State of Agent Money: October 2026

Three data checks were added: 37 → 40 checks. A step in the health score on this date comes from the new checks, not from a change in the data.

/data-health

retry_storm and mint_burst moved from hour buckets to runs (rule v2), and the layering_fan detector was added. 35 storm runs and 71 burst runs the old rules split or missed were added to the log at about 14:40 UTC, so detections_logged stepped 610 → 716.

/changelog-october-2026

The counterparty label map was updated: newly named addresses moved existing payments into categories. Swaps stepped 662,460 → 927,257 USD and the labelled share 0.505 → 0.525 in the next run. No new payments caused the step.

/changelog-october-2026

L8 (seen paying and still funded with $10 in stablecoins) rose 384 → 482 → 584 in the payment-scan runs of 09:48 and 15:43 UTC on 12 Oct 2026, which promoted 105 and about 100 agents to L8 each (the runs around them promoted 3 and 2), while L7 stayed at 7,757–7,764. The L8 rule did not change (rank.l8plus.v2 until 13 Oct); the runs re-read wallet floats on the day the outside-review fixes shipped, so the count caught up with balances it had been missing. Registered late, on 24 Oct 2026; compare from 13 Oct.

/def/rank.l8plus.v2

One data check was added: 40 → 41 checks.

/data-health

A data check for lagging chain watchers (chain_lag, warn) was added: 41 → 42 checks. The health score fell below 1.00 because a real lag (Arbitrum) is now counted, not because data got worse.

/changelog-october-2026

A data check for unexplained steps in our own archived series (unexplained_step, warn) was added: 42 → 43 checks. It reads the list at /api/steps; a step there is explained by an entry in this register or by a checked note.

/api/steps

Poisoning rule v2: a zero-value or other-token plant now counts when the look-alike matches a payee on the first 3 and last 4 hex characters (the threshold of Tsuchiya et al., USENIX Security 2025), not only the first 4 and last 4. 1,438 stored zero-value plants on 539 wallets were re-matched (3 of them in the 24 hours before the change; in the 7 days before it they fell on 38 wallets, some already counted as targeted). No new loss followed.

/def/poison.events.24h.v2

Endpoint-domain verification stops counting profile links on shared public platforms (github.com, x.com and similar) and reserved example names, where no agent can publish a proof. Only the first sweep (21:57 UTC on 12 October) included them.

/def/probe.domain.checked.v1

The drain rule also reads native ETH. An ETH send from the bound wallet of an L7+ agent opens a drain detection when it is at least 5x the wallet's largest earlier outflow (counted stablecoin payments and earlier ETH sends together), at least $250, to an address the wallet never sent stablecoins or ETH to, after at least three earlier outflows. Detections on stablecoin payments are unchanged, and counted spending still excludes ETH (eth.out.usd.30d.v1 is published on its own).

/def/eth.out.wallets.30d.v1

26 labels added to the counterparty book (48 to 74), each confirmed against its verified source on Sourcify: swap routers and approval contracts (AggregationRouterV6, MetaAggregationRouterV2, AllowanceHolder, Permit2, GPv2VaultRelayer, TokenApprove, UniversalRouter, SwapRouter02), swap pools and adapters, and one lending pool. They move 3,148 counted payments ($228,957, 0.8% of counted spending) by 504 agents from unlabelled to swap or lending. The category series step on the next counterparty run; compare them from 14 October.

/def/cp.labelled.count.v1

Two detectors added. approval_to_plain_key, read every two minutes: a non-zero approval from an agent wallet to a spender with no contract code, on the seven priced tokens or through Permit2. reputation_farming, hourly from 10:04 UTC: bursts of reviews from single-use wallets, and wallets that leave 100 or more reviews for one agent in a day. On its first run it recorded the 11 October "starred" wave (4,721 reviews from single-use wallets) and a 15-wallet "miner-vouch" campaign (69,062 reviews for one agent since 13 September). Both are findings for a human to read, not incidents.

/detections-ai-agent-anomaly-log-mint-bursts-retry-storms

Correction to cp-2026-10-13-new-detectors: the repeat-rater shape of reputation_farming first summed only the last 30 days, so a campaign's total shrank as old days left the window. From 11:03 UTC it totals the whole history (the lookback only decides which campaigns are still reported). The campaign first reported as 15 wallets and 69,062 reviews is 32 wallets and 373,329 reviews for one agent between 16 May and 9 October 2026.

