Agoranome Weekly: method v1.6

How every number in Agoranome Weekly is made, so anyone can check it or dispute it.

Definitions

Real demand. USDC payments on Base into the pay-to addresses of the 500 busiest x402 Bazaar sellers over 30 days, minus payments from catalog sweeps and from a seller's own pay-to address. Real buyers are the wallets with at least one such payment, each counted once.

Catalog sweep. A wallet that paid a seller in one session and, within an hour, paid 5 or more of the other sellers counted: a bot paying the catalog once, not a buyer. A wallet that paid a seller in two or more sessions, more than 30 minutes apart, is never a sweep there.

Not counted. A pay-to address that more than one seller lists (a platform's, whose payments can't be told apart), one that takes more than 20,000 transfers in 30 days (an exchange's), and Agoranome's own hosts and wallets. Every USDC transfer into a counted address counts, not only x402 payments.

Listings and sellers. Paid endpoints in the x402 Bazaar's index on the day of the week's snapshot (Mondays), and the distinct hosts that serve them. New sellers are hosts not in the week before's snapshot.

4-week survival. Of the listings that first appeared four weeks earlier (in that week's snapshot and not the one before), the share still listed this week.

Survival curve. The listings new in each of the eight weeks before (as for 4-week survival), and the sellers new in each (hosts in that week's snapshot and not the one before), pooled: of those whose week is at least n weeks back, the share still listed n weeks after they first appeared, for n from one to four. Only snapshots up to the edition's own week are read, so a final edition's curve never changes. A week whose snapshot or the one before is missing or empty starts no cohort.

Price index. p10, median and p90 of the USDC price per call across listings with 3 or more payers in 30 days, overall and for each category with 3 or more such listings.

Fair Offer share. The share of listings whose 402 offer carries a Fair Offer block, as the index keeps it.

Fair Offer share, read live. The share of the listings the Tuesday host scan asked, unpaid, whose 402 carried a Fair Offer block. Once the scan is in, the headline and the numbers under it use this one: the index's flag can drop the block.

Seller health. The same busiest sellers as real demand, as the Tuesday host scan read them: an unpaid request to each seller's 5 busiest listings and one picked at random, a failing one asked twice. A listing a buyer agent could pay answered with a 402 whose payment requirements decode. Price, pay-to and test-network counts are of those payable answers, against what the Bazaar index lists. A seller fails most of its sample when more than half of its listings asked couldn't be paid. Counts only: no seller is named.

SDK downloads. npm's and PyPI's public download counts over 30 days for the x402 SDK packages. Downloads include CI runs and mirrors, so they show how much is built, not who builds.

Concentration. The share of the paid calls the Bazaar index counts over 30 days that went to the top 1% of listings (rounded up, so at least one), and to the top 1% of sellers, each seller's listings summed. The index's call counts include catalog sweeps, so this is how concentrated calls are, not buyers.

Listing survival. Of every listing in the snapshot four weeks before, new or not, the share still listed this week: the nearest the Monday snapshots come to 30 days.

Monoculture index. In the newest Machine Choice Benchmark report, of the layouts two or more buyer-agent models answered (each sees the same offers in the same order), the share where every one picked the same offer, beside the share expected if each picked at random. Below 30 layouts the share isn't given. Beside it, how many of the panel's models moved their choices 25 points or more since the report before, and how many models changed between the two reports. When the models agree, a release can move a whole market at once.

Pay on delivery. Facilitators whose public /supported list, read on Tuesday, offers x402's auth-capture scheme on Base: a payment held until the route delivers, then captured or released. The facilitators read are listed on the method page; those that list it are named, the rest counted.

Our own indexes. Counts from Agoranome's own pages that Tuesday: the Buyer Safety Index's named clients, the admission feed's sellers, the MCP server report cards and the newest price-of-proof run, so each week keeps a dated record of them. Counts only: none is named here.

