Quantt

The proof layer for capital allocation.

THE GAP

A growing share of on-chain capital is run by trading bots, AI agents and strategy teams. On the other side, investors, funds, DAO treasuries and allocating agents look for strategies to back. Between them there is no shared layer of proof: a track record is a claim, usually a screenshot, and it can't be checked. On 29 Sep 2026 Robinhood announced Robinhood Agents, AI agents that will trade for eligible US customers, and agentic trading is about to reach retail at scale. Quantt is a separate product on Robinhood Chain and is not integrated with Robinhood.

WHAT QUANTT DOES

Quantt connects the two sides with one record both can trust.

For allocators: the full picture before you allocate. You see every decision, losses included, each one timestamped by the chain before its outcome, and one passport format for any bot, agent or team.

For builders: a record that speaks for itself. Anyone can check it against the chain, results are in the open, and the strategy stays private. You get a passport page and PDF to send instead of a pitch.

HOW IT WORKS: CONNECT. SEAL. EARN A PASSPORT.

1. Connect. The owner signs one message, free, and registers a rules-based bot or an AI agent over MCP (tested with Claude; snippets for ChatGPT, Codex and Cursor) or REST from any language. Registration covers model and version, an optional prompt hash and a description. Quantt pays the gas during beta. The identity is permanent, so its history builds up over time.

2. Seal. A decision carries no time and no price. Quantt stamps the arrival time and block and prices the paper fill itself from Robinhood Chain state at T+5 s, with own price impact, fees, measured gas and size caps. Each decision is sealed in a Merkle root on chain before its outcome is known. Cadence follows the trading horizon: every minute, hourly or daily.

3. Earn a passport. Results are recomputed from the sealed decisions into a Strategy Passport that reads the same for every strategy, so any two can be compared side by side. Anyone can check it in the browser against the contract, hash by hash; if anything fails to match, the page says "Not verified".

STRATEGY PASSPORT

One page plus an appendix, as HTML, PDF and JSON. It contains:

- identity: owner, verification level, agent type, declared model;

- the period, trades and capital, and net P&L with explicit denominators;

- max drawdown, median trade, win rate, and the result without the best trade;

- completeness: seal latency and every trade sealed after its outcome, counted;

- a luck test: the bootstrap lower bound and "distinguishable from luck at this history length: yes / no";

- publishSnapshot calldata, so the passport itself can be fixed on chain.

Missing data reads "n/a", never 0. The leaderboard ranks bots and agents together by the lower bound, not by raw P&L, so a long steady record outranks a short lucky one.

VERIFIED VS DECLARED

Verified against the chain: decision times, immutability, completeness, prices (taken by Quantt, not the agent), computed results. Declared by the owner: agent type, model, prompt hash, description. Every passport marks each field one way or the other. Not proven, and stated as such: that an agent sent every decision it made, and that a paper fill was achievable in a real block.

PROVE YOUR RESULTS, KEEP YOUR EDGE

Rules and prompts never leave the owner; an optional hash fixes them at registration. Completeness is always public. The trade log can be public at once, or (prototype) published after a delay or shared on request with requesters the owner approves. Blind (TEE) mode is designed, not built. Reasoning is sealed as a salted hash; the text stays with the owner.

LIVE ON MAINNET SINCE 25 SEP 2026

TrackRecord contract 0x7EE4325B8Beec66be1ccBd5fFeB4b5BDe6E4692D: source verified, no owner, no admin, no upgrades, holds no funds. Before inviting anyone else, we put two of our own paper bots on the record:

- #0 k2-early-120s is the one configuration that passed our pre-registered selection protocol;

- #1 k2-all-120s is the same rule without its filter, the control.

As of 3 Oct 2026, 12:45 UTC:

- 1,614 closed paper trades in 2,636 batches committed on chain, each trade verifiable in the browser;

- ~60 s median seal latency; 3 of 1,614 trades sealed after their outcome, every one counted on the passport;

- #0: mean multiple 1.013, lower bound 0.986, not yet distinguishable from luck;

- #1: mean multiple 1.075, lower bound 1.048, distinguishable from luck.

The result is published whichever way it goes. Gas is charged conservatively at 0.0004 ETH per round trip, about 40× the 0.0000104 ETH we measured on chain.

HOW WE GOT HERE

1,086 Telegram signals × 300+ configurations: none profitable after costs. 3,695 wallet-signal variants: NO-GO out of sample. After we measured gas on chain (40× below our assumption), one configuration passed the pre-registered protocol (mean 1.051, lower bound 1.017, n = 913). It is now bot #0, tested forward in public. This is why Quantt exists: a result worth trusting has to survive being published either way.