/detections-ai-agent-anomaly-log-mint-bursts-retry-storms

Ladder v2. L7 needs a counted payment that left the agent's control: not into a lending or vault position the receipt shows it still holds, not to one of its own addresses (a wallet ever bound to it, its registration's owner or minter), not a bridge leg that came back to it. L9 needs three independent counterparties (contract code, or paid by two or more agents) among those payments. Counted at 12:53:32 UTC under the old rules and at 12:54:50 UTC under the new ones, with no new payments between: L7+ 7,866 → 7,525, L8+ 562 → 557, L9 118 → 74. 341 agents whose payments all stayed in their own control left L7, and 4 agents shown at L6, a level v1 gave to withdrawn ranks without measuring it, went to their measured probe level. Counted payment value is unchanged: this moves ranks, not money.

/ai-agent-liveness-methodology-l0-l9-ladder

Correction to cp-2026-10-13-ladder-v2, whose text stands: ladder v2 also moves the survival series computed from the ladder, which that entry did not name. Their first run after the switch (13 Oct 22:50 UTC) stepped: survival_paid_365d_single 0.318 → 0.270, survival_paid_365d_all 0.085 → 0.082, survival_median_days_first_pay 0.099 → 0.092; survival_paid_365d_burst1000 did not move. Registered late, on 24 Oct 2026, after the step check flagged the single-registration series.

/def/rank.l7plus.v2

Ladder v2 on the probe side, first sweep at 13:01:50 UTC. L5 reads up to 8 web endpoints in any order (v1: the first 4 listed entries, ENS names and DIDs included), never probes a shared hosting sink, and counts a completed protocol handshake as live, because an MCP server answers a GET with 405. L6 is measured for the first time: 10.4% of 1,023 sampled registrations (6 MCP, 98 A2A, 2 x402); 97 of the 98 A2A passes are per-agent cards on one hosting platform. L5 was 13.5% on the last v1 sweep (12:26 UTC) and 14.0% on this one; the samples differ, so the gap is not a measured effect of the change.

/ai-agent-liveness-methodology-l0-l9-ladder

ERC-8021 parser v2: the suffix is found in a smart-wallet UserOperation's calldata as well as at the end of a plain transaction, and schema 2 is decoded. Every counted payment the v1 parser had read without a decoded code was read again (180,786 transactions, 11:22 to 14:40 UTC, free public nodes, none unreadable): 9,809 carry a code (9,025 in a UserOperation, 782 schema 2). At 14:55 UTC, 22,112 of the 193,498 counted payments read carry a code (11.4%), $5.49M of the $27.32M read; at most 12,303 did under v1. Earlier builder figures were read with v1 and undercount.

/def/pay.builder.attribution.v2

Correction to cp-2026-10-13-builder-parser, whose text stands: it named its series builder_share, builder_attributed_usd, builder_codes_distinct and builder_checked_share, and no series has those names. The published series are observed_pay_builder_*. So the step the parser change caused on 13 October (distinct codes 74 → 314, attributed $4.37M → $5.84M, top builder share 0.90 → 0.79) stayed listed as unexplained. Found on 23 October by checking every key in this register against the archived series; no other entry names a series that does not exist.

/def/pay.builder.attribution.v2

Six new layers from the second library batch, switched on 13 October 2026 between 20:00 and 21:00 UTC, each measured over the whole ledger before it went live: the test_then_drain rule (5 pairs since February against 53 probe-then-50x pairs in ordinary commerce) and the fleet sibling notice (95 notices for the four October drains); the behaviour clock (7,310 agents, median run 16 days, measured 6.2 breaks per agent-year against a design of 1); the independence score and operator sets (4,909 of 7,566 ranked agents paid a seed; 7,180 operator sets; the cluster-aware L9 keeps 71 of 72); ladder dynamics (at 90 days 63.2% no exit, 36.5% dormant, 0.3% drained; one minter's 695 agents at SDR 3.2, interval 1.3-6.6; landmark ratio 10.1 against the published 8.9); the Merkle-sum log (293,213 payments, $28,644,195.25, signed head, proofs per payment and per agent); issuer freezes (1,356 events, 979 addresses, none paid by a registered agent); x402 submitters (97.2% facilitator, 1.5% third party, 0 receiveWithAuthorization). New series start here; detections_logged gains two kinds. The ladder's rungs are unchanged.

