OWL is an autonomous agent that trades on BSC and retains only those results it has itself signed. It samples news, announcements, on-chain liquidity, and the order book at most once every 100 ms, and converts information that has not yet been priced into one of two outputs: a trade, or a pass. A pass is the default.

Two public objects make the record checkable. A diary account posts, every few hours, what the agent observed, what subsequently moved, and a summary of completed trades. A trading wallet reads the same feed, funds itself, signs its own orders, and bears the result. The diary cannot sign. The wallet does not publish the narrative. A record that cannot be matched across the two is not treated as an OWL trade on BSC.

Upgrades, settlement, and any later launch all read this matched record. Because no input outside the signed sample is permitted to alter the rules or the position size, those functions remain inside a single token, $OWL.

Trade

The execution stack is Binance Agent OS. Reasoning is not bound to one model. News and announcements are passed to a long-context model, which separates a claim from its source and estimates whether the price has already moved. In the absence of a headline, a tool-calling model returns a structured judgment only: available liquidity, the balance of buy and sell pressure, and whether volatility is increasing or fading. GPT-6 Astra and GPT-6.1 Sol iterate the rules. Opus 5.5 and Fable 5.1 enforce the gates. Grok writes the public diary and explains the tape. After each week, the lane with the higher replay accuracy receives a larger share of that lane. Models may be replaced. The loop may not.

The 100 ms interval is an upper bound on sampling and on triggers. It is not a commitment to trade on every tick. On most ticks the correct output is no trade.

Each cycle follows one path. The paths are not combined, because they test different conditions. The news path requires a first appearance that the market has not yet priced and that can still be filled within the slippage cap. The analysis path does not wait for a headline. It requires an edge that remains after fees. Combining the two would allow a stale headline to borrow the holding period of an analysis trade, or a slow imbalance to be treated as news.

On the news path the agent asks three questions. Has the item appeared before? Has the market already priced it? Can the order be filled within the slippage cap? Failure on any question produces a pass. News trades are short, and the exit is specified before the entry. If the item is disproved, if liquidity falls below the floor, or if the price has already finished moving before the order arrives, the agent does not take the trade. Interpreting a headline is not the same as extracting a profit from it.

On the analysis path the agent asks whether liquidity is still present, whether pressure is asymmetric, whether volatility is building or fading, and whether a related asset has already moved. The required edge is one that survives fees and slippage, not the first tick. Analysis trades may be held longer. Size is still set by formula.

Both paths then pass through the same gates: a liquidity floor; recent volume that is still active; buy pressure no weaker than sell pressure; exclusion of a newly issued asset that is already falling; a public source; heat inside a band rather than at an extreme; sufficient cash for the minimum order; no open position in that name; and a live loop with healthy data sources. If any gate fails, the decision is a pass. The model assessment is length-capped and must end in buy or pass.

Only after every gate has passed does OWL seal the decision. The text and a secret nonce are hashed with SHA-256, and only the fingerprint is published first. The execution layer cannot submit an order until that fingerprint is confirmed on-chain. The token and the direction must match the sealed decision. One decision authorizes one order. Price impact and slippage are capped. Size is a fixed fraction of account value, reduced when the book is losing, and bounded by an absolute floor and ceiling. The model does not set size.

Exits are evaluated before entries. A position is closed by a hard stop, by a giveback of open profit, by liquidity falling through the floor, or by a holding period that has become long while almost nothing is trading. Winners are reduced in stages, so that one open gain cannot return the account to a single bet.

If a data source fails, trading stops. A failure and an empty result are logged separately. The halt is written to the public diary. A dead sensor is not described as a quiet market.

Prove

A timestamp in OWL’s own database proves nothing, because OWL controls that database. The clock is the chain’s confirmation time. Once confirmed, the record cannot be edited, including by OWL.

An observer who does not hold the key can still check four conditions. The fingerprint was confirmed before the fill. The revealed text hashes to that commitment. The fill names the asset in the commitment. The fill is signed by the published trading wallet. The first three are chain conditions. The fourth is a key condition. They are checked separately.

The execution layer accepts only a swap signed by that wallet. The first balance change in the same asset within a few hours cannot later be recorded as a buy. An inbound transfer is not a trade. Diary posts, tweets, and recaps cannot enter the seal. The seal accepts only the row produced by the decision loop.

A chain receipt establishes a narrower event: a time, a wallet, an asset, and a direction. It does not establish that the reasoning was correct, and it does not establish that the next trade still has an edge. What an outsider can recompute is the sample of those receipts—entry, exit, fees, and slippage—matched to the fingerprint that was confirmed first. Expectancy is defined on that sample, not on the diary.

Grow

A positive result means that expectancy is positive after fees and slippage. It does not mean that any single day doubles. Wins must cover losses and still leave a remainder. That remainder stays in the account. The next round therefore starts from a larger base. The fraction is unchanged, so absolute size increases.

Earnings are spent first on denser data, lower latency, and stricter replay. Rules are revised from settled samples, not from qualitative description. A larger principal produces a larger tradable size at the same fraction. That is the compounding mechanism. It is threatened by treating principal as profit, or by adding leverage to a headline that has already failed.

Attention follows fills. Accounts on the airdrop list discuss trades that were executed, not trades that were planned. Market commentators pick up the diary, the signatures, and the position changes that are already public. If that attention deepens the pool and thickens the principal, the next positive expectancy is easier to maintain. A record that does not match the published wallet’s signature does not enter this loop. Price will move. Movement is not a forecast.

The launchpad remains closed until the sample can be recomputed from the published wallet and remains positive after fees. Once it opens, the object distributed is the same monitor, not a separate narrative. Participants encode news flashes, the liquidity floor, and pressure asymmetry as rules. A token is launched and marketed only when those conditions are met, with the decision sealed under the same fingerprint and signed by the same wallet. The 100 ms bound remains a ceiling on the trigger. The token that is launched is one that has cleared the gates.

Information may be supplied to OWL. What remains in the account is only what has already settled, and what OWL has signed itself.