Integrity digest, October 2026

Projects publish numbers. We check them against the chain, an independent index or the project’s own deployed contract, and record the verdict with the date and the method. This is the second monthly digest of that ledger, covering 20 September to 4 October 2026: what was checked, what held, and what did not.

That tally needs one piece of context: on 23 September we rebuilt the whole board, and nearly every row was re-cut and re-dated inside this window. The count is close to the full board, and most rows carry earlier work re-tested under stricter rules.

The rebuild started with an audit of our own ledger. Of 422 rows, only about 99 were a claim the project had made, present on the page we linked, with a verdict that followed from the evidence. The rest mixed in our own earlier review text shown as “They claim”, figures from third-party indexers, and quotes the linked page did not carry. Every row is now one of two kinds:

• A Claim: the project’s own sentence, quoted word for word from a page it controls and confirmed on that page by a script that keeps a dated snapshot.
• A Finding: our own measurement, in our voice, graded established, unverified or unverifiable.

Only a Claim can carry a correction verdict, so every correction in the tally above grades something a project published.

The corrections that changed a grade

NuNet’s fixed supply was minted through

NuNet’s docs say its Ethereum and Cardano tokens convert 1:1, “keeping the total token supply constant at all times at 1 billion NTX”. On 19 September a compromised deployer key minted 408.5 million NTX on Ethereum in a single transaction. The three chains now carry more than 1.4 billion NTX between them, and the docs page still states the constant.

The mint function has no cap guard, so the limit only ever lived in NuNet’s documentation. Supply Dynamics was cut hard on that evidence. On 25 September we found the same key still owns the upgrade path of the BSC token, which lowered its Censorship Resistance score.

Aethir’s ARR is a peak that has passed

Aethir’s Q3 2025 post says its Annual Recurring Revenue “has skyrocketed from $12 million to over $166 million”. Our own index of service-fee deposits on Arbitrum, converted at daily ATH prices, shows monthly fees falling from $12.5M in September 2025 to $3.0M in July 2026. Aethir’s own dashboard showed ARR of about $45.5M on 23 September, a long way short of $166 million.

Deposits counted in ATH rose over the same months, which looks like growth until the token price sits beside it. Across 11 months, ATH deposited moved against the ATH price with a correlation of -0.90, the pattern of compute priced in dollars and settled in tokens. Token Terminal’s separate monthly series agrees with our index to within 3% across the last five months, so the dollar decline isn’t a quirk of our conversion. Both rows quoting the ARR are graded outdated.

Giza’s fee loop has gone quiet

Giza’s July 2025 tokenomics report says “every uptick in the AUA feeds fee flows back to the treasury”, AUA being assets under agent. DeFiLlama’s open on-chain adapter reads zero protocol revenue across the trailing 30 days, 7 days and 24 hours, against $56,685 for the trailing year. The loop ran and then went silent, and the yearly figure still carries the earlier months, which makes it read healthier than the position is.

Flux runs 39% fewer nodes than its docs say

Flux’s docs say that as of mid-2025 it has “over 10,000 nodes” run by more than 560 independent providers. The deterministic node list held 6,108 nodes on 18 September. The providers half fits the data, because the nodes resolve to about 880 payout addresses, an upper bound on operators. Flux dated its own figure, so the verdict is outdated, and the same page links a live dashboard that could keep the figure current.

io.net claims twelve times what its explorer lists

The FAQ data on io.net’s homepage says its network “includes over 30,000 GPUs and 80,000 CPUs”. io.net’s own inventory API listed 2,447 devices, 1,199 of them active, when we read it on 13 August. The explorer is self-reported, so we use it as a self-contradiction and anchor the verdict on Messari’s Q1 2025 count of 6,720 daily active GPUs, the latest independent measurement. Multi-GPU hosts can’t close the gap, since only 128 of those devices are NVLink, SXM or DGX class.

ZetaChain sells a lock that stopped locking

A ZetaChain post on 17 September says “Lock ZETA, receive credits, and spend them on AI usage. Locked ZETA comes out of circulating supply.” The lock contract paused new stake on 24 June and claims on 3 July, almost three months before the post. Anuma’s live Token Dashboard reads ordinary delegation instead, which bonds ZETA without removing it from supply, and grants credits in Anuma’s off-chain ledger at a rate the company sets. About 83,691 ZETA still sat in the paused contract on 29 September.

Vana’s timelocks aren’t deployed

Vana’s security docs say its core contracts “are upgradeable through governance with timelocks”. All eight are UUPS proxiesUUPS ProxyA pattern for upgradeable smart contracts where the upgrade logic sits in the implementation contract itself. Cheaper gas than the older Transparent proxy, but the implementation must be written carefully or the contract can be bricked.A power tool where the swappable head also contains the swap button. Pull the head off, snap on a new one with a new button, you can keep using the tool. But if the new head forgets to ship a swap button, the tool is now stuck on whatever head you just attached.UUPS Proxy in the glossary → whose upgrade authority sits with one 3-of-7 Safe, with no timelock and no on-chain governance anywhere in the path. The permission contract’s logic has been swapped several times since deployment, so three signatures can change what the chain enforces on users’ data, and have done so before.