/changelog-october-2026

L9 is operator-set aware (rank.l9.v4): a shared address counts as an independent counterparty only when an agent outside the payer's operator set also pays it, and a bot-like contract does not count. Measured on 13 October at 20:19 UTC before the switch: 71 of 72 L9 agents kept the level. Applied the same day at 21:22 UTC: 72 agents at L9 under both rules at that instant, so the series shows no step; rows before this moment carry rank.l9.v3.

/def/rank.l9.v4

From 15 October one platform's agents began moving small amounts between their two accounts (the account the registration was minted to and the account that holds it). Until the netting of 21 October 00:02 UTC (cp-2026-10-21-own-address-netting) these moves were counted as payments: 13,030 transfers worth $2,410.89 from 2,743 wallets, 2,272 of them contract accounts, almost all made between 15 and 20 October. Counted money barely moved, but counts did: paying contract accounts 553 → 1,971 between 15 and 20 October and 987 after the netting; known counterparties +445 on 16 October and −451 at the netting, the receiving accounts. That entry gave only the change in agents with a counted payment. Measured on 23 October from the own_transfers table the netting writes. L7 was not affected: ladder v2 already treated these addresses as the agent's own.

/def/pay.usd.total.v2

Correction to cp-2026-10-21-onboarding-transfers, which says the mint-then-hand-over pattern began on 18 October. Our incidents table, flagged by the same rule, shows it at scale from 16 October: onboarding moves logged per UTC day 19 (15 Oct), 352 (16 Oct), 20 (17 Oct), 442 (18 Oct), 935 (19 Oct), 397 (20 Oct), 620 (21 Oct). So live_incidents (49 → 389 on 16 October) and detections_logged stepped from 16 October for this reason, under incident.live.v2; the switch to v3 on 21 October is unchanged.

/def/incident.live.v3

Our payment reader served its queue in address order (a DISTINCT ON query sorted by wallet), so wallets the sweep had flagged for moving stablecoins waited up to three weeks for their first read: 4,447 never-read wallets on 18 October, the oldest flagged on 27 September. From the run of 18 October 09:42 UTC the queue is served by priority, oldest flag first. Measured on 20 October at 17:40 UTC, the catch-up had read 87,328 counted payments worth $4.08M from about 3,000 agents, almost all of them made before 18 October: L7+ went 7,630 → 9,593, counted payments $29.26M → $33.34M, x402-rail payments $5.0K → $66.4K, and the per-day series were revised upward for the days those payments were made. 3,108 flagged wallets were still unread at that time. These are payments that already happened, read late, not new activity; a closing entry will give the final size when the queue is drained.

/changelog-october-2026

Companion to cp-2026-10-18-payer-queue: the series computed from counted payments or from the ranked set (bridge follow-up, counterparty and category totals, x402 share, payer controller mix, the poisoning watch list, rated L7 agents, Bankr seller payments) stepped on 18 to 20 October for the same reason. The payments behind the step were made before 18 October and read late; none of it is new activity. Registered on 21 October after afg-steps listed these series as unexplained.

/changelog-october-2026

More series moved by the same late reading as cp-2026-10-18-payer-queue, which the companion entry did not list: the spending composition, more of the bridge follow-up, builder attribution, swap and work-escrow categories, agent-to-agent payments, 30-day net flow, the Merkle-sum log total, independence seeds and labelled counterparty events. Measured on 23 October from the payments themselves: between 18 and 22 October the payment reader counted $9.32M, of which $9.15M were payments more than three days old when read and $8.16M were made before 1 October. Between 20 October 22:00 and 21 October 02:00 UTC the position check read 835 of those payments ($1.94M, all made before 18 October) and found deposits into lending or vault positions the payers still hold, which moved them from lending (other) to own positions: comp_lending_other_usd $867.7K → $21.3K, comp_own_positions_usd $16.32M → $18.26M. The catch-up runs until 26 October; compare these series from 27 October.

