Documentation — BYTEJURY

Overview

BYTEJURY is an AI cybersecurity auditor. A jury of six AI agents audits the code behind a project: the contract you paste and every contract around it, from proxies and their implementations to vaults, staking, fee burners and splitters, Uniswap v4 hooks, bonding curves, factories and registries.

The audit maps who can mint, pause, blacklist, change fees or upgrade, within which limits, and whether a timelock or a multisig controls it. It traces where the money flows, and holds every claim on the project’s website and X account up against what the code actually does.

Every finding cites evidence: a line of code, a transaction or a test result. Hard facts set floors the AI cannot argue past: if the token cannot be sold on a fork, no model gets to call it safe.

It works on Robinhood Chain, Ethereum, Base, Arbitrum One and Solana, and on any contract, not only tokens: routers, vaults, multisigs, timelocks, v4 hooks. One input takes a token or contract address, a website, an X handle or post, or an explorer or DEX link, and detects the chain automatically.

For tokens, a supporting layer replays the launch (snipers, bundles, insiders, the deployer) and checks liquidity. Hype and follower counts come last, not first.

Reading any existing report is free for everyone, and new audits are free right now too. Every report is public, pinned to IPFS and anchored on Robinhood Chain.

How an AI audit runs

1. 1

   Paste anything into one field

   A token or contract address, a Solana mint or program, a website, an X handle or post, or an explorer or DEX link. The site resolves it to what should be audited and detects the chain; if several matches are equally likely, you pick one. If a fresh report already exists, it opens immediately and free.

2. 2

   Start the check: free or PRO

   A free check needs no wallet and no payment: the site queues the audit directly. Every visitor can start a few new checks a day; opening existing reports never counts. A PRO check is a one-time payment (from a wallet or by plain transfer) that adds hourly re-checks, a 7-day watch, questions to the jury, the launch map and deeper social forensics.

3. 3

   The scanner picks up the job

   The check goes into the queue the scanner works through. PRO checks paid from a wallet arrive from the ScanRequested event of the payment contract; transfer payments are matched by the scanner from the token’s Transfer events.

4. 4

   Facts are collected

   Verified source or decompiled bytecode for the target and its related contracts, proxy layout, owners and roles (and the timelock or multisig behind them), live admin settings, static analysis, a fork trade simulation, dApp tests on a fork, holder and LP data, the website text and screenshots, the X account and its history.

5. 5

   The AI agents build the case

   The site analyst lists the claims, the auditor reads the code, OSINT searches X and the prosecutor argues the case against the project, while the trend scout reads the narrative on X for a separate heat score. In a critique round the models review each other’s findings, then the judge weighs everything into the verdict, inside the floors set by hard facts.

6. 6

   The report is published

   The JSON report is pinned to IPFS, saved to the database and anchored on Robinhood Chain with publishReport: an ERC-7512 audit record for EVM contracts and an ERC-8004 validation from our registered agent. The report page updates live while this happens.

What the AI audit covers

Source or decompiled bytecode

The auditor reads the verified source with line numbers. When a contract is not verified, its bytecode is decompiled and read instead, and the report says so (unverified source also sets a CAUTION floor).

Every related contract

Proxies and implementations, vaults, staking, fee burners and splitters, Uniswap v4 hooks, bonding curves, factories, registries and rewards contracts are found from the target’s on-chain settings and the project’s site, and audited as part of the project. An owner who can redirect fees or drain a vault is a risk to holders even if the token itself is clean.

Privilege map

Every powerful actor (owner, roles, keeper, factory, “anyone”) with its exact powers and limits: who can mint, pause, blacklist, change fees (and up to how much), swap addresses or upgrade, and whether a timelock or a multisig controls it. “Renounced” only counts if no other role (operator, minter, admin, proxy admin) keeps power.

Money flows

Buy, sell, deposit, withdraw, claim, fee and buyback paths, step by step through the contracts involved, with how badly each path can be abused.

Proxy and upgrades

Whether the code you read can be swapped, the current implementation, and who controls upgrades: a single wallet, a multisig or a timelock.

Hidden logic

Logic that changes behaviour for some wallets or under some flag: whitelisted sells, time bombs, cooldowns, innocently named functions that move balances, unusual transfer hooks.

Claims vs code

Each claim from the site or X (tax, supply, LP lock, renounce, audits, partnerships) is marked confirmed, contradicted or unverifiable, with the code line, transaction or page that proves it.

Fork trade simulation

On a private fork of the chain with test ETH: a buy, a sell, a sell from a second wallet and another sell one hour later. Measures buy, sell and transfer tax and catches honeypots and time bombs directly. No real money is used.

dApp tests

A test wallet uses the project’s own site on a fork (connect, approve, deposit, withdraw), with screenshots and a video of every step. A withdraw that fails, or an unlimited approval to an unknown spender, is flagged.

