Changelog, November 2026: corrections first

Changelog · November 2026 Corrections first, each dated with its measurement. Agent income had been read for 2,640 of 46,509 ranked wallets; it is now read for all of them, and the register says what the larger figure consists of. Base began making a block every 0.2 seconds and we corrected every figure that had assumed 2; the Watch had scored a fifth of counted payments while the security page said all; a slowed receipt reader held 303,057 small payments back for a day and a half. Built: a free self-check for any agent, the shape of the whole rank under the free 25 rows, MCP 1.9, the Agent Pass as a JWT and a Verifiable Credential, and Watch v3.

7 Nov 202611:35 UTC

Agent income, taken apart by sender, every run

The breakdown is a published figure now. income.composition.v1 splits counted income by who sent it, in the same run as the income figures, every three hours; the parts add up to the total exactly. At 11:31 UTC, of $57,824,959.57: from other agents' wallets 0.9% ($545,243.33); from a work escrow 19.0%; from vaults 18.9% (withdrawals from the agents' own positions, not sales); from swaps 3.8%, bridges 2.3% and lending contracts 1.9%; from x402 sellers' addresses 0.05%; from senders we have not named 53.2%.
Nine vaults named. The full income read showed USDC vaults among the largest senders. Each is now labelled from its verified contract name (MetaMorpho, VaultV2, PlasmaVault), with its asset read on chain. 68 counted payments ($5,307.25) into them move from unnamed to vaults on the spending side (cp-2026-11-07-labels).

7 Nov 202610:36 UTC

Correction: agent income was read for 2,640 of 46,509 ranked wallets; now it is read for all of them

What we said and what was true. The income figures (income.usd.total.v1 and its family) said every ranked wallet was re-read about twice a week. The reader took 60 wallets a run and the 50 newest transfers of each: by the morning of 7 November it had read 2,640 of 46,509 ranked agents' wallets, so every income figure we published was a floor for a small part of the ranked set. Spending, levels and registrations are read from every bound wallet and are not affected.
The fix. Every ranked wallet's incoming USDC, USDT, EURC and DAI was read from the chain by recipient, from the first Base registration to block 54,209,693; from there the sweep, which reads every stablecoin transfer on Base, records income as it goes. At the first full run (10:28 UTC): income $57,759,142.00 (was $3,761,583.76 at 07:57 UTC), 3,310,142 transfers (was 55,634), 30,101 agents with income (was 2,519), $13,885,198.37 over 30 days (was $912,254.40). Checked two other ways: one work escrow's 6,048 payouts to bound wallets in the week of 30 October, read on chain, are all in the ledger now (it held 142), and 25 randomly chosen ranked wallets read through a second index agree with the ledger to the cent.
What the figure is. Money received, not money earned. Payments from other agents' wallets were 0.9% of it. One work escrow sent 18.9%, and in the week we checked 5,935 of its 6,048 payouts went to wallets that also paid into it that week; one contract sent 23.4% to 145 agents; one address sent 9.8% to a single agent in 342,359 transfers between 22 October and 6 November; three USDC vaults sent 12.5% as withdrawals. Net flow over 30 days moved from -$8.09M to +$4.87M for the same reason.
Two rules, corrected. The hop test judged an income transfer on the sender's movements instead of the agent's; it now judges the agent's wallet, as it does for payments. A transfer from another wallet of the same agent is now a self-transfer, as on the spending side. Register entries cp-2026-11-07-income-coverage and cp-2026-11-07-income-hop-rule in /api/changepoints; compare these series from 8 November.

7 Nov 202608:23 UTC

Watch v3: 65% fewer notable alerts on routine traffic, the same catch rate

One reason, one condition. The Watch (behaviour.out_of_character.v3, from 08:17 UTC) no longer counts a payee's EIP-7702 account type as a reason when at least 100 registered agents paid that payee before. Pre-registered at 08:10 UTC, before any v3 run, with the rule for shipping fixed in advance. On the evaluation data it scores exactly as v2: 13 of 15 scoreable drains caught at notable, 301 of 319 synthetic drains, 37 control alerts over 47.7 agent-years. Replayed on the 2.69 million payments scored in the previous 7 days, 853 of 1,308 notable scores fall to none and all 244 major scores stay. Register entry cp-2026-11-07-watch-v3.