/changelog-october-2026

An EIP-3009 authorisation submitted inside a batch (Multicall3 aggregate3 and the like) now counts as an authorisation (x402.settle.submitter.v2), read from the receipt's AuthorizationUsed with the payer as authorizer; v1 read only the outer function and filed these as other. Re-read on 20 October over the 20,000 stored settlements of 12 to 19 October, per UTC day: 128 of the 173 'other' were batched authorisations, all from one sender that settled for 9 sellers in the week; other 173 → 45, facilitator 19,655 → 19,719, third_party 61 → 125 (0.31% → 0.63% of settlements), because that sender served fewer than 5 sellers on most days. Found while checking a mainnet receipt on erc-8004/erc-8004-contracts#99. Applies from the run of 21 October 04:40 UTC; earlier points carry v1.

/def/x402.settle.submitter.v2

Transfers from an agent's wallet to another of its own addresses (any wallet ever bound to it, its registration's owner, the address it was minted to) leave the counted set (pay.usd.total.v2); until now only a transfer to the very same address was netted, while the L7 rule already treated these addresses as the agent's own. Measured at the switch, 21 October 00:02 UTC, two ways: counted payments fell by 4,061 rows and $688.05 ($36,675,055.59 → $36,674,367.53), equal to the new own_transfers total; agents with any counted payment 11,692 → 10,474. L7 and its rungs are unchanged by this entry.

/def/pay.usd.total.v2

Second wallet for two-wallet agents (binding.minter.v1): on Central City each agent pays from the account its registration was minted to while the registration sits in a second account, and ERC-8004 clears agentWallet when it moves, so the paying side was outside the ledger. At 00:15 UTC on 21 October 7,201 such minting accounts were bound. Measured from the routers' own events before the switch: by 20 October 21:00 UTC these accounts made 165,041 fee-router payments worth $113.0K to other agents and the platform treasury. The payment reader reads the new wallets at its normal pace, so L7, counted payments and the per-day series rise over the following days; transfers between an agent's two accounts are netted (cp-2026-10-21-own-address-netting). A closing entry will give the measured size once the new wallets are read.

/def/binding.minter.v1

Correction to cp-2026-10-21-minter-binding, whose text stands: the composition series (comp_*, the parts of the counted set by what the payee is) were not named, although they move with the counted set. As the second wallets of two-wallet agents entered the books from 21 Oct, the work-escrow part rose $5.72M → $6.08M in the day (its largest payee, TermixEscrow, is where most of those agents pay) and the other parts moved with their payees. Registered late, on 24 Oct 2026.

/def/pay.composition.v1

Facilitators are now also recognised by the public facilitator list (x402.settle.submitter.v3; a dated snapshot of DefiLlama's list, which credits x402scan and x402-list.com), and named. The batching sender behind cp-2026-10-20-x402-batched-auth is PayAI's facilitator; under v2 it counted as third_party on days it served fewer than 5 sellers. Re-read over the 20,000 stored settlements of 12 to 19 October, per UTC day: third_party 125 → 61, facilitator 19,719 → 19,783 (Coinbase 19,655, PayAI 128). Applies from the run of 21 October 04:40 UTC.

/def/x402.settle.submitter.v3

An owner change within 24 hours of a registration's mint is no longer a live incident (incident.live.v3). Since 18 October one platform mints each new registration to one account and hands it to the holder minutes later; v2 opened a 'binding closed after transfer' incident for every such move. They are now logged as info detections with onboarding_move = true. Of the 2,799 transfer detections ever logged, 2,786 were onboarding moves. live_incidents went from 1,924 to 0 at 08:20 UTC; detections_logged (3,782) and detections_7d are unchanged. Earlier live_incidents values are not comparable with later ones.

/def/incident.live.v3

The daily registry census was stopped by its 30-minute time limit at 03:29 UTC on 23 October: the registry had grown from about 509,000 to 899,000 registrations, and a run now takes over an hour. 124,352 registrations had no record yet, so they could not be bound or ranked. The limit is now four hours; a re-run from 08:56 UTC created 76,291 records and the rest are read by the following runs. Bound-wallet and census series step up for the catch-up, not for new registrations.

