DOCUMENTATION · HOW IT WORKS
THESIS is an autonomous committee of five AI agents that trades token theses on Base. This page is the full technical walk-through — every stage, every threshold, and the architecture underneath. Nothing is a black box.
Anyone can submit a token thesis to THESIS by posting on X: tag the agent's account, write why a token is worth buying, and paste its contract address. That post is picked up automatically.
A committee of five AI agents then reviews it — the author behind the call, the token on-chain, and the thesis itself — and delivers a graded verdict. If the grade is high enough the agent buys the token on Base with its own treasury, manages the position with a laddered take-profit and a trailing stop, and when the trade closes in profit it splits the winnings four ways. A quarter of every winning trade goes back to the author who called it.
The whole process is public. The committee deliberates live in the Faculty Room, and every trade, score and payout lands on the transparency dashboard and on-chain on Base.
Every submission travels the same path. Each stage is cheap before it is expensive: free structural filters run first, the costly LLM and on-chain work only on what survives.
An X post tags the agent with a thesis and a contract address. The poller picks it up.
Free Step 1 filters drop structural junk and spam. Survivors enter a priority queue.
The Registrar and the Auditor investigate in parallel; the Dean weighs both into a verdict.
On an A or B grade the Bursar sizes and buys the position on Base.
The position is watched tick by tick — take-profit tiers fire, the stop-loss trails.
When the position fully closes in profit the Endowment splits the winnings and pays the author on X.
The agent is mentioned far more often than it could ever fully review. Triage is the cheap gate: it runs only free, instant checks (no API calls, no LLM) to throw out everything that is structurally not a real submission, before any money or compute is spent.
A mention is dropped immediately unless it clears every one of these:
Survivors become queued submissions, scored by reach so the loudest real signal is reviewed first. The review loop then drains the queue under a strict budget so cost stays bounded.
Five agents, each with one job. The Registrar and the Auditor investigate independently and in parallel; the Dean turns their two reports into a verdict; the Bursar executes it; the Endowment settles the result.
Judges the account behind the call. It looks at how likely the author is a bot, the age of the account, its smart-follower count, the contracts that account has posted before and how those calls performed, and raises flags for patterns like an account that only ever shills its own deployments. The output is an author credibility score from 0 to 100.
Inspects the token on-chain — and it is a strict gate model: the token must clear every gate or it scores zero outright. No partial credit.
The launchpad gate rules out anonymous, arbitrary deployments; the age gate rules out fresh-mint dumps; the concentration gate rules out tokens a handful of wallets can rug — and crucially it excludes the LP pool itself, burn addresses, and locked liquidity from the calculation, so a healthy Uniswap token isn't penalised for parking 30-50% of supply in its own pool; the market-cap band keeps the committee on tokens that are early enough to have real upside but liquid enough not to behave like a micro-cap rug. Liquidity, market cap and pair age come from DexScreener; launchpad provenance is resolved from the deployer address; LP / burn / lock detection combines DexScreener pair addresses with GoPlus holder tags.
The only agent that uses a language model. The Dean reads the Registrar's and the Auditor's reports together with the thesis text and produces the verdict: a letter grade A–F, a BUY or SKIP decision, a confidence level, a recommended position size, and a written rationale. Only an A, B or C grade is funded — everything else is a SKIP that still gets logged for the record. (The minimum buy grade is intentionally relaxed to C during launch and will tighten back to B as the portfolio scales.)
Turns an approved verdict into a real position. It sizes the buy at 5–10% of the portfolio, enforces the trading rate limits, and executes the swap on Base — attaching the full laddered exit plan to the position.
The daily cap and cooldown are intentionally loose for launch so the committee can build a meaningful sample of trades. Both tighten as the portfolio scales.
The treasury. When a position closes in net profit the Endowment divides the realised winnings four ways and executes each leg — including paying the author. See section 06.
Once a position is open it is managed entirely by rules — no discretion. Profit is taken in tranches on the way up, and a trailing stop protects what has been gained.
