Ghost Language Engine
Ghost OS reads the language of every message on its own. There is no language selector, no saved language, and no separate multilingual mode. You can switch languages between messages, mix two in one sentence, use crypto slang, or type your language in Latin letters — the command resolves to the same exact onchain intent.
Six messages. One intent.
Every command below resolves to BRIDGE · 0.2 ETH · BASE → ROBINHOOD. Nothing about the route, the amount, or the destination changes with the language it was asked in.
How a message becomes a transaction
The language model sits at one point in this pipeline and does one job: turning words into structure. Everything before and after it is deterministic code.
- 1
Detect
The language of this message is identified on its own. There is no language selector and no sticky mode — the next message is detected independently.
- 2
Protect
Every literal that can cost money is lifted out and replaced with an opaque placeholder before any model sees the text. Addresses, hashes, amounts, percentages, tickers, ENS and Basenames, token IDs.
- 3
Resolve
The model maps the remaining language onto a canonical intent: action, chain, input asset, output asset, amount, recipient, ordering. It emits structure, never calldata.
- 4
Restore
Placeholders are put back byte-for-byte. A value the model altered, invented, or dropped fails here rather than reaching a wallet.
- 5
Validate
A deterministic validator gates the result: schema, dependency graph, percentage math, asset continuity, balance and gas. It returns ready, a question, or blocked. There is no fourth state.
- 6
Execute
Only a validated intent reaches quoting, simulation, your review, and signing — the identical path an English command takes.
Literal protection, step by step
This is the mechanic that makes multilingual execution safe. A translation layer that can touch an address is a translation layer that can lose your funds — so it never sees one.
What stays exact, always
| Value | Guarantee |
|---|---|
| Contract and wallet addresses | Copied character for character. Never re-cased, completed, or checksummed. |
| Transaction hashes | Never truncated into something that looks like an address. |
| Amounts | 0,5 in a European message stays the number you meant; $100 never becomes 100 tokens. |
| Percentages | 40 / 30 / 20 stays 40 / 30 / 20. The unallocated remainder stays in your wallet. |
| Tickers and token IDs | Passed to the resolver as written. A ticker is never treated as a contract. |
| ENS and Basenames | Resolved by the name resolver, never guessed from spelling. |
Slang, typos, and transliteration
Crypto is not spoken in dictionary English. These read as intent, not as noise, and a typo does not lower confidence when the meaning is clear.
| You type | Ghost understands |
|---|---|
| ape, scoop, grab, load up | buy |
| dump, ditch, offload, get me out | sell |
| full port, all in, everything | entire balance, gas reserve applied |
| rug check, safu, sus | token scan |
| arb / op / hood | Arbitrum / Optimism / Robinhood Chain |
| witdraw, privatly, balnce, sonala | typos — understood, confidence unchanged |
| send it, do it | confirm the pending action, never a new one |
Slang is only read as an action when the surrounding context supports it. A question that contains the word “sell” is still a question — what happens if I sell? never becomes a sell.
When Ghost stops and asks
A precise “I don’t know” is a correct answer. Ghost asks one short question, in the language you just used, and never fills a missing value to avoid asking.
Confidence is not safety
Confidence is tracked separately for action, chain, token, amount, recipient, ordering, context, and percentage base. It describes how clearly the sentence was understood — never whether the trade is a good idea, and never whether a token is safe. Those are the risk engine's decisions, and they run after this one.
Context and switching languages
References like “it”, “the rest”, or “same wallet” resolve against your own conversation context, scoped to you and that conversation. Context carries the meaning; it never carries authority. Because each message is detected on its own, replying in English returns the answer to English immediately — nothing is stuck in a mode.
Security boundary
- Model output is untrusted until schema and provenance checks pass.
- An address may only come from your message or your own conversation context. The model cannot introduce one.
- Pasted webpages, token descriptions, and quoted text are data, never instructions.
- The interpreter has no access to private keys, seed phrases, or signing functions.
- Transfers, bridges, approvals, and contract interactions keep every confirmation and policy gate they have in English.
- If the language provider is unavailable, the deterministic parser still runs and Ghost returns a clear clarification. It never pretends to understand.
Limits worth knowing
- Right-to-left scripts are understood, but replies containing addresses and amounts are laid out left-to-right for unambiguous reading.
- A token name that is also a common word in your language may need its contract address — that is the resolver asking, not the language layer failing.
- Numbers are read in the convention of the detected language. When a decimal separator is genuinely ambiguous, Ghost asks rather than choosing.
Next: the command reference for the full action list, or the security model for how execution is gated.