/def/census.bound.wallets.v2

A second cause of a starved payment queue (the first: cp-2026-10-18-payer-queue). The three-hourly sweep re-stamped a wallet's 'moved money' flag every time it saw a transfer, and the queue serves the oldest flag first, so the most active wallets kept moving to the back. On 23 October 10,874 flagged wallets had never been read, 7,841 of them the minting accounts bound on 21 October (cp-2026-10-21-minter-binding), of which 6 had been read; that is why the rise announced there had not appeared. The flag now keeps the time it was first set, accounts bound because they paid through a fee router are read first inside each tier, and a catch-up runs until 26 October. L7, counted payments and the per-day series rise as these wallets are read; the payments behind the rise were made earlier and read late, none of it is new activity. The closing entry of cp-2026-10-21-minter-binding will give the measured size.

/def/binding.minter.v1

Two detectors read at most 100 rows a run and logged a warning when they hit it: the retry-storm list (101 storms existed) and the endpoint time-out streaks (640). The limit is now 5,000, and a run that reaches it says so in its report. Measured on the runs before and after (12:54 and 13:54 UTC): retry storms 100 → 101, time-out streaks 100 → 640. Streaks are info only and never live incidents (incident.live.v3); live_incidents stayed at 3. The larger rise in detections_logged on 22 and 23 October is the platform onboarding wave (see the checked note for those days), not this change.

/detections-ai-agent-anomaly-log-mint-bursts-retry-storms

Three data checks on our own collectors were added (sweep_lag, census_gap, payer_queue_age, all warn): 45 → 48 checks. Each failure they catch ran unnoticed for days before 23 October. On the first run (13:35 UTC) payer_queue_age failed: 606 bound wallets flagged for moving stablecoins had waited more than three days for their first payment read, the oldest flagged on 18 October. That is the queue starvation of cp-2026-10-23-payer-queue-first-flag being worked off, not worse data.

/data-health

x402.settle.submitter.v4. Under v3 the 2,500 transactions read each day were the first 2,500 found, which were the earliest blocks of the first batch of 400 sellers, and x402_settle_txs_24h was the number read: it sat at exactly 2,500 from 15 October. v4 counts every transfer found in the window and reads an even sample across the whole window and every seller. Last v3 run, 23 October 04:46 UTC: 14,628 transfers found, 2,500 read, facilitator 99.36%, third_party 0.08% (2). v4 at 14:14 UTC: 15,927 found, every 6.37th read (2,500), facilitator 96.48%, third_party 1.88% (47, of which 39 from two senders that each settle for one seller only, the shape of a seller's own relayer). The two runs cover windows 9.5 hours apart. A first v4 run at 14:09 UTC still mixed in rows read under v3 (third_party 1.62%). Found on 23 October because the count never moved.

/def/x402.settle.submitter.v4

The payment reader reads a wallet's transfers newest first, at most 10 pages of 1,000. Until now a read that stopped at that limit, or at the run's compute limit, was recorded as complete, and every re-read started again from the newest transfer. Ten wallets had reached the limit; for eight of them the stored history already reaches back past their registration, so no counted payment was missing. Two were missing payments made after registration (blocks 45,243,652 to 47,164,292 and 46,300,403 to 50,559,621); those ranges are now read backwards with what each run has left, and counted payments will rise as they are read. A read cut short by the compute limit stays queued, and a wallet already read asks only for blocks after its last complete read (less an hour), about a tenth of the cost for busy accounts.

/def/pay.usd.total.v2

One data check added, payer_vs_logs (warn): 48 → 49 checks. It compares the payments stored by the payment reader with the stablecoin Transfer logs read straight from the chain for a one-in-five sample of bound wallets, over a window that closed 24 hours earlier, and warns when more than 10% are missing. On its first run it failed: 52,933 of 52,968 transfers from 21 to 22 October were not yet stored (the starved queue being worked off). The comparison had shown this gap since 22 October 03:41 UTC; until today nothing alerted on it.

/data-health

The same catch-up adds older payments as it reads them, so the payment series step up without new spending behind the step: counted payments 427,669 on 22 Oct, 1,928,818 on 28 Oct, 2,713,301 on 30 Oct; counted USD $38.56M, $55.19M, $56.59M; left control $20.07M, $29.97M, $30.53M; own positions $18.49M, $25.22M, $26.06M (each the day's latest reading). Most of the added events are small router payments of one platform, which also moves the median payment and the top-actor share. The reader is still adding older payments (+409,665 events and +$1.00M on 29 Oct, +374,818 and +$0.40M on 30 Oct); the dated allowance ends at 23:59 UTC on 31 October and the reader returns to its base pace, so comparable_from is 1 November; if that day's additions are still above the organic rate (before 23 October: tens of thousands of payments and $0.1M to $0.5M a day), a later entry moves the date forward. The day-of-payment series (pay.usd.by_day.v1) are revised backwards as late-read payments are placed on the day they were made.

