Arkeo was meant to be a blockchain and AI marketplace for on-chain data. What exists is a live chain and a working marketplace, built and maintained, with eighteen of eighteen audit findings resolved. The AI integration is designed and waiting.
Innovation Theory signed on as Arkeo’s platform development contractor in June 2024 and was still diagnosing production breaks in March 2026. Everything on this page comes out of six documents: the contract, a third-party security audit marked for public release, a first-person status brief written to stakeholders, a 2026 plan, a services list, and an engineering task doc. Where those stop, this page stops.
Infura and Alchemy sell blockchain data, and most of the decentralized web quietly runs on them. If they go down or say no, the apps go with them. Arkeo’s answer was a market instead of a vendor: providers set their own terms, builders pick their own providers, and payment and reputation settle on chain. Below, the person who built it describes it, and so does the firm that was paid to attack it.
Arkeo is a decentralized marketplace for blockchain data. It lets anyone provide node data services and lets builders choose the most reliable and cost effective providers, with payments and reputation handled transparently on chain so apps are not forced to depend on a single centralized RPC provider.
Phil Cowan · Arkeo Engineering & Support Status, Dec 2025
Arkeo is a free-market, decentralized network that provides access to blockchain data. Its intended objective is to provide a decentralized option for communicating with any blockchain (including but not limited to) and its data. It is comparable to Infura and Alchemy, which are currently the industry’s leading providers of blockchain data.
Monethic · third-party security audit, Feb 2025
The AI layer is designed and still ahead of them.
The 2026 plan names Arkeo MCP Integration as a priority: decentralized data for AI, the plug that would make this network an AI data source. It hasn’t shipped, and this page won’t pretend it has. Where the AI is real in this engagement is the delivery; it’s how the engineering and the maintenance got done.
Innovation Theory, LLC signed on as Contractor under a document titled “Arkeo, Platform Developer Support Contract, v1.0,” addressed to Arkeo’s counterparty at ShapeShift. Colorado law, Jefferson County jurisdiction. Three months covering pre-launch, launch, and post-launch scaling, with Phil as Product Lead and primary point of contact.
The contract says why it’s written loose: “This is a flexible proposal that tries to accommodate the nature of uncertainty around the launch.” And one of the named deliverables wasn’t a feature: close out the issues found in the audit reports and bring the platform to a public testnet stage.
Mainnet went live on chain-id arkeo-main-v1, running cosmos-sdk v0.50.x and ibc-go v8. By July the team had built the Arkeo Master Services List, a catalog of blockchain data service types running to service ID 386, of which the first 51 were flagged as already added to the chain. The gap between those two numbers is why the Dynamic Service Registry exists.
Autumn was operations: a managed-server architecture overview, a relayer port map, a pay-as-you-go proxy load script, and the first real provider sentinel configs.
A marketplace dashboard with provider discovery, filters by service type and location and time range, and earnings and spending views. Provider and Subscriber Data Engines that ship as Docker images and get an operator live in minutes. A hot-wallet path that routes value from outside Cosmos onto the chain in about twenty minutes, fifteen of which are the swap. Four demo videos recorded December 22.
The status brief that went to stakeholders says where things stood, in its second paragraph: “Be aware this is all in beta and in progress.” And it names the gate in front of all of it: “The videos and descriptions of the apps in place are not able to be used on the live chain until we push the next engineering upgrade.”
The unglamorous part of the job, and the one worth the most. Every finding came back Resolved with a fix commit recorded in the final report, and every fix was independently retested by the firm that found the bug. This was the chain’s second audit, not its first.
Requested Jan 15, 2025 · work began Jan 24 · draft Feb 11 · retests Feb 21 to 28 · final Feb 28, 2025, marked “Final, for public release.”
The chain could have handed out free rewards.
A same-address transfer in the THORChain claim message double-counted a claim record, so a claim could pay out twice. Nobody catches that in a demo; it surfaces in the treasury months later, unless an audit gets there first. This one did.
Local time on validators is non-determinism.
Validators reading their own clocks can disagree about state, and a chain that disagrees with itself stops.
Reward distribution was computing the wrong number.
Incorrect logic in validator payout. Nobody notices a payout bug until somebody is underpaid enough to check, which is why an outside firm reading the code beats an inside team trusting it.
A testnet signature would have replayed on mainnet.
The chain ID was missing from the signed message, so a signature captured on testnet would have replayed against real value. It came down to one missing field, which is usually how it goes.
We didn’t grade our own homework.
The fixes went back to Monethic and were retested before the final report was written. Eighteen Resolved is the auditor’s word, not ours.
