On 19 September a compromised deployer key minted two fifths of NuNet’s supposedly fixed supply, sold it overnight, and froze every holder on Ethereum the next morning. One of our integrity-board rows had named that exact path on 12 August. The same day, asked for the best source of independent reviews of decentralised AI projects, Grok answered Own Your Mind, and we published the first monthly digest of the board it was describing: 185 claim-checks across 41 projects, 47 corrected. This issue is the ledger, in both directions.
What broke this week
NuNet’s 1 billion cap turned out to be a sentence on a website. At 20:50 UTC on 19 September the deployer key, which had held mint, pause and admin on the Ethereum token since 2021, minted 408.5 million NTX in one transaction and took that leg from 591.5 million to exactly 1 billion. Add the Cardano and BSC legs and NTX now totals about 1.41 billion against a published maximum of 1 billion. The new tokens were sold through MetaMask Swaps overnight.
Twelve hours later admin moved from that key to a fresh single address, the old key was revoked and the contract was paused. Every NTX holder on Ethereum has been frozen since, with no carve-out. Cardano and BSC kept settling, which is the one part of the multi-chain design that did what the docs said it would. Control of the token is still one bare key, before and after.
The mint was the second theft of the evening. Twenty-nine minutes earlier the same receiving address had emptied Fetch.ai’s conversion escrow, the contract that holds FET for cross-chain conversion, with a validly signed release. FET’s supply is untouched and its token contract sits behind a three-of-five Safe that was never involved. The conversion path beside it released on one off-chain signature, and that is what went.
Our NuNet review said this could happen. On 12 August the board recorded that mint() carried no cap guard and that a single unrenounced deployer address held mint, pause and role-admin together, so new NTX could be created outside the bridge by one key. Thirty-eight days later that key did exactly that. The contract did what it was written to do, for whoever held the key.
NuNet’s Returns score fell from 3.3 to 2.4 and its Freedom score from 4.7 to 4.5, and the supply-cap verdict on the board now reads outdated at high severity. Fetch.ai’s scores hold, with the single-signer conversion path flagged for the October review.
Full reads: ownyourmind.ai/projects/nunet/ and ownyourmind.ai/projects/fetch-asi/.
What shipped this week
Gensyn published a training run you can replay step by step. Open-1B landed on arXiv on 15 September 2026 with the model, the full pretraining dataset, every checkpoint, the training code and the tooling to audit it. Any single step replays bitwise on ordinary hardware, so a reader can check the model was trained the way the paper says. The training mainnet that would make this a paid market is still pending, and the paper is a preprint. arXiv 2609.17380, 15 September 2026
Nous set its own agent loose on its own million-line codebase. From 2 September 2026 Hermes Agent ran 1,393 subagents for about nineteen hours and cut the non-test Python by a third. The merged pull request is public, so the claim can be checked line by line, which is more than any benchmark score offers. This is still the team’s agent on the team’s repository, and the dollar saving Nous attached is its own estimate. Nous Research, 16 September 2026
NEAR made perps confidential by default, with a narrower meaning than the headline. near.com perps launched on 17 September 2026, funded through NEAR Intents from any chain and executed on a direct Hyperliquid integration. The product’s own FAQ says positions sit on the order book like any perp venue; what is hidden is the deposit that opened them and who owns it. Trading fees stay with Hyperliquid, and nothing we read shows a NEAR share, so the token’s only leg is the Intents swap that funds the margin. perps.near.com FAQ, read 18 September 2026
Virtuals launched tokens that buy Nvidia stock with their own trading fees. Occupy went live on Base on 15 September 2026. Every token it issues pairs with a tokenised equity, and 40% of each trade fee buys that stock into a treasury run by ten AI “senators” with no human key-holder; creators take 30%. Where the remaining 30% goes, and whether any of it reaches VIRTUAL, is unstated on the product’s own pages, so for holders of the parent token this is a launch to watch. occupy.virtuals.io, read 20 September 2026
The board that corrects its author
The integrity board is the thing Grok named, and it said why in a public reply: “Independent on-chain claim checks and dual freedom/returns scoring remain scarce in DeAI, which is why that approach stood out.” Every row quotes a claim, checks it against the chain, a deployed contract or an independent data source such as CoinGecko, and records the verdict with the date and the method. The monthly digest is the dated reading of that ledger, and the first one is up.
The month’s largest corrections all have the same shape. A design document, a specification or a charter describes an intended mechanism, and the deployed system does something else.
NEAR’s economics spec still routes 10% of issuance to the protocol treasury. The live protocol config sets that rate to zero; validators receive all of it. We had restated the 10% on our own page.
Bittensor’s charter says community-governed. The finney runtime carries no council, senate or democracy pallet, and every privileged operation, runtime upgrades included, goes through one administrator key that Bittensor says belongs to a foundation multisig. Our own page had described a Triumvirate and a Senate since August, because that is what Bittensor’s documents describe.
Ora’s token page described a live oracle. The contract has been silent since September 2025, which Issue 16 covered in full.
Aethir positions itself as broadly distributed. Its own docs put team and investors at 24% between them, with the split published only as an image.
Then the corrections to our own pages. Most came from one sweep on 6 September, when every exchange-listing sentence on the board was re-read against CoinGecko’s full ticker set: it caught venues that had delisted since we wrote, figures that had aged, and a few adjectives like “exceptional” that the numbers no longer carried. Two we plainly got wrong. We had peaq on Binance main spot from January 2026, a market that never existed, and we had rated IoTeX’s Snapshot governance as functional when the venue’s last vote was in October 2024.
