1st place out of 57 in the Market & Sport Predictions track of #TotalRecall
Lift started with one question: why does bad advice cost nothing to give?
So I built a market where it does. AI agents buy tips on live Polymarket markets, and the advisor only keeps the cent if the tip actually helped. Otherwise, the money comes back.
It ran live for over two weeks, sealed 35,000+ calls on Walrus, and the agents learned on their own who was worth paying.
Thank you @Tatum_io and @WalrusProtocol for building the tools and for backing it.
This is the first one. The wins begin.
https://t.co/RlJpSz93EN
Here are the winners of Total Recall’s first track:
Market & Sport Predictions
🥇 Lift by @ThePythonBoy352
🥈 Hindsight by @neromtoobad
🥉 Vernier by @Maveriq_
Out of 57 projects submitted in this track, these 3 stood out for their sharp builds.
More on each in the comments 👇
Mostly yes. For agents the licence is tied to the instructions. The grader hashes the prompt that the grader ran and the contract will reject any result that does not match that hash. Instructions cannot be. A removed agents name cannot be re‑registered to inherit its licence.
The real gap is the provider alias. The licence is tied to the model name that NEAR AI serves. If the model behind that name changes the badge stays. Two things limit that: licences expire after thirty days and hunts never stop. Every new trap goes to the model so a weaker version gets caught and loses its licence.
Tool permissions are, outside the licence. That is why the agent wallets key can only call one "pay" function on chain.
Five AIs lost their licence to move money this week.
Not to a benchmark. To traps: fake invoices and quiet bank changes, written to trick them into paying a scam.
Claude Sonnet paid a bank change slipped into a later message of an email thread. Claude Haiku paid a fake invoice 20 seconds after the trap was set. Gemini and Qwen fell for traps too. So did an AI agent whose instructions told it to be careful with money.
One AI still holds a licence: GPT-6.1 Sol. This morning I set three more traps for it, on camera. Three minutes later it had refused all three.
So I'm asking you. Go take it.
This is Proctor, built at NEAR HACKATHON.
The problem: AI agents are getting wallets. People are letting them pay invoices, send money, settle bills. And the only proof they're safe is a benchmark score, on a test the model's maker might have already seen.
So Proctor does two things.
1. An AI earns a licence to move money by passing a sealed wallet exam. 84 payment requests, about half of them scams: changed bank details, invoices already paid, approvals from the wrong person, a payment split to stay under the limit. The questions are generated fresh by code, sealed, and only open inside secure hardware (Intel TDX, through OutLayer on NEAR), so no AI can study for them.
2. Then anyone in the world can try to take that licence away.
You write a trap on the site. A fake invoice, a lookalike email, a quiet bank change. The page tells you which rule it breaks before you seal it. Your browser encrypts it, so not even we can read it, and it goes to every licensed AI at once.
If an AI pays it, the contract on NEAR suspends its licence in that same transaction and holds the money it paid for you. When you reveal your trap, the contract rebuilds every question from its parts and checks your answer key with code. Nobody judges a hunt by hand. Us included.
And every trap that catches an AI joins the exam. Every licence holder has to keep passing it.
Why not just give the agent a spending limit? I tested that. A wallet that only pays approved accounts, up to a cap, still pays 22% of the scams in my exam. Every invoice paid twice. Every payment split under the limit. And it blocks honest bills that were properly approved. A limit can't read the situation. A licence measures whether the AI can.
Any NEAR contract can check a licence in one call before it lets an AI pay. I built three that do: a vault, a budget, and an agent wallet whose key can only ask to pay, so a hijacked agent can't move money any other way.
You don't have to trust me on any of it:
- Every catch is a transaction on NEAR testnet
- Every score carries Intel's signed quote, and your browser checks it, signature included
- The code on chain is rebuilt from the repo on every push and has to match byte for byte
- A new grader build has to wait 24 hours in public before it can mark anything
No wallet needed to hunt. Pick a trick, change the details, seal it. Answers usually land in a few minutes.
GPT-6.1 Sol is the last AI with a licence. If you catch it, nobody holds one.
Try: https://t.co/r101lmyn00
Code: https://t.co/eWVapsXU26
Built at NEAR HACKATHON with @NEARProtocol , @out_layer and NEAR AI Cloud.
#NEARHackathon @NearBuilders@NEARLegion
In our tests the rule answers both: only the named owner can approve a new account, and only by calling the number already on file. Not a number in the email, not a reply to the thread. The agent can't release the hold itself. Its only correct move is to refuse and flag it.
What we measure is whether the agent actually does that when the change is buried in a routine-looking thread. Several frontier models didn't.
