MEI Methodology

The Machine Economy Index (MEI) is a composite index — in the same family as a stock-market or economic-health index — that tracks the growth and vitality of the machine economy on a 1–100 scale. It is published on a fixed schedule, and every score is stored with a methodology version tag so readers can compare like with like.

The index measures four components. Three are infrastructure rails: Payment (how machines pay each other), Physical (the infrastructure machines use to do work), and Legal (the frameworks machines operate within). The fourth is Macro — realized demand, captured through enterprise AI-use rates and developer adoption. The three rails are the analytical framework; the index measures the rails plus the demand component.

Aggregation follows a three-level structure. The index is built from 14 scored metrics — 4 on the Payment rail, 6 on the Physical rail, and 4 on the Macro rail. Individual metrics are combined arithmetically within each of those three rails — category means first, then a rail mean. Within each rail, all included metrics are weighted equally: each metric contributes the same amount to its rail score, and the rail score is the simple average of all normalized metrics in that rail. Equal within-rail weighting is transparent and reproducible — no hidden per-metric weights — and avoids overfitting to current conditions. The four top-level components are then combined geometrically. The geometric step operates only across those four components; within a rail, everything stays arithmetic.

The Legal rail carries no scored metrics; it is represented directly by the Legal Rail Readiness Score (LRRS), a separate composite with its own coverage model — explained in its own section.

The four components are aggregated as a weighted geometric mean with an elasticity of substitution of 1 and a floor of 1 on the [1, 100] scale. That floor ensures no component can collapse the composite to zero; the geometric form does the rest of the work.

The geometric mean is chosen because the framework treats the rails as required layers, not substitutes for one another. A simple arithmetic average lets a strong component mathematically compensate for a near-absent one — a high Payment score can pull up a composite even when Legal readiness is close to the floor. The geometric mean does not allow that: a low component pulls the index down more than it would under an average. Because a machine economy needs all these layers present, the index uses a geometric mean, which — unlike a simple average — does not let a strong area paper over a weak one.

For transparency, the geometric and arithmetic means are published side by side. Their difference is the balance gap — a diagnostic of how uneven the four components are. A larger gap means more imbalance between rails and demand. The component bars above show this live: the geometric headline, the arithmetic reference, and the gap between them.

Each of the four components receives equal weight — 0.25 — stated explicitly as a null hypothesis. Rank-sum weights are used for LRRS categories because a defensible importance ordering exists among legal instruments. Equal weights are used for MEI components because not even an ordering is defensible across Payment, Physical, Legal, and Macro — any ranking would be contestable, and geometric aggregation already enforces that no component is dispensable. That null-hypothesis argument stands on its own; it does not depend on the sensitivity analysis below.

Robustness band (§6.7). The headline is accompanied by a robustness band — not a measurement-error confidence interval, but the 5th–95th percentile range from 10,000 Monte Carlo draws over component weights (uniform on the simplex, truncated to [0.15, 0.35] per component) and normalization goalposts (log-uniform [0.5, 2] scale on each metric's bounds). LRRS has no goalposts — its underlying is already 0–100 coverage. At current values the fixed weight grid (equal weights, all 24 permutations of the 35/30/15/20 vector, plus ±10pp shifts) spans 28.2–37.6 (width 9.4 points); the Monte Carlo robustness band is [29, 37] (width 8.5 points). R2 (band width ≤10): passes. R3 (weakest rail unchanged in ≥95% of draws): legal in 100.0% of draws. R1 requires 30 days of post-launch history; skipped until sufficient series exists. Live values: GET /api/v1/mei (score_low / score_high).

Every major index is opinionated and declared by its creators, then trusted because it is transparent, consistent, and independent. The S&P 500 reflects S&P's decisions about which companies belong and how to weight them. Gartner invented the Magic Quadrant's axes. The Net Promoter Score's creator chose the 0–10 scale. The UN's Human Development Index weights life expectancy, education, and income equally. None of these were democratically decided. The MEI follows the same model and goes further: its methodology is not just published but derived — every parameter traces to a stated premise, and the one unavoidable judgment (equal weighting) is labeled as a null hypothesis rather than disguised as knowledge. When the methodology is refined, a new version is published with a full explanation, and old scores are retained.

Every metric is normalized to [1, 100] before it enters aggregation. Two regimes apply, depending on the shape of the underlying series — the bounds table below shows which each metric uses.

Log-space min-max is used for unbounded or wide-range metrics — volumes, counts, capacity — where values span orders of magnitude. The log transform makes normalization unit-invariant and appropriate for quantities that grow exponentially rather than linearly. Linear min-max is used for naturally bounded metrics such as ratios and percentages, where the raw value already lives on a finite scale.

