Promptvault

Vault open · Solana mainnet · slot —

The model can be public. The prompt should not be.

Promptvault measures how much of your prompt an inference node could recover, then gives you the locks: redaction, AES-GCM sealing, secret shares that read as noise, and a hash you can commit on Solana mainnet without saying what you wrote. Every number is computed in your browser.

combination · reading slot…

Measured leak, raw hop

100%

Measured leak, one share

0 bits

Cipher

AES-256-GCM

Server sees your prompt

never

01 · Leak meter

How much of this prompt could a node recover?

Type or paste a prompt. The meter finds what is sensitive in it, then actually runs the attacks: it reads the raw text, breaks a substitution-cipher "obfuscation" with letter frequencies, and tries to guess from a single additive secret share. The percentages are measured on your text, not quoted.

    02 · Redact and seal

    Send the shape. Keep the secrets.

    Redaction swaps every finding for a placeholder, so a hosted model can still do the work. The map back stays in this tab. Sealing encrypts the original with AES-256-GCM under a key stretched from your passphrase with 310,000 rounds of PBKDF2. Paste a model's answer and the placeholders turn back into your values, locally.

    What you send (redacted)

    
                
              

    Rehydrate a reply

    
              

    Seal the original

    The sealed envelope appears here: salt, IV and ciphertext, all base64.

    03 · Secret shares

    Split it so no single holder learns anything.

    Shamir's scheme over GF(256): every byte of your prompt becomes a point on a random polynomial. Any k shares rebuild it exactly; fewer than k carry zero information. The histograms show each share is flat noise, and the chi-square test says so in numbers.

    Rebuild from the shares you tick

    Tick shares above to rebuild.

    04 · Through the swarm

    What each node holds, hop by hop.

    Raw: every node handles real values derived from your prompt. Scroll to move the request along.

    05 · Compare the locks

    Six ways to run a model, weighed your way.

    These are kinds of inference, not named companies. The attributes are our assessment of how each design works, labelled as a model. Move the weights to see which design fits what you care about, and what a private token really costs in bandwidth.

    Bandwidth of a private token (model)

    06 · Your deposit box

    A box only your wallet can open.

    Your wallet signs a fixed message; the SHA-256 of that signature becomes an AES key that never leaves this tab. Prompts you file are stored encrypted in this browser. To commit to a prompt without showing it, hash it with a salt and write the hash on Solana mainnet as a memo. Later, reveal prompt and salt and anyone can check the hash.

    LOCKED

    Connect a wallet at the top right, then unlock.

    File the current prompt

      Commit on mainnet

      Salt—
      SHA-256(prompt ‖ salt)—
      Memo—
      Your SOL—

      Check a commitment

      Review · Solana mainnet

      Write a commitment memo

      Permanent and public. The memo is written to mainnet forever. It contains only the hash, not your prompt. Nothing is sent to anyone; you pay the network fee only.

      07 · How it works

      Four locks, all in your browser.

      L1

      Find

      Pattern and checksum detectors: emails, phones, card numbers that pass Luhn, base58 strings that decode to Solana keys, API-key shapes, URLs and IPs.

      L2

      Seal

      Web Crypto PBKDF2-SHA-256 stretches your passphrase; AES-256-GCM encrypts and authenticates. A wrong passphrase fails loudly, never silently.

      L3

      Split

      Shamir sharing in GF(256) with random coefficients from crypto.getRandomValues. Below threshold, every message is equally likely.

      L4

      Commit

      SHA-256 of prompt and salt, written with the SPL Memo program from your wallet, confirmed with getSignatureStatuses and readable by anyone.

      08 · Questions

      Plain answers.

      Does Promptvault run a model?

      No. It is a toolkit for the prompt side: measuring, redacting, sealing, splitting and committing. The swarm and bandwidth figures are labelled models of how private inference works.

      Is my prompt sent anywhere?

      No. Every tool runs in this tab. The only network calls are mainnet reads through our read-only RPC proxy and, if you choose, a memo transaction your wallet signs. The memo holds a hash, never the text.

      Why does one share recover "something"?

      An attacker who sees nothing can still guess the most common character and be right sometimes. The meter shows that blind baseline next to the share result: when they match, the share taught the attacker nothing.

      Where does the deposit box live?

      In this browser's local storage, encrypted with a key derived from your wallet's signature. Clear site data and it is gone; we hold no copy. Wallets that sign deterministically (Ed25519 does) reopen the same box.

      What does a commitment prove?

      That you knew this exact prompt at the time the memo landed. Revealing the prompt and salt later lets anyone recompute the hash and compare it with the memo on mainnet.

      Is there a token?

      Only if one is shown in the hero and the footer. Until then there is nothing to buy, and no feature here needs one.