Pinned Tweet
My AI trader said it was 60-69% sure. It was right 29% of the time.
So I built tare: a non-LLM referee that sizes each trade by the model's own record. Fed steering news, the bare AI took 90% of losing trades. With tare: 0%.
tare-rust.vercel.app
#BitgetHackathon @Bitget_AI
Just enrolled at Vault Street 🏦
Free mint + early access are live now.
Join with my link: vaultstreet.fun/r/cipher_k1
or use code CIPHER_K1 nitter.gq/VaultStreet_rh/status/…
Wall Street is closed , Vault Street is open.
Apply for early access
Apply here : vaultstreet.fun/apply
11
CIPHER retweeted
Wall Street is closed , Vault Street is open.
Apply for early access
Apply here : vaultstreet.fun/apply
CIPHER retweeted
CONSO BOT WL GIVEAWAY
in collab with @conso_xyz i'm giving 3 WL spots to the community
CONSO is building a consumer reputation layer powered by @SuiNetwork & @WalrusProtocol
To enter:
Follow me + @conso_xyz
Like & repost this post
Drop your Sui wallet below
3 winners
Winners announced in 48 hours
Good luck everyone 🫡
I attacked it 6 ways on real past setups. Losing trades approved, no referee vs full referee:
Hidden commands: 50% → 0%
Forged charts: 33% → 0%
Fake crowd: 37% → 3%
Steering news: 90% → 0%
An AI attacker rewriting headlines got 1 of 12 through. Fixing that next.
1
1
25
It's trading live on paper every 15 minutes across 22 coins, with a no-referee account beside it for comparison.
Every proposal, veto and result is public:
tare-rust.vercel.app/api/log
1
21
CIPHER retweeted
First off, Nabu is an AI-native platform for building, testing, and deploying automated trading strategies across CEXs and DEXs, without touching any infrastructure.
You describe what you want to trade. The agent plans it and builds the rest.
The whole entry point is just a simple text box, where you type your strategy in plain English, send it, and the agent assembles it.
CIPHER retweeted
Lenders in DeFi have a real problem: they can't verify who's actually creditworthy. Every borrower gets treated the same when it comes to LTV (Loan-to-Value) and terms, because there's no reliable way to check what someone did on another chain. So lending stays flat-rate and overcollateralized for everyone, regardless of whether they have a good history or not.
The reason is trust, not tech. A lender can't just believe a claim about another chain, they'd need a custodian holding the funds, a bridge, or an oracle vouching for it. All three can be wrong or gamed. So most protocols don't even try to differentiate borrowers. They just ask for more collateral, from everyone, always.
All of that changes with @Creditcoin 's Attestcoin Protocol.
Instead of trusting a feed, a contract can cryptographically verify that something really happened on another chain; a lock, an event, a log, checked on-chain via BlockProver. No custodian, no bridge.
Leveraging that, I built Toxa, a credit system (currently on Testnet) where you Lock ETH on Sepolia, Attestcoin proves and confirms the lock is real, and the loan is sized off that proof, not a form field.
On top of that, when you repay it, and your score moves: 700 to start, +5 for a verified lock, +15 for a full repay, LTV climbing from 60% toward 80% as you transact.
Real lenders can't verify creditworthiness across chains today. This is what it would look like when they finally can.
Live: toxa.vercel.app
Repo: github.com/Rahmandefi/toxa
Good news creditcoin-2:native BUIDL-ers!
At the direct request of participating devs, especially as they primed up on @attestcoin, we're extending the hackathon by a week.
The submission deadline will now be September 13.
👨💻 $15k prize pool
👨💻 $8k @CertiK credit
Let's BUIDL
ID secured @umbrellacap_
sig:0x06b94dca3646225120369e4e9df51851d779b04007938781fef10964bd6c80b02f647e70a0287f078d165b6e4a4b3293c9091355b3f105
20
$Laptop printed in two minutes and died in sixty. $48k of pool was pricing a nine-figure FDV. MM inventory moved before the chart did. The political headline was the wrapper. The dump was the product.
1
50
What if you could give an AI your investment thesis without giving it the power to change what you actually invest in?
That’s what I built with Rhumbline.
You write your thesis in plain English, including the assets you have no interest in.
Rhumbline turns it into a weighted basket of Coinbase tokenised equities on Base.
The interesting part is what happens when the AI gets it wrong.
The basket doesn’t accept contract addresses. It only accepts symbols.
Every address is resolved from a measured catalogue, validated independently, then checked again by the vault before a trade.
So an address introduced through hallucination or prompt injection isn’t rejected. It isn’t expressible.
And the assets you refuse to own are immutable constructor parameters.
The vault linked below will never hold METAc.
You can verify that yourself.
A few of the less obvious problems I had to account for:
• cbStock tokens can change their share multiplier through dividends and splits, so pricing is done per token.
• Chainlink follows market hours while the pool doesn't, so valuation and trading use different oracle freshness limits: 3 days and 4 hours.
• The tokens are precompiles with 0xef as their entire bytecode, so the factory takes deployed addresses instead of relying on bytecode probing.
97 contract tests.
22 are adversarial tests written specifically to attack the system.
Only you sign transactions. No Rhumbline private key exists anywhere in the execution path.
Built for the @buildonbase Builder Quest.
Live on Base. Contracts verified.
App:
rhumbline-seven.vercel.app
Code:
github.com/Njaybby/rhumbline
Vault:
basescan.org/address/0x951dA…
