Where the grades moved

Five grade changes landed in the window, each on settled evidence and each recorded in the project’s score change log:

• Render Returns fell from B to C: emissions outrun job-payment burns and the 5% fee is off-chain.
• Virtuals Freedom fell from D to F: a single EOA can upgrade the core agent registry with no timelock, which is the same finding that grades its governance claim inflated.
• FLock Returns fell from C to D on revenue that has all but stopped and a value-accrual case resting on its own reported fees.
• Ocean Returns rose from F to D: the supply is capped on chain, the owner is renounced and vesting is complete.
• Vana fell from C to D on both Freedom and Returns: ten addresses hold over four-fifths of native VANA, and neither Coinbase nor Kraken lists it.

This month’s review moved 12 scores across ten projects, and Vana’s two were the only ones that changed a grade. Heurist’s Token Distribution and Liquidity scores both fell, and it stays at F.

Corrections we owed the reader

Some of the month’s corrections overturned our own words rather than a project’s. They sit in each project’s score change log, never on the board, because a correction to our text is ours to carry. The ones that changed what a reader would have taken away:

• Our first digest. The September digest led on a tally of corrections, most of which were fixes to our own earlier text counted as if projects had erred. We removed those tallies on 23 September, when the rebuild made the split visible, and corrected its ORA, Aethir and Bittensor lines to what those projects had published.
• Olas. Our August governance trace checked two retired GovernorOLAS contracts and missed the current one, which has held proposer and executor roles on the Timelock since 15 June. The verdict on Olas’s Timelock claim stands, because the delay still reads zero, but the evidence under it was incomplete.
• Giza. Our token record said the owner Safe could mint more GIZA. The mint is capped in code at 1 billion and fully issued, so it can’t, and this correction runs in Giza’s favour.
• Allora. Our score rationales credited permissionless workers and open topic creation. Mainnet enforces worker, reputer and topic-creator whitelists on all 23 topics, and the Freedom score fell.
• Flux. Our review repeated the line that Flux runs on no centralised data centres. The ten largest IP blocks host about a third of nodes, so we withdrew the line and Infrastructure fell a point.
• Aethir. Our token-contract record was corrected twice in two days. The second pass found the supply capped in code and fully minted, plus an owner-only whitelist over the contract’s remaining balance that the first pass had missed.

Venice, Nosana, Nillion and OpenMind carried smaller data, listing and attribution fixes in the same fortnight. Phala’s withdrawn revenue verdict, on 17 September, falls in the previous window.

The September digest said three of our own corrections were flagged for this review. Heurist’s scores moved in this review. IoTeX’s Governance score had already moved on 7 September, so that digest’s line on it was out of date when it went out. Nillion’s Censorship flag, on a token a 3-of-5 Safe can replace, is still open and carries into the next window.

What the month says about the method

Half of this month’s corrections share one shape: a page that was accurate once kept saying the same thing after the chain moved on. Those rows are graded outdated, a verdict about upkeep, which is a different failing from a figure that was never true.

• Aethir’s Q3 2025 momentum was measurable at the time, and monthly fees have since fallen by more than 70%.
• Giza’s fee loop ran, and it read zero across the trailing 30 days.
• ZetaChain’s lock contract accepted stake for about five weeks, and the post promoting it came almost three months after it paused.

Most of the rest are governance and control claims graded against deployed contracts. Vana, Virtuals, Olas and Bittensor each describe a limit on their own power that the deployed code doesn’t impose. That pattern held from September, and the rebuild sharpened it, because a verdict now has to quote the sentence it grades.

Our own errors carry a lesson too. The Olas and Aethir self-corrections both came from checks that read part of a set and stopped: the retired governor contracts without the live one, and some of a token contract’s functions without the rest. The fixes that held came from enumerating the whole set, every role grant and every proxy.

The next window

The next digest covers 5 October to the November review. Four things are already queued:

• Nillion moves Blacklight to Ethereum mainnet, set for 5 October by Nillion’s own schedule, and its Censorship flag is due in the same review.
• Olas is watched daily: the Safe losing its Timelock roles, or the delay rising off zero, would overturn its verdict.
• NuNet’s supply is re-read across all three chains for further mints or burns of the excess.
• OpenServ’s docs describe a token and treasury multisig with five board-level signers; the owning Safe reads a threshold of two, and that sentence is a candidate Claim.

The board updates as rows are re-tested. The digest is the dated reading.