Decoding · November 18, 2026 · 7 min read

Spotting new accounts in the raw transaction stream

A handful of instructions create Solana accounts, and each says something different about whether the account is really new. Here is the list, what each proves, and how to confirm it afterwards.

To detect new Solana accounts from raw transactions, watch a short list of instructions: System CreateAccount and CreateAccountWithSeed, the associated token program's Create, and SPL Token's InitializeAccount variants. Each one is a request, not proof. The transaction can fail, and CreateIdempotent succeeds on existing accounts, so confirm the account with RPC once the block settles.

Instructions that create accounts, and what each proves#

ProgramInstructionTag in instruction dataWhat it tells you before execution
SystemCreateAccount0, as a 4 byte little endian valueA new account is requested at the given address, with lamports, space and an owner program
SystemCreateAccountWithSeed3, 4 bytesThe same, at an address derived from a base key and a seed
SystemTransfer2, 4 bytesLamports to a destination; if that address holds nothing yet, this is how a plain wallet first appears
Associated Token AccountCreate0, 1 byte, or empty instruction data, which the program also treats as CreateA token account for a wallet and mint is requested; the program fails if it already exists
Associated Token AccountCreateIdempotent1, 1 byteA token account is requested "if missing"; it succeeds either way
SPL TokenInitializeAccount, InitializeAccount2, InitializeAccount31, 16, 18, 1 byteA previously allocated account is being set up as a token account

The tags come from each program's own instruction source, listed below. System program tags are four bytes because its instructions are serialised with a 32 bit enum index; the token programs use a single byte.

System program: CreateAccount and CreateAccountWithSeed#

These two are the clearest signal. The instruction names the funding account, the new account, the lamports to deposit, the space to allocate and the program that will own it. CreateAccount fails if the target address already holds lamports, so a successful CreateAccount does mean a new account.

The owner field is the useful part for most watchers. An account created with the token program as owner is about to become a token account or a mint; one owned by another program is that program's state. Reading SOL transfers uses the same System layout, so one decoder covers both.

Token accounts: why CreateIdempotent proves nothing new#

Most wallets and bots open token accounts through the associated token account program, which derives the address from the wallet, the token program and the mint. Associated token accounts explains the derivation, which lets you map a token account back to its wallet without any RPC call.

Two instructions do this, and they differ in one way that matters here. Create fails when the account exists. CreateIdempotent succeeds anyway and does nothing. Bots send CreateIdempotent before almost every trade as a safety step, so seeing it says nothing about whether the account was new. Treat it as "this wallet touches this mint", not "a new holder appeared".

The token program's own InitializeAccount instructions are the lower level path, usually after a System CreateAccount in the same transaction. SPL Token instructions lists their tags.

A wallet's first transaction rarely says create#

Ordinary wallet addresses have no create step. A system account exists as soon as it holds lamports, so the first sign of a new wallet is usually a plain Transfer into an address you have not seen before. To call it new, you need history: your own index of addresses already seen, or an account lookup made before the transfer settled. Neither comes from the transaction itself.

What raw data cannot tell you here#

Account creation often happens inside another program. The associated token program creates the token account by calling the System program itself, and many programs create their state accounts the same way. Those calls are inner instructions, recorded only in the transaction's metadata after it runs. A decoder working on raw shreds sees the outer request, not the inner creation, so for program owned state you detect the program's own instruction and learn the layout from its documentation.

Raw data also has no result. A creation inside a failed transaction did not happen, and a transaction in a skipped slot never happened at all.

Confirming that the account exists#

Write the detection as a provisional row, keyed by the transaction's first signature and the new address, using the status model from provisional rows. Once the block settles, check the address with the getAccountInfo RPC method at the commitment you choose: it returns the owner, lamports and data length, or nothing if the account does not exist. The reconcile guide runs that check in batches.

When instruction level detection is the wrong tool#

If you need every new account on the network, including those created inside other programs, detection from outer instructions will miss many of them. Settled data with inner instructions, or an index built from account updates, fits that job better. Instruction level detection is the early signal for the cases above, not a census.

Keep reading#

Combine this with the token program instruction reference for the token side. To run a watchlist on live traffic, Standard Shreds delivers the raw packets; send us your IPv4 address and UDP port after ordering.

If the accounts you care about belong to new token launches, the comparison of sources for new Pump.fun tokens puts each place a launch appears in order.

Token accounts created under Token-2022 bring their own extensions, which the Token-2022 notes walk through.

Sources#

Raw shreds on your own server

Ultra Low Latency Shreds is $200 and Standard Shreds is $100. You pay in crypto, tell us which IPv4 address and UDP port should receive the stream, and we wire it up by hand.

Read nextObservability for decoders: the metrics that catch silent failure