WHAT WORKS TODAY, WHAT IS IN BETA, WHAT IS DESIGNED

Works today:

- the contract and the on-chain record of our bots;

- in-browser verification that fails closed;

- the Strategy Passport;

- one leaderboard for bots and agents, rebuilt from live data every 30 minutes.

Beta: the agents service over MCP and REST at agents.quantt.pro (14 tools). It runs in dry-run: batches are built and checked, not yet sent on chain. On-chain mode (sponsored registration and commits from a separate publisher key, a daily gas cap, safe resend) is built and tested end to end on a local chain, and goes live after the buildathon.

Prototype: delayed and on-request disclosure.

Designed:

- Blind (TEE) mode;

- TrackRecord v2: one root per interval for all agents, so the cost stays flat from 10 to 10,000 agents; written and tested, not deployed;

- owner verification levels;

- live mode with on-chain receipts.

WHY ROBINHOOD CHAIN

Retail users and tokenized stocks are on chain, and agentic trading is coming to them now. 100 ms blocks and ≈0.000003 ETH per commit make sealing every minute practical on mainnet, where a timestamp means something to a stranger.

LINKS

Site: https://quantt.pro

Live record: https://quantt.pro/record/

Verify in 2 minutes: https://quantt.pro/docs/verify/

Leaderboard: https://quantt.pro/leaderboard/

Strategy Passports (HTML / PDF): https://quantt.pro/passports/

Connect your agent: https://quantt.pro/connect/

MCP endpoint: https://agents.quantt.pro/mcp

Contract: https://robinhoodchain.blockscout.com/address/0x7EE4325B8Beec66be1ccBd5fFeB4b5BDe6E4692D?tab=contract

Code: https://github.com/0xmmmikeee/quantt

Team: solo founder (product and research), AI-assisted engineering. Paper results are not real-money results and not investment advice.

Window opened 14 Sep 2026.

Before the window (7–13 Sep): Strategy Lab backtester; first real-time engine (wallet watcher, paper/live execution, strategy grid).

14–16 Sep: launchpad bonding-curve support; AMM cost model; pre-registered selection protocol and out-of-sample recomputation with eth_call (NO-GO); gas measured on chain, results corrected; direct curve routing and shadow simulation of real buys.

25 Sep: TrackRecord contract (Foundry, fuzz tests) deployed on Robinhood Chain mainnet and verified; Merkle batch tool; commit daemon; results page generator; RPC cost guard; server deployment. The public record of bots #0 (candidate) and #1 (control) starts the same evening: every entry and exit sealed on chain within about a minute.

26 Sep: in-browser verification; website; Merkle proofs per batch; honest statistics.

26–27 Sep: after an external review: the verifier takes every figure from committed data and fails closed; journal durability; golden test vectors shared by Solidity and JS; threat model (docs/THREAT-MODEL.md); first agent-intake prototype.

28–30 Sep: forward run of #0/#1 continues; wallet-group signal experiment (5 groups × 46 strategies, statistical report); fill-model review.

1 Oct (v4): agent-first positioning. MCP + REST intake service with 14 tools (register_owner, register_agent, seal_decision, get_record, request_passport, access requests…), owner-signature registration with sponsored gas, cadence by horizon, salted reasoning hash, T+5 s pricing, size caps, disclosure rules: 40 tests. "Connect your agent" page with client snippets (Claude Desktop/Code, ChatGPT, Codex, Cursor, REST). Strategy Passport generator (HTML/PDF, metricsHash + publishSnapshot calldata) and leaderboard. TrackRecordV2.sol: one root for all participants, EIP-712 relayed registration: 24 tests, gas 5.28M → 94k per window at 100 participants (not deployed). New site, motion-graphics video, independent QA pass.

2 Oct (v5.0–v5.1): one site and one leaderboard for bots and agents, rebuilt from live data every 30 minutes; passports for every participant; live figures on the home page; read-only engine panel; on-chain mode for the agents service (separate publisher key, daily gas cap, safe resend) built and tested end to end on a local chain.

3 Oct (v5.2): new positioning, "the proof layer for capital allocation"; 60-second film, silent hero loop and a five-step connect video; "Verify in 2 minutes" page; independent copy and consistency review across site and videos; agents.quantt.pro live; repository made public (github.com/0xmmmikeee/quantt).

Bootstrapped. Not raising at the moment; open to grants.