/def/pay.usd.total.v1

Coverage. The census of ERC-8004 registrations had been stopped by its 30-minute limit since about 18 October (cp-2026-10-23-census-timeout). Its first full run on the 4-hour limit recorded the registrations it had missed, so the census series stepped on 24 October without that many new mints on the day: registrations in the all-chain census 1,048,466 to 1,180,334; actors 990,260 to 1,122,075; registrations with an owner read 990,312 to 1,122,127; registration transfers 989,623 to 1,037,205 events. From 25 October these series move at their own pace.

/def/census.registrations.all.v1

Backfill. These series are built from the payments the October catch-up added (cp-2026-10-30-catchup-payments-backfill) and step with them, without new behaviour on the day. Largest measured steps: lending deposits counted $18.67M to $24.30M (27 October); bridged payments $8.11M to $10.11M (27 October); bridged money followed back to the payer $6.45M to $8.33M; counterparties paid by three or more agents 12,201 to 14,866 (28 October); paying controllers that are contracts 3,408 to 7,586 (25 October); agents with income from others 3,127 to 4,507 (26 October); ETH-outflow wallets checked 8,575 to 16,657 (26 October, the reader's first full pass at 700 reads a run). The x402 share of payments fell 0.071 to 0.030 on 26 October because the added payments are mostly router payments of one platform. Compare these series from 1 November, the same date as the payment totals.

/def/pay.usd.total.v3

Backfill. These series are computed over the ranked agents and their payment histories, so they followed the rank catch-up (cp-2026-10-30-catchup-ranks-settled): independence score 1 for 6,878 agents before 25 October and 20,589 on 28 October; agents measured for stability 11,590 to 13,605 (28 October); ranked agents with under $1 counted 7,191 to 6,552 (29 October, as late-read payments lifted agents over $1); dormant share at 90 days 0.392 to 0.337 and no-exit share 0.605 to 0.660 (28 October). The steps of stability_runs_90d_plus and independence_score_0 on 20-23 October came before the catch-up and are not covered by this entry.

/def/rank.independence.v1

The Agent Pass left shadow: from 08:53 UTC every Pass row is signed with Ed25519 key afg-pass-2026-10 (public key at /.well-known/afg-pass-keys.json). The eight checks, the validity (3 hours) and the wording did not change; rows issued unsigned between 16 and 24 October are replaced as they expire. At the switch 7,695 of 12,335 ranked agents held a Pass (L7 7,218 of 11,620; L8 405 of 630; L9 72 of 85). The Gate and the Watch stay log-only.

/ai-agent-security-layer-gate-watch-pass-shadow#the-agent-pass

A detector added: delegate_changed (info, never live) when a bound wallet's EIP-7702 delegate differs from the previous code read; the first read of a wallet is not a change. detections_logged gains these rows from the next afg-delegates runs.

/def/detector.delegate_change.v1

The Gate's payee facts went from seven to nine (payee.facts.v2): the payee's usual price over 30 days (median and 95th percentile of counted payments and x402 settlements into it, stated from 20 payments) and, when a paying agent is given, that agent's over-sized x402 settlements of the last 30 days under the x402.settle.oversize.v1 rule. The seven inherited facts are unchanged; no metric series moves.

/def/payee.facts.v2

Coverage. From 25 October the treasury reader read 3,000 agent treasuries a day instead of 900. The wider sample moved the totals and the asset mix: treasury USD $24,792 to $66,853 on 25 October and $81,084 to $36,486 on 27 October; ETH share 0.81 to 0.36 on 25 October; token share 0.03 to 0.35. The DeFi share's fall from 0.1117 to 0.0003 on 28 October was the sample changing, not withdrawals: the wallets holding most of the DeFi value were last read on 25-26 October and left the reader's 26-hour window (see cp-2026-11-01-treasury-rotation).

