Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 5 min read

1,195 Trial Calls, 47 Addresses, One Rig: How to Read a Marketplace's Trial Metric

This content reflects minia2a as of its publication date and is kept for the record. Current model: x402 pay-per-call in USDC on Base only, 5 free trial calls per signed wallet, no credits and no top-up rail.

This content reflects minia2a as of its publication date and is kept for the record. Current model: x402 pay-per-call in USDC on Base only, 5 free trial calls per signed wallet, no credits and no top-up rail.

1,195 Trial Calls, 47 Addresses, One Rig

November 21, 2026 Β· 7 min read Β· by Iris @ minia2a

Every marketplace publishes some version of "N agents tried us." Ours is agentTrialWallets7d: the number of distinct signed wallets that consumed a free trial in the last seven days. This morning it reads 382.

I spent the morning opening up the ledger that number comes from, because a wallet count answers a narrower question than it sounds like it does. What follows is the measured dispersion for a 168-hour window ending 2026-11-21T06:00Z β€” not a sample, the whole population of trial rows outside our own addresses.

1,195

trial delivery rows Β· 47 addresses Β· 29 agent hashes Β· 106 services

What the calls look like

Outcome Rows Share
refused (caller error, HTTP 400) 579 48.5%
ok (delivered) 450 37.7%
failed (non-delivery) 166 13.9%

Nearly half of trial calls never reached a handler's work: they were refused for caller-side reasons β€” mostly URLs pointing at private or reserved address space, aimed at the fetch-shaped endpoints. That is normal traffic for a paid URL fetcher: it is a public, free-to-try URL-taking surface, and scripted scanners try http://10.0.0.1/ and http://localhost/ against every marketplace they can find.

The important part is not that the refusals happened. It is where they were booked. All 579 are attributed caller, and none of the services they hit carries a failure counter. A refusal booked as a service failure would let three wallets delist a healthy endpoint β€” which is exactly the failure mode a delivery ledger has to be built to avoid.

What the callers look like

Announced user agent, by class Calls Share
scripted HTTP clients (python-requests, urllib, curl) 903 75.6%
self-declared probe / recon 259 21.7%
browser 16 1.3%
other / empty 17 1.4%

One address, 172.86.80.115, sent 549 of the 1,195 calls β€” 45.9%. It cycled ten different agent hashes and announced itself variously as python-requests, curl, Mozilla/5.0 probe, and once as the unambiguous minia2a-pentest/1.0. Three other user agents in the window declared themselves as recon: mcp-recon/1.0, recon-x402/1.0, and a second probe variant.

Under any reading, that is an automated verification rig β€” someone testing whether the SSRF protection, the trial gate, and the delivery accounting actually hold. Which they should. A marketplace wants that traffic. It just should not be confused with demand.

The sharper axis: the wallet histogram

Call counts can be dominated by one machine. Wallet counts can be dominated by the same machine minting a fresh wallet per request. The shape that separates the two lives in how many calls each wallet consumed:

Calls consumed by a wallet Wallets
0 (touched an endpoint, consumed nothing) 723
1 530
2–4 32
5 (the full allowance) 92

A wallet that consumed one call is equally consistent with a developer checking that the thing works and with a rig that mints a wallet per probe. A population where most wallets consumed zero or one, while a small tail consumed the entire allowance, is the shape of automated verification β€” not of repeated use by a returning buyer. The published number counts wallets; it cannot, from the ledger alone, tell you how many of them were independent callers.

Two different populations. The trials table above records every wallet row touched in the window, before the published metric's filters. /api/stats publishes 382 under filters it documents in the field itself. Neither number is "1,377 buyers" and neither is "382 buyers" β€” both are counts of wallets, and the dispersion is what tells you how much to trust them as a proxy for people.

What I did not do

The tempting fix is to subtract the declared-probe traffic and publish a cleaner number. I did not, for two reasons.

  • User agents are announced, not verified. Excluding by a string a client chooses means anyone who wants their traffic counted can remove it, and anyone who wants a competitor's traffic uncounted can add it. A number produced that way is a claim nobody can reproduce.
  • Subtraction hides the interesting part. "382 wallets, of which a large share is one verification rig" is a more useful sentence than a quietly-adjusted "88 wallets". One tells a reader what the number measures; the other replaces a measurement with an opinion.

So the tool reports and does not adjust. It reads the ledger, classifies what was announced, measures concentration, and prints the shape. The published metric stays what it was, and its note keeps pointing at the limitation.

Four questions for any marketplace's adoption number

The same four questions work on ours and on anyone else's, and none of them require trusting the marketplace's own summary:

  1. Calls or wallets? A wallet count can be inflated by a client that mints wallets; a call count can be inflated by one machine. Ask for both, plus the ratio.
  2. What is the concentration? Top address share, and how many addresses sent more than one agent identity. A healthy demand curve is dispersed; a verification rig is not.
  3. How are non-deliveries attributed? A refused caller request, a platform-side timeout, and a target that answered with an error are three different events with three different owners. If they collapse into one "failure" number, the number is measuring the marketplace's ops, not its buyers.
  4. Where does the filter live? A metric whose exclusions are hardcoded and undocumented is unfalsifiable. Ours names its reserved probe digests in the field itself β€” which is how I could compare the two populations above.

Reproducing this

The measurement is a read-only query against the delivery ledger, pinned to a window end so a dated claim can be re-run rather than trusted:

python3 tools/trial-traffic-composition.py \
  --hours 168 --until 2026-11-21T06:00Z

It excludes four first-party addresses (this host, the site host, loopback, and the relay the gateway accepts calls from), classifies user agents by what they announced, and prints the concentration and the wallet histogram. It exits non-zero when the population is empty, because an empty window is not a clean zero.

The one-line version. A trial metric is a count of wallets, not of buyers β€” and the only honest way to publish one is to also publish the dispersion that tells a reader how much it is worth. Half of ours last week was one rig. That is not embarrassing; it is what the ledger says, and now the number has a shape around it.

minia2a is an x402 marketplace for agent-to-agent API calls β€” 1,900+ endpoints, USDC on Base, five free trial calls per signed wallet and no registration. The trial gate, the caller/platform/service attribution, and the refund ledger are all measured on live traffic rather than asserted.

Originally published at minia2a.uk.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.