7 Nov 202608:07 UTC

The Agent Pass, portable: a JWT and a W3C Verifiable Credential

Same Pass, two more forms. A Pass that holds is now also served as a JWT (/api/pass?agent=<id>&format=jwt, EdDSA, key in /.well-known/jwks.json) and as a W3C Verifiable Credential in the JWT encoding (&format=vc, issuer did:web:agenticfinancegraph.com, /.well-known/did.json). The eight checks, the key and the three-hour expiry are those of the Pass (agent.pass.portable.v1). Checked from the live site with two standard libraries: jose verifies the JWT, did-jwt-vc verifies the credential, and both reject a copy with one changed claim. A registration without a Pass that holds gets a 404. MCP 1.9.1: agent_pass names the new forms.

7 Nov 202607:55 UTC

EIP-7702 share, v2. The share of agent wallets running delegated code was taken over every bound wallet whose code we had read. Since 30 October our wallet investigator reads about 10,000 more of those a day, most of them registrations that never paid, so the share fell from 15.5% to 8.1% and rose to 9.9% with the order of reading, not with the market. From 07:53 UTC today it is taken over ranked agents' wallets, all 46,446 of them read: 2,050 delegated, 4.41% (wallet.7702.rate.v2; register entry cp-2026-11-07-7702-rate-ranked, which also corrects the 30 October entry's "compare from 30 October"). The live data page said one in six; it now carries a dated note.
Steps. Since last night the step check found five more steps in our own series, and one was still open. All six now have a dated explanation with its measurement: two approvals series moved when wallets that had approved one escrow contract on 2 November were counted as payers; one relayer's bursts for a single x402 seller; the ETH-outflow reader's first full pass; the wallet-code backlog above; and a plant mix in the poisoning watch, where fake-symbol plants now often arrive exactly one day after the payment (checked on a public Base node).

7 Nov 202607:26 UTC

An independent ERC-8004 study beside our figures, with dates

arXiv:2606.26028, dated. The registry page already cited the study's live-endpoint figures (3%, 4% and 15% of registrations on Ethereum, BNB Smart Chain and Base). It now says they come from data up to 13 May 2026, that our September readings (2.7% to 8.2% a day) used a different rule, and that since mid-October nine in ten or more of the sampled registrations that answer do so from one platform's hosting, so the 91.4% read at 06:15 UTC today describes what the registry is made of, not agents at work. The method page carries the same point as a reference note.
Reviews. The trust-features page sets the study's finding that feedback is rarely grounded in verifiable interactions beside our fixture of 1 November (840 of 631,230 Base feedback entries, 0.133%, follow a payment from the rater), and says why the share written by the ten largest reviewers is not a Sybil test.

7 Nov 202600:04 and 00:09 UTC

MCP 1.9.0, two late register entries, and 44 steps in our own series explained

MCP 1.9.0. pay_check is the front door ("start here before your agent pays or hires another agent"), and every tool that overlaps it says when to use it and when not. agent_record, agent_statement and evidence_pack_quote accept agent as well as id; the canary tools say first that they need a Desk session; ladder_dynamics takes compact; an answer for an agent we do not rank says what to try next (try_next).
A late register entry: the receipt reader slowed. A payment counts only after its receipt is read. From 30 October 17:33 to 1 November 10:09 UTC a settings file on our server held the reader to 18 minutes an hour instead of 50. After the fix it read 303,057 counted payments ($5,794.67) from 3,723 agents that had waited, 225,871 of them ($3,588.36) from October; L7 rose from 27,897 to 29,906 on 1 November. Dated in the register as cp-2026-11-01-receipt-reader-slowed. A second late entry, cp-2026-10-31-bazaar-full-read, dates the Bazaar reader's switch from its first 25 pages to the whole catalogue (31 October 17:24 UTC), which widened the set of x402 sellers our seller figures count: 2,602 sellers read on 31 October, 2,877 on 2 November.
Steps explained. Our step check (/api/steps) listed 45 steps in our own series from 23 October to 4 November that nothing in the register explained. 44 now carry a dated explanation: the entry above, or a checked note with its measurement (most follow the October catch-up; others are bursts of owner changes, one bad probe sweep, an outside indexer's own figures, a burst of tiny payments by newly ranked agents on 3 and 4 November, and one agent's payments into a lending adapter). One stays open until it is measured: poison_copy_lag_median_min.