All bounds are published and versioned. The table below shows every metric's floor, ceiling, unit, and regime, tagged with boundsVersion 1.0. This is the replicability commitment — anyone with the raw data can reproduce the normalization step.

The “launch N₀” column shows where each metric sits at launch on its 1–100 scale. These positions are a deliberate snapshot calibrated against verified anchor values — not an accident, and not an implication that the metric is “good” or “bad.” Some metrics launch low (nascent) and some high (mature) by design; the bounds are set to capture realistic future growth. Per-metric specifics — floor, ceiling, cadence, and source — are in the table.

Every metric must pass six tests before it becomes an MEI input. These criteria were applied strictly during the v1.0 redesign — only a fraction of the metrics considered during development qualified; the rest were excluded as named gaps or contextual signals rather than index inputs.

• Construct validity (the “freeze test”): the metric must genuinely measure machine-economy activity — if the machine economy froze, the metric would stop moving.
• Primary-source verifiability: the value must come from a primary, verifiable source — on-chain data, a regulator, an official statistical agency, or a company's own filing — not an aggregator or estimate.
• Reversibility: the metric can go down as well as up; it reflects real current activity, not a cumulative count that only grows.
• Single-category assignment: each metric maps to exactly one category, with concentration disclosed — no double-counting.
• Non-redundancy: the metric adds information not already captured by another metric.
• Defined cadence: the metric has a known update frequency — daily, biweekly, quarterly, or annual — shown in the bounds table.

Some dimensions of the machine economy do not yet have a metric that passes all six inclusion criteria — or cross a disclosed construct boundary the current proxy cannot reach. Rather than filling those slots with weak proxies, the MEI and LRRS name them explicitly in one register. Each entry states what Tier-1 source would close it.

MEI — Payment Rail

• Closed-rail machine payments — verifiable transaction-level data from private settlement rails (cloud billing, API metering). None exists in public infrastructure today.
• Non-USDC x402 modalities — a recurring primary-source series for x402 V2 rails beyond on-chain USDC (USDm, ACH, card).
• Quality-adjusted x402 volume — a free, recurring, reproducible wash-filter methodology. Gross on-chain USDC is the v1.0 input by design.
• Lightning / L402 machine payments — L402 (Bitcoin Lightning) is a real, active machine-payment protocol. We cannot measure it: Lightning payments settle off-chain by design, so no public, verifiable volume series exists. This is a construct boundary, not an omission. It closes if a Tier-1 verifiable series ever emerges.
• Settlement liquidity — machine-payments-specific stablecoin float, distinct from total stablecoin supply (fails the freeze test).

MEI — Physical and Macro

• Captive silicon — hyperscaler-designed accelerators form real AI compute capacity our merchant-market proxy cannot see. Closes with a vendor-neutral, AI-specific SEC-grade series (logged candidate: TSMC HPC-platform revenue).
• Energy / machine-economy power draw — no Tier-1 high-cadence official series on AI/datacenter electricity consumption exists. That absence is why it is a gap, not a metric.
• Embodied machines / robotics — no Tier-1 recurring deployment count for autonomous machines. Robotics remains contextual until such a source is confirmed.
• Realized agentic commerce volume — actual transaction volume of autonomous agentic commerce. Current figures are vendor estimates.
• Agentic labor demand — no official labour body publishes an agentic-AI occupational series; the taxonomy does not yet exist.
• Metered machine-services consumption — pay-per-use consumption of machine services (e.g. LLM throughput). On the verification queue; no confirmed free API at v1.0.

LRRS — construct and coverage boundaries

• Digital-asset property law — a disclosed v1.0 construct boundary, not a sixth category. Instruments such as the DIFC Digital Assets Law map to no existing LRRS category.
• Constituent-state watchlist — declared non-coverage boundary (Ruling #7). US states, provinces, and Länder are not sovereign-designated zones; their instruments are excluded and tracked, not scored.
• Multilateral machine-economy governance — UNCITRAL's Model Law on Automated Contracting (adopted 11 July 2024; UNGA-endorsed) addresses machine contracting directly: a contract formed by an automated system may not be denied validity or enforceability on the sole ground that no natural person reviewed or intervened in any action carried out in connection with the formation of the contract. Zero states have enacted it. The category therefore scores zero — not because the law does not exist, but because no legislature has passed it.

These are real parts of the machine economy the indices do not yet measure. Naming them is a credibility choice: the platform states what it cannot yet quantify rather than substituting unverified numbers.