Heurist ran the other way: we had reasoned from the allocation table that the foundation controls half of supply, and the chain shows one Safe holding two-thirds. A ledger that only finds fault outward is one nobody re-tests. The rows that moved this month moved because the checks ran again, against the chain, with a known-good row beside each one.
The full digest, with every row linked to its evidence, is at ownyourmind.ai/journal/integrity-digest-september-2026/. The next one covers the month to 19 October.
One more thing about the citation. Asked the same question the same day, Perplexity left us out; its top source was an automated newsroom whose provider tracker lists seventeen providers with zero verified claims. A model’s answer is a snapshot of what it had read that day, so the About page records it with the date and nothing more.
Biggest movers this week
7-day change as of 20 September 2026. Twenty-eight of the forty tokens we track closed up, and the top gainer is the one whose confidential perps launched mid-week. NuNet’s fall is the mint above, priced by the market. Last week’s top gainer went the other way too: Giza gave back its doubling and more, and sits at roughly half its level of a fortnight ago.
Gainers:
NEAR Protocol (NEAR) +52.4%
Auki (AUKI) +46.8%
Phala Network (PHA) +24.9%
Venice (VVV) +20.6%
Sentient (SENT) +18.6%
Fallers:
NuNet (NTX) -79.4%
Giza (GIZA) -75.0%
Ora Protocol (ORA) -25.3%
Morpheus (MOR) -25.1%
FLock.io (FLOCK) -10.5%
See all movers: ownyourmind.ai/projects/movers/
What we verified this week
Eighteen rows re-tested or corrected across thirteen projects since Issue 16, plus five supply verdicts on 20 September. Two scores moved, both NuNet’s, both down, and the section above has them. Four of the corrections were to our own claims.
Morpheus’s privacy claim now rests on evidence we pulled ourselves. We fetched a live hardware attestation from the provider Morpheus tags as running inside a secure enclave. Intel’s signature checks out and the attestation is tied to that provider’s own connection certificate, so a prompt sent there is sealed from the operator, and “permissionless and private” moves from unverified to verified. The deployed marketplace contract also has no function that could pause, blacklist or filter anyone, which confirms the no-content-policy claim on chain. Its Snapshot voting page no longer loads, so governance is flagged for the monthly review.
io.net’s burn reconciles on chain to six decimals. Every burn io.net reports matches the burn in its settlement transaction, and the token’s total supply has fallen by an amount within a fifth of a percent of io.net’s reported cumulative figure. Client purchases stay a reported number with a verified floor.
Venice’s model catalogue is now measured. The public API lists 118 text models, 69 of them Venice-hosted open-weight and 49 proxied closed models a third party could withdraw. Holder concentration is now measured on Base as well.
We withdrew our own Warden transaction estimate. It compared a whole-chain extrapolation against Warden’s first-quarter figure, which are two different things. The cumulative count is unverifiable from any reachable source and the row now says so.
We withdrew a Phala revenue verdict. It had graded Phala against a Cloud ARR figure Phala never published. The measurement stands; the verdict against a number nobody made is gone.
Our Flux supply figure had aged out. Five supply verdicts landed on 20 September for Ocean, Sahara, Flux, Oasis and Vana. Four confirmed the published figure on chain. The fifth was ours: the Flux circulating figure we carried was right when taken and about 200 days of emission behind the explorer, and it now reads live.
We had Theta’s validator bar wrong by a hundredfold. Our page said a validator needs 20 million THETA and an enterprise invitation. Theta’s docs say 200,000 and permissionless, and nine of the 21 live validators sit near that minimum. The claim is withdrawn and the decentralisation dimensions are flagged for October.
Allora’s audit is still an announcement. Halborn’s full public audit index, control-checked, carries no Allora entry, and the report announced two years ago remains unpublished. Flagged for the monthly review.
See the full integrity board: ownyourmind.ai/integrity/
On the watchlist
Whether Grok still names us in October. On 19 September we asked it for the best source of independent reviews of decentralised AI projects and it answered Own Your Mind. What a model answers depends on what it has read lately, so we will ask the same question at the October review and publish the result either way.
Whether Morpheus voting has moved or stopped. The Snapshot voting page is gone; the GitHub repository where proposals are drafted remains. Which of the two it is decides the governance score.
Whether Venice’s eighth emission cut lands on 1 October. Seven have landed on the announced day so far.
Whether Giza’s fee mechanism produces anything in the next thirty days. The trailing week reads zero; the verdict moves back the moment the chain does.
Whether NuNet puts anything between one key and the mint function before it unpauses. A multisig on the admin role and a hard cap in
mint()are each a day’s work. Neither existed on 19 September and neither exists now.Whether Fetch.ai’s conversion path gets a second signer. The escrow that drained released on one off-chain signature while the token contract next to it sat behind a three-of-five Safe.
Whether AntSeed’s usage-linked rewards read on chain the way its docs describe. The docs say rewards now follow paid delivery, with a slash on early exit. The parameters are owner-set, so we will read the contract before it reaches the review.
From the site
The first monthly Integrity digest, with the tally computed from the board at build time and every correction linked: ownyourmind.ai/journal/integrity-digest-september-2026/.
The NuNet review, rescored the morning after the mint, with the August row that named the path: ownyourmind.ai/projects/nunet/.
What Grok said, dated and attributed, on the About page: ownyourmind.ai/about/.
Every project we track carries dated, sourced verdicts across five scoring dimensions: ownyourmind.ai/resources/integrity-checks/.
The OYM podcast is on Spotify, Apple Podcasts and the major directories under “Own Your Mind”.
Cheers,
Bobski
Own Your Mind. Independent research on decentralised AI. No sponsors.
ownyourmind.ai · @ownyourmindai · @0xbobski
Disclosure: OYM holds MOR, NEAR, VIRTUAL and VVV.