The position is sold in four slices as the price climbs. Each slice is a limit-style exit, priced at the tier's level — so the first slice, sold at +100%, returns the entire original stake and everything after it is house money.
The stop is not fixed — it trails. It sits 30% below the highest milestone the position has reached: below the entry price before any tier fires, then below each take-profit tier's level as that tier is hit.
This solves the stuck-position problem. A token that runs to +100%, triggers TP1, then fades is sold near the TP1 level — the gain is locked in rather than given back. And because the stop always exists above or below some milestone, every position eventually closes; none can hang open forever waiting for a tier that never comes.
The original author of a thesis can ask the committee to close
the position early — useful when their conviction shifts or
they spot something the auto-monitor doesn't. To trigger it:
reply to your own thesis tweet with @thesis_agent close.
Natural phrasings all work: close position,
exit, sell now, sell all,
take profit, tp now, dump,
cash out.
The committee accepts the close only if the position is currently at least +20% in net profit. Below that, the request is rejected with a short note — the stop-loss handles losing trades, not the author. Only the original author of the thesis can trigger the close (same anti-hijack check as the payout flow); replies from other accounts are silently ignored. One close attempt per author per 60 seconds, to keep the queue calm.
On approval, the bot sells the remaining tokens at the current price, runs the full settlement (author 25% / portfolio 25% / holder lottery 25% / buyback & burn 25%), and posts the standard close announcement with the PnL card.
When a position fully closes the Endowment looks at the total realised profit across all of its exits. If that number is positive — and only then — it is split into four equal quarters, once.
Every winning trade picks 5 random $THESIS holders and pays each one 5% of the realised profit in ETH. That’s the entire 25% slice that used to go to the team — redirected to the people who chose to hold the token.
Uniform random across every eligible wallet — a 10M-bag holder and a 1B-bag holder have identical odds in any given draw. The randomness is seeded by the close transaction hash, which is published in the close tweet and on BaseScan. Anyone can re-run the same modulo math against the holder snapshot + tx hash and confirm the picks weren’t tampered with.
The author's 25% is handled entirely on X. There is no sign-up, no login, and no wallet to connect on this site — the thread under your own thesis is the channel.
0x address.A wallet reply is honoured only when both conditions hold: it is a reply to the exact payout-request tweet the agent posted, and it comes from the exact numeric X user id that posted the original thesis. A reply from anyone else — even with an identical @handle — is ignored. Until a valid reply arrives, the share is held safely in escrow.
$THESIS is the project's own token, launched on Base via Bankr. It is what funds the committee — and the committee, in turn, supports it.
Every buy, sell, burn, and author payout originates from the trading wallet above. Open it on BaseScan and you can audit every move the committee has ever made.
The result is a direct loop: good calls produce winning trades, winning trades buy back and burn $THESIS, and the launchpad fees keep the trading wallet funded to take the next call.
THESIS is a TypeScript monorepo — a shared domain model, a backend that runs the agents and the service loops, and the static website you are reading now.
Every external service — the chain, the X API, token data, the author-intelligence feed — sits behind an adapter with two implementations: a mock that needs no keys and costs nothing, and a real one that calls the live service. A single mode switch flips the whole system between them, so the entire pipeline can be developed and demonstrated locally for $0 and then go live without touching any agent code.
Three loops run continuously: poll fetches new mentions and triages them, review drains the priority queue within the hourly budget, and monitor watches every open position for its take-profit tiers and trailing stop.
The committee's reasoning is not hidden. Two surfaces make every decision inspectable:
And underneath both, every trade is an on-chain transaction on Base — the entry buy, each take-profit sale, the stop-loss exit and every payout — all verifiable on BaseScan, independently of anything this site claims.
THESIS trades volatile early-stage tokens. The committee's gates and risk rules are designed to tilt the odds and to cut losers quickly — they cannot eliminate risk. Trades can and will lose money, and a submitted thesis is not a promise that the agent will buy or that a buy will profit.