6 Nov 202623:19 and 23:34 UTC

The whole rank, free; one alert, start to finish; every detection in the packs

The shape of the whole rank now sits under the free 25 rows of the ranked page (rank.shape.v1). At 22:57 UTC: the top 25 agents hold 40.8% of the ranked agents' counted payments, the top 100 60.4%, the top 1,000 77.1%; 7,279 agents have paid $1,000 or more and hold 97.3%; the median ranked agent paid $9.15 over 31 payments. Measured from the roster and checked against a direct query of the payments: identical.
Names and labels. Declared names in the ranked table are quoted as the agent's own words, and a registration without a name says so. The edition pages' bar labels were cut at 34 characters with no mark; they now end with an ellipsis and show the full name on hover. No figure changed, and the PDFs stay as published.
One alert, start to finish. The home page walks through one public detection: registration 59642 paid $5.00 to an address it had never paid, then $1,994.58 to the same address 128 seconds later (21 September), read again from a public Base node on 6 November. Its history shows the same habit before and after, which is why an alert is a question for the owner, not a verdict.
Packs see every detection. The incident pack, the drain-shape pack, the Desk's alerts, the self-check and the agent page now read the detections logged on an agent from its own record. Before, they read the shared list of the newest 2,000 detections, about four days, so older ones were missed while the incident pack said "all logged detections". The drain-shape pack also reports the test-then-drain pair beside the drain rule, which cannot see that pair by design.

6 Nov 202618:00 to 19:01 UTC

Corrections: Base makes a block every 0.2 seconds, and approvals count each balance once

Base block time. Base began making a block every 0.2 seconds at block 53,556,127 (5 November, 18:00:01 UTC). Thirteen of our producers had assumed 2 seconds. The published x402 seller inflow over 24 hours read $41,846,800 at 16:18 UTC on 6 November, because USDC burns sent to the zero address were counted as seller income and placed in hours that had not happened yet; corrected to $48,328.32 at 18:29 UTC. One conversion now serves every producer. Every effect, measured, is in the register as cp-2026-11-06-base-block-time.
Loop returns after the switch. 1,561 payments of one agent ($966,060.27) that went back to their own payer were counted as spending until 19:01 UTC, because the receipts they are matched against were dated by the old rule. Counted payments moved from $57,305,192 to $56,368,474.
Approvals. The dollars behind standing unlimited approvals now count each wallet's balance of a token once (approval.unlimited.usd.v2): on 6 November at 00:48 UTC, $17,936.16 instead of $20,958.52 summed per approval (-14.4%). A reader's question caught it (cp-2026-11-06-approvals-usd-once).

3 Nov 202616:56 UTC

Self-check: everything the ledger holds about one agent

One free call at /self-check, through the MCP tool self_check, or by asking our A2A endpoint "self-check agent N": an agent's record and level, the Agent Pass, the Watch's reading of its newest payment, the token approvals its wallet gave in 30 days, its detections in 30 days and, at L8 and above, what a stolen key would control. It ends with the items worth a look, each with its source. Facts, never a verdict: a clean list does not mean safe.

2 Nov 202611:58 UTC

Correction: the Watch scored a fifth of counted payments, not all of them

What we said and what was true. The security page said every counted payment of ranked agents was scored. From 16 October to 1 November the scoring worker scored 538,825 of 2,402,995 counted payments (22.4%), from 4,402 of 24,779 agents. Unscored payments now go through a work queue, and very large histories get their own hourly pass (cp-2026-11-02-watch-coverage).

1 Nov 202611:38 and 12:02 UTC

October is closed and anchored on Base, and the security layer answers in advisory mode

The October edition is final (11:38 UTC); the hash of its file is a leaf of our signed statement log.
Anchoring is live. Once a day a zero-value transaction on Base names the newest signed heads of our statement log and our sum log, so anyone can check that our history was not rewritten without trusting our server. The first is in block 53,371,050 (anchor.v1, /api/anchors).
The Gate, the Watch and canaries left shadow at 12:02 UTC: answers are advisory, every call is recorded, nothing is blocked by us, and the Gate's default policy is log-only (cp-2026-11-01-security-layer-advisory). The Watch's notable line fires mostly on one payee that thousands of agents pay; that is published beside the figure.
Treasury coverage. The treasury reader had kept reading much the same wallets. From 1 November it takes never-read wallets first, 3,000 a day, so every L7+ wallet is read within about 15 days; compare its figures from about 16 November (cp-2026-11-01-treasury-rotation).

What we still refuse to say

That a detection is a verdict. That synthetic precision is live precision. That a registration is an agent. That a routing hop is spending. Every figure on this site carries the time it was measured and a definition you can argue with; a figure that turns out wrong is corrected here, dated, beside the measurement that found it.

Previous month: October 2026.

Questions people actually ask

Why publish a correction instead of quietly fixing it?

Because a figure that was wrong was quoted while it was wrong. The correction is dated, the old rule and the new one are both described, and the measurement that found the fault is public, so a reader can see what changed and when.

What is a checked note on /api/steps?

When one of our own series steps and no entry in the register explains it, the step stays open on /api/steps. We close it with a register entry when we changed something, or with a checked note citing the measurement when the inputs or the world moved. A step we cannot explain yet stays open; nothing is closed by guessing.

Corrections first. Then what we built.