conflict-disclosure regime — because Basis audits Virtuals, and says so. Venue evaluation under a disclosed conflict is itself a test of the methodology.

The rejected designs are part of the record because the rejections are the firewall: no pay-per-label, no revenue routing or buyback-burn, no token-gated access or corrections, no governance over findings, no airdrops to covered projects, no prediction markets on forthcoming labels. The full analysis, including attack scenarios and open operator decisions, lives at basis.watch/token and in the public research document.

11 · Governance and conflict minimization

Basis's governance principle is asymmetric by design: process can be governed; findings cannot. Methodology constants — status taxonomy, CRI weights, source grading, sybil heuristics — change only through versioned methodology events with public diffs. If parameter governance ever exists (queue weights, bond sizes), it operates on an enumerated allowlist with findings constitutionally out of scope. Conflict rules are published: no treasury positions in covered assets, disclosure of any commercial relationship in affected reports, and a publisher-key custody plan that migrates to a multisig before mainnet anchoring.

12 · Roadmap

Five phases, each gated on the integrity of the previous: the launch registry (live), coverage expansion across projects and x402 sellers, marketplace integrations that put claim checks inside agent loops, a researched tokenized verification market, and long-horizon protocolization — attestations, reproducible evidence bundles, and a reviewer network. Status labels distinguish live, planned, research, and draft; the roadmap page is the canonical version.

13 · Risks and limitations

Basis publishes its own failure modes. Heuristics can be gamed by sophisticated actors; the methodology labels adjusted figures as floors, not proofs. Coverage is finite and selection is editorial; the registry records what it has examined, not everything that exists. Sources can disappear; check dates are displayed and re-checks are scheduled. A single-operator registry carries key-person and capture risk; the protocolization phase exists to retire it. And any future economic mechanism carries the risks documented in §10 — which is why it remains research, fenced behind public review gates, rather than product.

14 · Conclusion

The agent economy will be load-bearing infrastructure; its claims deserve load-bearing verification. Basis's bet is that the credibility produced by labeled evidence, open methodology, anchored history, and free corrections compounds — slowly, then decisively — into the trust layer agents and humans both route through. The registry is live. The record is public. The numbers have a referee.

basis.watch · 2026-06-09 · report hash 0x68d9c398d05bfaa9… · this document is versioned with the repository and supersedes nothing in the published methodology