Entry rate. Of the sellers (hosts) in the Bazaar snapshot four weeks before and not in the one before it, the share whose listings show at least one payer this week, as the index counts payers over 30 days. The index's payers include catalog sweeps.

HHI. The Herfindahl-Hirschman index of the index's paid calls over 30 days: the sum of each listing's (or seller's) squared share. 1 is one taking every call; near 0 is calls spread evenly.

Cue load. The share of every listing whose description makes each kind of claim the Buyer check labels: a count, a standing claim, an endorsement, urgency, an authority claim, a pledge or an instruction to a buyer agent. Urgency and instructions are steering cues. A claim is counted, never judged true or false here.

Vendor agreement. In the newest Machine Choice Benchmark report, of the layouts that models from two or more vendors answered (each sees the same offers in the same order), the share where the vendors picked the same offer. A layout where one vendor's own models split is left out. Given from 30 layouts compared, overall and by category, beside the share chance alone would give. With it, Krippendorff's alpha (nominal, 1 minus observed over expected disagreement) between the vendors and across the panelists, and the effective number of independent buyers each is worth, m / (1 + (m - 1) alpha). A panelist that reads offers as claim labels sees the same layouts without the sellers' words. No vendor or model is named.

Cue advantage in rank. For each category, an exploded (rank-ordered) logit over the week's category searches on a directory's public search: rank 1 picked from every listing shown, rank 2 from the rest, and so on, on log 30-day calls, log price and whether the description makes each kind of claim. Gamma is a claim's coefficient: above zero, listings making it are placed ahead of listings with the same calls and price. Given with its 95% range, from five sellers with the claim and five without; four weeks above zero becomes a sentence. The directory is named only after notice.

Small cells. No share in the fairness section rests on fewer than five sellers: such a cell says "under 5 sellers" and no more, so no seller can be picked out.

Sources

The x402 Bazaar's public discovery index (Coinbase CDP), snapshotted every Monday; USDC Transfer events on Base, read from a Base RPC node every Tuesday; the Tuesday host scan's unpaid requests to the busiest sellers' listings and their security.txt files; npm's and PyPI's public download counts (pypistats.org), read every Monday; the Machine Choice Benchmark's raw trials, facilitators' public /supported lists and Agoranome's own index pages, read every Tuesday. Nothing is estimated or modelled, and no figure is a forecast.

Other trackers

Each edition sets real demand beside the counts other public x402 trackers showed the same week: what each one includes, and why it differs. Their figures (payments, buyers or dollars, as each shows them) are read by hand from their own pages, with the day read and the window they use, and are theirs: quoted with a link and credited, under the licence they name when they name one. No tracker's dataset is copied. A tracker not read that week says so. None is called wrong: most count gross, every payment that settled, and real demand is the part buyers paid for.

Gross and real

Other public trackers count gross: every payment that settled, with tests, catalog sweeps and sellers' own payments left in, and every package download. Agoranome Weekly leads with what's real: payments from buyers, net of sweeps and sellers' own, and what sellers actually do. Others' totals stay as credited lines, for context. The Bazaar's own seller and listing totals, and npm's and PyPI's download counts, are quoted that way, as one credited line each.

Seller safety table

Seller health's counts as columns, a row a week: JSON lines, CC BY 4.0. Beyond the section's counts it has how many TLS certificates the Tuesday host scan read, how many didn't verify and how many expired within 30 days of the reading, and how many of every Bazaar seller, not just the busiest, the daily safety checks read against OFAC's SDN list that week, and how many matched. It holds our own checks and OFAC's list only: Google Web Risk's results stay out of it, and so does RDAP, because their terms don't clearly allow them to be passed on. For the same reason, Seller health's Web Risk counts are shown on the edition's page with Google's notice, and left out of its CSV and the dataset. Counts only: no seller is named. A week is final once a later week has counts.