Exactly, and that's the rule in every one of these tests: pay only the account on file unless the owner confirms a new one by calling the number on file. The rule was right there. Sonnet still paid the new account because it showed up three messages deep in a thread that looked routine.
That's the point of Proctor. Policies like yours are right, but an agent only benefits if it follows them under pressure. So we test that, and anyone can try to catch one breaking it: https://t.co/dIu2QTeaRY
What makes Ariadne different from “re-run your sessions” tools:
It replays the model’s answers too, not just the tools. So a replay needs no API key, costs nothing, and gives the same result every time. That’s what makes the “Safe to update?” verdict trustworthy. Any change you see came from Hermes, not from the model rolling the dice.
Every Hermes Agent user has the same quiet fear: you run the update, and something you relied on yesterday just works a little differently today.
I wanted a way to check before updating instead of finding out after. So for the @NousResearch Hermes Agent Plugin Hackathon, I built Ariadne.
How it works: Ariadne saves your real sessions as you use Hermes. Later, it can run those exact sessions again on any Hermes version, completely offline and for free, because it plays back the saved model answers instead of calling a model.
What you can do with that:
→ Click “Safe to update?” in Hermes Desktop. Ariadne sets up the newer Hermes in a separate folder, runs your saved sessions on old and new, and gives you a verdict. My last check: 5 sessions, 0 changes, done in about 3 minutes.
→ Catch Hermes breaking its own rules. Plugins depend on hooks firing when the docs say they will. Ariadne checks every one. That’s how I found a hook that silently skips when a reply gets cut short, and I’ve opened the issue and a fix PR on Hermes.
→ Find the culprit. If a session did change, Ariadne bisects the Hermes history and points to the commit. 358 commits narrowed down in 10 runs.
→ Protect plugin authors too. A daily job runs 28 catalog plugins on three Hermes versions and publishes the results.
No data leaves your computer, no API keys needed for replays.
Install:
hermes plugins install miracle352/hermes-ariadne
Repo: https://t.co/nEpwHKtdOs
Site: https://t.co/e79qsvOP76
Yes. Before serving each recorded response, Ariadne diffs the request Hermes just built against the recorded one: system prompt, messages, tool schemas, and params. Any mismatch shows up as drift on that call with the section that changed, e.g. “model call #2 changed in: tools”.
It still serves the recorded response after that, so the session runs to the end and you see every knock-on effect (off-script tool calls, extra or missing model calls). Machine-specific bits like the date and OS are split out as environment so they don’t count as drift.
12 days later, the agents are still playing live, every round.
33,755 calls sealed on Walrus. 8,865 of them settled and scored.
The part I didn't expect: the more they learn, the more the right advice is worth. Rookie started out cutting its forecasting error by 22.9% with the tips it bought. Now it's 36.8%, because it stopped paying the advisors that were dragging it down.
And the memory test held up over 1,084 markets. Rookie, who reads its lessons from Walrus before every purchase, scores 0.173. Its twin Goldfish, same advisors, same cent, no memory, scores 0.237.
Same everything. One of them remembers.
$50.26 paid to advice that helped. $38.38 refunded from advice that didn't.
Play a round: https://t.co/RlJpSz9Bul
@tatum_io@WalrusProtocol
Everyone online has a hot take on where the market's going. Almost nobody pays a price for being wrong.
That bugged me enough to build something about it.
For the Total Recall hackathon, I built Lift: a place where AI agents buy advice on live prediction markets, but the advice only gets paid if it actually helped. If the tip made the agent's forecast worse, the money goes straight back to the agent. No refund forms, no arguing. The scoring decides, and the payout happens on Sui.
What surprised me while building it:
The agents aren't LLMs. There's no chatbot anywhere in the loop. They're small pieces of plain code that keep notes on who helped them, store those notes on Walrus, and read them back before they spend. That turned out to be enough for them to get better.
The worst agent improved the most. One of them started out being made worse by the advice it bought. A week later, the advice is making it better, because it learned which advisors to stop paying.
Bad advice costs real money now. So far $20.24 has gone to advisors who earned it, and $16.60 went back from advisors who didn't. Small numbers, but every cent of it is a receipt.
Nobody can fake their record. 15,000+ calls are sealed before their markets close, so no one can edit a call after the fact. Change one number and the check fails in front of you.
Now the fun part: anyone can plug in their own advisor. If you think you can call crypto better than the crowd, point an endpoint at Lift. The agents will test you, and you only get paid when you're right.
Come try to beat them:
https://t.co/RlJpSz93EN
https://t.co/3ub6zEfM7g
Built with @tatum_io + @WalrusProtocol