The LRRS category bars in the LRRS section reflect live coverage — categories at zero indicate no jurisdiction in scope yet has an operational framework for that category.

On-chain-only Payment Rail. The Payment Rail measures publicly verifiable on-chain machine-commerce activity — for example on-chain USDC x402 volume. Off-chain or non-public machine payments (including Lightning / L402) are not captured. This is a deliberate scope choice: verifiable on-chain data over unverifiable off-chain estimates.

x402 gross on-chain volume and wash/bot disclosure. The x402 volume metric uses gross on-chain USDC volume. Some of this may include wash trading or bot activity. The MEI discloses this as a concentration and quality risk rather than applying an opaque “quality adjustment” — because gross on-chain volume is primary-source verifiable and reproducible, whereas any adjustment factor would be an unverifiable judgment. Transparency about a known risk over an opaque correction.

Payment-rail concentration at composite level. At launch the Payment Rail is substantially a two-protocol rail at the MEI composite level: x402 metrics account for roughly 16.7% of the headline composite via a single protocol source; ERC-8004 registry activity accounts for another ~8.3%. Together they dominate the 25%-weight Payment component.

OECD-scoped connectivity. Machine connectivity is measured via OECD M2M subscription data, which covers OECD member countries — not global. The series is annual with an 18–30 month publication lag (~2-year phase lag: a 2026 IoT inflection registers around 2028). As a stock (point-in-time subscription count), a machine-economy freeze would plateau the metric rather than drop it — the same stock-behaviour caveat applies to Filecoin raw byte power.

Nvidia Class-M exception. The Physical Rail's compute-capacity metric uses Nvidia's SEC-reported Data Center segment revenue (not consolidated company revenue — gaming and other segments are excluded by construction) as a proxy for merchant-market AI compute capacity formation, not total AI compute capacity. Captive hyperscaler silicon (TPU-, Trainium-, MTIA-class) forms real capacity that this merchant-market proxy cannot see. A Class-M exception to the usual primary-source-per-metric rule — the best available verifiable proxy today. Quantified wrong-direction sensitivity: if true AI compute capacity rises 30% while Nvidia's merchant share slides 90%→65%, Nvidia revenue falls ~6% and the normalized score drops ~1.5 points (68.0→66.5 on published bounds) even as true capacity rises — a disclosed construct boundary, guarded by a pre-committed 60% share construct-review trigger (monthly sweep). It launches at a relatively high position (mature) and an early bounds rebase is anticipated by design as the metric matures.

GDP vintage. LRRS GDP weights use a frozen IMF WEO vintage, tagged per methodology version, so the weighting is reproducible and does not drift between versions.

LRRS scope. The LRRS universe is 48 jurisdictions covering the large majority of world GDP: 30 are actively monitored, and 18 EU member states are included so EU-level instruments can propagate — not every country. The coverage bars' ceiling line reflects this in-scope GDP total.

Launch positions are designed. The index launches with metrics at calibrated positions — some nascent and low, some mature and high — by design. These are a deliberate snapshot, not a judgment that a metric is good or bad.

Nominal-bounds drift (USD log metrics). Three metrics use log-space normalization on nominal USD bounds: Nvidia Data Center segment revenue, x402 on-chain USDC volume, and Akash compute spend. Bounds are nominal by design; rebases absorb structural shifts. They do not adjust for inflation — at ~3% annual inflation, normalized scores drift upward from price level alone even if real activity is flat: approximately +0.68 index points per year on Nvidia, +0.32 on x402 volume, and +0.24 on Akash spend (quantities held constant).

Cadence and staleness mix. Metrics update at different cadences, from daily to annual. Between updates, the most recent verified value carries forward. Slower-cadence metrics — for example quarterly or annual series — are inherently less fresh; that mix is disclosed rather than hidden. The LRRS is different: milestone cell statuses are hand-curated and verified on a monthly sweep, not fetched on a timer — so legal coverage may lag a regulatory event by up to one sweep interval.

Robustness band. The headline is reported as a point estimate plus a robustness band (e.g. MEI 32.6 [29, 37]) — the 5th–95th percentile range from the §6.7 sensitivity analysis over weights and normalization goalposts. This is not measurement uncertainty; it states how much the score moves under defensible alternative methodological choices. Values are stored as score_low / score_high on each daily MEI API response and shown on the Index page.

Versioning policy. The methodology is version-tagged (currently 1.0). When refined, a new version is published with a full explanation, and old scores remain stored under their original version — so the index is never silently re-baselined. This follows the standard practice of major index publishers.