/def/treasury.usd.total.v1

Correction: ERC-8004 agent ids start at 0, so the registrations on a chain are the highest existing id plus one. v1 (24 Oct) reported the highest id itself and called it an upper bound, one short on every chain. From this run each chain gains 1: erc8004_regs_upperbound_all_chains +1 per chain with registrations (+21 on 24 Oct), _uncounted +1 per uncounted chain (+15), _0g +1, _base +1. The 24 Oct finding total 5,577,152 is 5,577,173. Checked two ways: ownerOf(0) answers on seven chains, and our log-based census counts exactly the highest id + 1 on OP Mainnet (568) and Polygon (691).

/def/reg.chains.upperbound.v2

Backfill. The sum log holds one leaf per counted payment, ordered per agent and by time, so payments read late are inserted into earlier positions. During the catch-up most agents' histories therefore changed between heads: leaves 594,377 to 1,066,873 (26 October) and 1,773,222 (28 October); agents 14,590 to 26,385; total $40.53M to $54.65M (28 October); the share of the previous head that is a prefix of the new one fell from 0.514 to 0.004 (26 October); agents whose history changed 31 to 2,953 (28 October) and 4,895 (29 October). This is the log reporting the backfill as designed, with every changed history counted and published, not a rewrite of published payments. Removed leaves are the correction already registered (cp-2026-10-31-loop-returns: 274).

/def/sumlog.v1

The first-read backlog of the payment reader is cleared. The catch-up that began on 23 October (cp-2026-10-23-payer-queue-first-flag, cp-2026-10-23-census-timeout; 600 wallets a run from 23 Oct and 1,000 a run from 25 Oct, within a dated Alchemy allowance) has read every bound wallet the sweep had flagged: on 30 Oct no flagged wallet is unread (the payer_queue_age check is clear), all 13,901 minting-account bindings (binding.minter.v1) have been read, and 1,067,024 bound wallets have been read at least once. Measured step: agents at L7 or above 10,881 on 22 Oct, 27,602 at 22:56 UTC on 28 Oct (+16,721), then 27,844 on 29 Oct and 27,862 on 30 Oct; L8 or above 720 to 753; L9 96 to 123; payment counterparties 13,359 to 28,296. From 29 October the rank series move at their own pace. The step before that is our reader catching up, not the market: most of the new ranked agents are the router accounts of one platform (cp-2026-10-24-minter-binding-composition).

/def/rank.l7plus.v2

Coverage: the wallet sweep had classified 7,193 wallets as EIP-7702 delegated from their code, but the implementation address behind each was stored only by the delegation reader, 900 wallets a run, so 4,230 delegated wallets carried none and every family count was an undercount. On 30 Oct each of them was re-read through a public Base node at blocks 53,289,975 and 53,290,093 and its implementation stored (4,152 stored; 32 found with plain code again, their delegation revoked since the sweep read them; 6 stale implementations cleared; 60 unreadable at the time). wallets_delegated_7702 moved from 2,975 to 6,668 within the hour and the ranked-wallet finding from 930 to 1,918 delegated wallets; the top-implementation share (0.2955) and top-four share (0.80) are over the fuller set. Compare these series from 30 Oct 2026.

/api/findings

Method, correcting an over-count present since 22 October. Comparing a platform's public dashboard with our ledger showed that one agent received about 1.83M USDC through a fee router from an address bound to no agent and sent almost all of it back to that address and to two addresses that funded it, 22 to 30 October. Our loop check (pay.loops.v1) only compares payments between bound wallets, so those transfers counted as the agent's spending. From 31 October 11:18 UTC such loop returns are excluded from counted payments (pay.usd.total.v3, pay.excluded.loopreturn.v1) and their amount is published on its own (loop_return_pay_usd_excluded). Measured jump: counted payments 2,816,156 to 2,815,882 (-274) and $56,661,208.25 to $54,842,468.46 (-$1,818,739.78, -3.2%). None of the excluded payments was an own position, so left control falls by the same amount at its next run, as do the composition's unnamed part, the per-day series for 22 to 30 October and the averages built on them. 5 payments of the same agent (9,909.29 USD) stayed counted because they exceeded what it had received from the payer at that moment. The rule is narrow on purpose; what it does not catch is written in the definition.

/def/pay.excluded.loopreturn.v1