Marked “Final, for public release.”
Finding 16 asked for a specific service to be added and a duplicate removed; the July 2025 services list shows exactly that. The remediation is visible in an unrelated artifact five months later.
A full end-to-end data request through the Subscriber Data Engine, measured on a real run. Most of it is the upstream node answering. The decentralized layer isn’t what makes it slow.
Queue wait, height check, contract fetch and contract select all measured 0 ms on that run, on cache hits.
The same brief put the ceiling right next to the number, in capital letters:
The MAJOR limitation with this is that all incoming request are being forced into a serial handling in a row to avoid the nonce collisions and transaction failures.
Status brief · December 2025
The problem with this is most applications will send multiple threads to a node for various application handling. It’s a real problem.
Status brief · December 2025
Pay-as-you-go contracts carry a signature and a nonce per transaction, which forces requests into single file. The fix was a lane inside the listener and a queue that releases transactions in clean series. It works and it caps throughput, and both went to the client in writing, along with the proposed remedy: up to five parallel lanes as a dial the user controls.
The March 2026 bridge failure got the same treatment. Root cause: the app never registered IBC client-recovery handlers with the governance module, and the client keeper was constructed without a governance authority. The fix is one constructor change plus an upgrade proposal and a governance vote. “Minimal surface area: only modifies IBC keeper constructor.” And the reason to trust it: “This is the canonical ibc-go v8 pattern used by Cosmos Hub, Osmosis, and Juno.”
The chain is real. The market hasn’t shown up yet. This is the part a case study usually deletes; the 2026 plan put it in writing, so it stays here too:
Thin liquidity makes the token hard to trust.
Arkeo Marketing Goals · 2026
Providers earn in ARKEO. When the token is undervalued and hard to convert, running infrastructure feels like charity work.
Arkeo Marketing Goals · 2026
I know how easy it is to lose heart of faith in where we are going as a project and what we hope the token price can be; I’m in the same boat as building the Arkeo identity is a daunting task.
Status brief to stakeholders · December 2025
Through all of that, the chain and the marketplace kept working, and the problems kept getting written down where the client could read them. “I haven’t built this USDC functionality in- but it sounds slick.” “I was logging too much- I cleaned that up and it’s much leaner and reliable.” Those sentences went to stakeholders during a hard stretch, which is when nobody writes them.
There’s no shipped AI product in this story, and we’re not claiming one. The AI is in how the work got done: a small team ran twenty-two months of engineering and maintenance with AI in the loop, and the outputs are the receipts: an audit closed out finding by finding, a marketplace front end in beta, a production bridge failure traced to a constructor argument. The AI integration on top of the data is designed and waiting on funding.
This page comes out of the contract, a third-party security audit and its retest, the task documents, a December 2025 status brief and the 2026 marketing plan. The latency figures — 138 milliseconds end to end, about seven of them ours — are from that status brief. Arkeo is named here with permission; fees, invoice amounts, token and treasury figures, wallet and contract addresses, endpoints, and the names of individual provider operators are not published, here or anywhere.
The audit report is theirs to publish. Monethic found the eighteen, Monethic retested the fixes, and the final report is a third-party document Arkeo cleared for public release.
Where the evidence stops. The bridge failure is traced to root cause with the fix specified in the task doc, and that is the last of it in our hands — whether the March 2026 upgrade passed is not in our evidence. The AI layer is designed and waiting on funding; nothing on this page claims it shipped.
Everything above this line is in a file somebody can open. Nothing below it is. This is how we would run the same engagement today, with AI in the work — written as an opinion, because that is what it is.
Eighteen audit findings, all resolved, is a workload AI genuinely shrinks: static analysis, test generation around each finding, and a model that will read a codebase for a class of bug faster than a person can grep for it. The reward distribution computing the wrong number, and the testnet signature that would have replayed on mainnet, are both the kind of defect a well-pointed agent can be made to hunt. Twenty-two months of documentation written as we went is now a thing that writes itself.
It would not have hired the auditor.
The line on this page is that we didn’t grade our own homework, and that rule gets more important with AI in the builder’s seat, not less. A model will tell you its own code is fine, in detail, at length, and it has never once been embarrassed. Non-determinism from local time on a validator is exactly the class of problem a confident summary glosses over. Somebody independent still has to look.
Most AI plans sit exactly where Arkeo’s does: designed, waiting on somebody to say what’s actually buildable now. That’s the assessment. Two days, no charge, answered in writing.
Let’s find outOr see the rest of the record: every wave, same job.