Agent-Infrastructure Index — Methodology

The source-of-truth for the methodology cited by every signed snapshot of the PayPerByte Agent-Infrastructure Index. Version 1.

methodology: byte-agent-index/v1 · url: payperbyte.io/agent-index/methodology

What the Index is

Three daily, on-chain-signed snapshots of the decentralized agent-runtime stack: Compute (GPU/compute an agent rents), Memory (storage/memory for agent state + RAG), and Tools (networks exposing an agent-callable tool surface). Each snapshot ranks a vetted top-8 of providers, scores them, and tags every metric with its provenance. The product is honest, sourced, timestamped measurement that compounds into a first-party time-series — not a relayed dashboard.

Inclusion bar

A provider is included only if it is (1) real and mainnet-live, (2) genuinely decentralized or honestly flagged with its decentralization tier/maturity, and (3) has at least one pollable data source we can track over time. Providers that fail (testnet-only, deprecated, presale, no public tool, auth-walled-only, vaporware metrics) go on a public watchlist with explicit re-entry triggers. Selection comes from an adversarially-verified research pass (one skeptic per provider, skeptic-default-exclude).

byte_score (0–100)

A composite of relevance-to-agents × data-trackability × legitimacy × decentralization, assigned in the verified research pass. v1 publishes that research composite (reproducible from our dataset). The compounding signal is score drift over time as providers ship tooling, reconcile metrics, or decentralize. Rank reflects the score with a documented data-tractability tiebreak where scores are within ~2 points.

Provenance — every metric carries it

Each metric is {value, provenance, source, as_of, unit?, computed?, basis?, flags?}. Provenance grades, strongest → weakest:

• onchain-verified · public-api · derived (computed from inputs we also store)
• dashboard-scrape · self-reported-unaudited · third-party-verified · internal-research

A metric with no live source this snapshot is listed in tracked_pending — never a silently-nulled value.

Honesty rules (the differentiators)

• Active, not cumulative. Counts publish the concurrent-active figure; the inflated registered/cumulative count is published separately and flagged (e.g. Akash 65 active vs 1759 registered).
• We compute the honest number. Utilization etc. are derived from raw inputs we also store, never the provider's self-report; the computation's limitation is flagged.
• The LLM never originates a number. Deterministic adapters produce metrics; local-first inference only normalizes scraped text, and anything it touches is tagged dashboard-scrape/self-reported-unaudited.
• No vanity metrics as fact. Self-reported headline figures are flagged, not republished as verified.
• Confidence is data-quality, not network health. Economic risk is carried in caveat flags, not in confidence.

Refresh

Daily signed snapshot, internally tiered: daily on-chain/price core · weekly supply/pricing/liveness · monthly scoring + decentralization re-rate · event-driven launch/death flags · quarterly full re-verification (promotes watchlist names, demotes decliners). Probe spend is capped; a failed poll emits scrape-failed rather than a stale value.

Sourcing & licensing posture

Built on on-chain RPC + open APIs + MIT-licensed explorers wherever possible; vendor dashboards and key-gated APIs are derive-and-attribute only. Token price is on-chain via Pyth (attested on Pythnet, served through Hermes with a signed update); FDV is price × on-chain totalSupply (labelled FDV, never “market cap”); market cap + circulating supply stay pending a non-commercial-licensed source. There is no CoinGecko anywhere. Provider metrics read directly from the underlying chain — e.g. Render burn/mint from Solana, Autonomys from its public substrate RPC — so no vendor license is required. Per-row source + licensing_risk make provenance auditable if terms change.

Not advice