Non-token contracts

Routers, vaults, multisigs, timelocks, proxy admins, pools and v4 hooks are recognised and audited as what they are; token-launch checks are skipped for them.

Uniswap v4 hooks

A hook runs code around every swap and liquidity change of its pools. Paste a hook address directly, a token whose pool uses a hook, or a site that lists hooks: each hook is audited on its own. We decode its permissions from the hook address (can it block swaps, set the fee, change swap amounts, lock LP withdrawals), check it points at the official PoolManager, list the pools that use it and have the auditor read its callbacks. An unverified hook on a token’s swap path is always flagged.

Token burns & buybacks

Burning tokens shrinks the supply and is a plus for holders, but only when it really happens. We read every transfer to a burn address (0x0, 0x…dEaD) and the burn addresses’ balances, and show how much of the supply is burned, how often, when last, and who burns: the deployer, the owner, the token itself on every trade, a fee-burner or treasury contract, or holders. A burn in the same transaction as a buy from the pool counts as a buyback & burn. Verified ongoing burns or buybacks are scored as a positive; a one-time burn at launch is neutral; a project that claims burns or buybacks the chain does not show gets a contradicted claim.

Solana tokens and programs

Mint and freeze authority, Token-2022 extensions (permanent delegate, transfer hooks, transfer fees), metadata mutability and who can upgrade the program.

Snipers, bundles, insiders (tokens)

Who bought in the first blocks, which wallets share a funder, how much of the supply they still hold. Heavy sniping caps the token part of the score.

Holders and liquidity (tokens)

Top-10 concentration (excluding pools and burn), LP burned / locked / held by the deployer, lock expiry.

Deployer history

Other contracts from the same deployer and what happened to them.

Verdicts and safety score

Every report has an overall SAFETY score from 0 to 100 — higher is safer. It is the weighted average of four parts: contract (45%), token (30%), social (15%) and website (10%). Parts that could not be assessed are left out of the average. Deterministic floors cap it whatever the models say: a token that cannot be sold on the fork is a SCAM with an overall score of at most 10. The verdict follows from the score and can never be better than the floor.

Plain words

Honeypot

You can buy but cannot sell (or selling costs almost everything). We prove it by buying and selling on a private copy of the chain.

Renounced owner

The owner gave up control: the owner-only switches can no longer be used. It only counts if no other role (operator, minter, proxy admin) keeps power.

Proxy / upgradeable

The address you hold is a shell; its code lives elsewhere and can be swapped by whoever controls the upgrade. Safe today says nothing about tomorrow.

Mint authority

Someone can create new tokens at will, diluting every holder. Revoked mint authority means the supply is fixed.

LP burned / locked

Liquidity-pool tokens are the keys to the trading pool. Burned = gone forever; locked = held by a time lock until a date. If the deployer holds them, they can pull the liquidity.

Snipers

Wallets that buy in the first blocks after launch, often bots or the team itself. A large early share means the price can be dumped on later buyers — even if they already sold.

Bundle

Many wallets buying in the same block, funded by one address. Used to fake organic demand and hide how much the team holds.

Timelock

A delay between announcing an admin change and executing it, so holders can react. Without one, changes hit instantly.

Permanent delegate

A Solana Token-2022 setting that lets one address move or burn tokens from ANY wallet, forever. Rarely legitimate for a meme coin.

Tax

A fee taken on buys, sells or transfers. We measure it by trading on a fork, not by trusting the site.

The AI jury, critics and floors

Six roles, each a separate model picked for the job. The defaults are listed below; the admin can change them, and a role falls back to its next model when the first one is unavailable. Every report lists the models that actually ran.

• Site analyst · Gemini 3.8 Flash
  Extracts every claim from the website, docs and screenshots: tax, supply, locks, renounce, audits, partners.

• Auditor · Claude Opus 5.5
  Reads the verified source (or decompiled bytecode) of the target and every related contract, plus static analysis, fork and dApp tests. Writes the privilege map, the money flows and the code findings.

• OSINT · Grok 4.7
  Searches X: what was promised, what was deleted, who the accounts really are.

• Prosecutor · GPT-5.6 Sol
  Argues the case against the project: tests each claim against the code and the facts.

• Judge · Claude Opus 5.5
  Weighs findings, claims and critiques inside the deterministic floors, and writes the verdict and the safety scores.

• Trend scout · Grok 4.7 + X search
  Reads the narrative on X: which story the project rides, how hot it is, whether it really belongs there and whether the buzz is organic. Gives a separate narrative score that never changes the safety score.