System health

Concentration and listing survival come from the Monday Bazaar snapshots. The Tuesday hosts job reads the rest at its end: the monoculture index from the newest Machine Choice Benchmark report, and the /supported lists of these facilitators, unpaid, and of the one the Lab settles through:

• https://x402.org/facilitator
• https://facilitator.payai.network
• https://facilitator.x402.rs

A facilitator that lists auth-capture on Base is named; one that doesn't, or doesn't answer, is counted. Our own indexes' counts are read from their pages the same day. Counts only: no seller, client, server or model is named.

From 9 November 2026 the same job scores the delivery odds (the chance a seller delivers at its listed terms, which the deliver feed gives each day). Each settled paid probe of the seven days before is paired with the newest odds kept for its host before the probe, at most 30 days old, so a probe is scored once and never against odds that already counted it. A probe counts as delivered when it answered 200 at the price it offered. The Brier score is the mean of (odds - outcome) squared; the prior alone (the category's base rate, or the cautious 0.8 before there is one) is scored on the same probes. The evidence weights move only after 4 weeks in a row in which the odds beat the prior alone (a lower Brier score) on at least 10 scored probes. Then the learning ledger (/machines/learning) moves one weight by at most a quarter, within published bounds, if that would have scored better on the last four weeks, and moves it back if the next two weeks score worse than the prior.

The research notes carry the choice model of buyer agents as kept that Tuesday: a conditional logit fitted to every priced benchmark decision, its market coefficients per standard deviation of each feature with 95% ranges, its log-loss on held-out layouts (one in five, by a hash of the layout) beside picking at random, and its calibration slope, the factor on every utility that best fits the held-out picks. The pick-odds fact answers only from 100 held-out decisions and a slope between 0.8 and 1.2.

Fairness

From method v1.6 (9 November 2026). Entry, cue load and the HHI come from the Monday Bazaar snapshots; the sweep share from the Tuesday chain count. The Tuesday hosts job reads the rest at its end: the base rates by category, the delivery odds' score, parity once the parity probe runs, and agreement between model vendors from the newest benchmark report. Each share carries its Wilson 95% range: for a count over the whole index it shows how far the share could move by chance alone at that many listings. No share rests on fewer than five sellers. The components are published side by side and never folded into one score, so each reader weighs them. The usage a new listing needs to reach the shown list comes from a research model that stays unpublished, so it isn't given.

Cue advantage in rank (from method v1.6): each Monday the insights engine runs every category's buyer searches on a directory's public search and keeps each result's rank. For each category the Weekly fits an exploded logit, the order of results read as a series of picks, on log(1 + 30-day calls), log price and one indicator per kind of claim the description makes, under a weak normal prior (sd 3); the ranges come from the curvature at the fit. Whatever a search gives every result alike cancels, so the category needs no term. Only listings the week's snapshot knows with a price are read; Agoranome's own are left out. No gamma is given from fewer than 5 sellers with the claim or without it, or from fewer than 5 searches. A claim whose range stays above zero for 4 weeks running becomes a sentence; the directory is named only after notice. A claim is counted, never judged true or false.

x402 100

The 100 x402 Bazaar sellers with the most real buyers over 30 days, each counted by the who-pays read in each of the 4 weekly snapshots before the rebalance, drawn on the first Monday of each quarter. The level is the basket's real payments over 30 days, chain-linked from 100: each week it moves by the same basket's change in real payments since the week before, and a new basket links in at its rebalance. Members are listed A to Z, never ranked. The first basket is drawn on 2027-01-04, and the editions show it from that week. The members are named because being in the basket is good news; each seller's own counts stay on its report card.

Versions

A change to the method is announced in the edition before it takes effect. Every final edition is then restated under the new version, and says so.

Corrections

A seller who thinks a count is wrong writes to research@agoranome.com. Corrections are logged here with dates.

Neutrality

No paid placement and no sponsors; a grant pays only for named public work and is in the open book. Agents pay for the Lab's tools per call, and nobody can pay to change a number. Sellers are never ranked here; a seller's own counts are on its report card. Agoranome sells tools to x402 sellers and buyers. Its own hosts and wallets are left out of every count, and no figure here is its revenue.

The newest edition