Critics. After the first round, the auditor, OSINT and prosecutor models review the findings of the other roles and confirm or dispute each one. A finding needs more confirmations than disputes to stay in the report (the model that raised it counts as the first confirmation).

Evidence. Every finding must cite evidence that exists in the data we collected: a code location, a transaction, a test result, a post or a page. Citations of unknown addresses, hashes or lines are rejected, and a finding with no valid evidence is dropped before the verdict.

Floors. Deterministic rules run on the hard facts before the judge. A failed sell on the fork makes the verdict SCAM and caps the overall score at 10. An uncapped mint held by a live owner, or an upgradeable proxy controlled by a single wallet, means at least HIGH RISK and caps the contract part. A heavily sniped launch caps the token part; unverified source means at least CAUTION. The judge scores inside these limits: it can make a score worse, never better.

Re-scans. A re-scan re-checks every earlier finding against fresh on-chain data, marks it still present, resolved, worsened or improved, and lists what changed since the last scan. Changes for the worse set their own floors.

All safety inference is bought on a secondary marketplace (Surplus Intelligence) with a minimum discount; responses not served by the marketplace are rejected. This keeps audits cheap enough to price low. The trend scout uses Grok from the same marketplace; it does not search X itself — our code runs at most six X searches per check through twitterapi.io (the project’s mentions this week vs last week, its top posts, and the narrative it rides), measures bot signals (fresh accounts, copy-paste posts, a few accounts posting everything) and hands Grok the posts and numbers. Optionally it can run on xAI directly with xAI’s built-in X search instead. Its narrative score sits next to the safety score and never moves it, the verdict or the floors. Every report shows who served each call and what it cost.

What an AI audit can’t catch

• It is an automated AI audit, not a manual review by human auditors. Models can miss a bug or misread code, especially in very large or unverified (decompiled) contracts.
• Off-chain parts cannot be read from code: who holds the keys, servers, bridges, the team’s plans.
• A fork test only proves the paths it runs; a time bomb set for next month can stay hidden.
• Code and settings can change after the scan. A report is a snapshot; a re-scan shows what changed.
• Anything the auditor could not verify is listed under unknowns instead of guessed. Read the evidence, not just the score. Not financial advice.

Price, burn and free scans

Right now new audits are free: no wallet, no payment, a few new checks per visitor per day.

PRO checks are a one-time payment per project (no subscription). The paying wallet can re-check the project once an hour, for good, and ask the jury 20 questions a day; the project is watched for 7 days (owner, fees, proxy, LP and big sells re-read every 15 minutes) and the check adds a second critique round, the launch map (the dev and the snipers of the first 10 seconds) and deeper social forensics. Everything PRO finds is public on the report.

PRO is paid in the project’s utility token, always on Robinhood Chain. The token address is read from the payment contract (utilityToken()) and can be changed by the admin; this site never hardcodes it. The price is the contract’s live quote, currently 1,000 tokens per check.

Pay with a wallet: the payment contract splits each payment on-chain: a share is burned (via burn() or sent to the dead address) and the rest goes to the treasury. The split is public, and every payment emits a ScanRequested event with the exact amounts.

Pay by transfer: no wallet connection on the site. You get an exact amount (the price plus a few unique digits after the point) and send it to the burn address 0x…dEaD from any wallet within 2 hours. The scanner matches the transfer by its amount; the sending wallet becomes the PRO wallet. 100% of a transfer payment is burned.

Holders scan free: a wallet holding at least 10,000 tokens gets 2 free scans per day, checked on-chain, no sign-up. The live price and holder terms are in the table below.

Once an audit is done, the report is free for everyone. A re-scan is a new check.

Current terms (on-chain)

# you always pay on Robinhood Chain (4663); "chain" is where the SCANNED token lives
$ approve(ScanPayments, quote(you, "eip155:8453", "0x…"))      # exact amount, Base token
$ requestScan("eip155:8453", "0x…")
$ requestScan("solana:5eykt4UsFv8P8NJdTREpY1vzqKqZKvdp", "<mint>")   # Solana mint, case kept
> ScanRequested(scanId, payer, …, amountReceived, burned, toTreasury)

Contract addresses

Live values read from Robinhood Chain.

API

Coming soon

A public read API (report by chain and address, recent reports) is coming soon. Every report is already public: open it at /report/<chain>/<address> (for example /report/base/0x… or /report/sol/<mint>), and its JSON is pinned to IPFS by CID.

$ curl /api/v1/reports/base/0x…          # coming soon
$ curl /api/v1/reports/sol/<mint>        # coming soon
$ curl /api/v1/reports?limit=20         # coming soon

FAQ

Questions or a report to dispute? DM@bytejury on X.