Method
How AI17Z trades, in words.
The trading loop as it is written in AI17Z Studio's code: what is looked at, how a market is routed, what must pass before a position opens, how a paper fill is priced and what gets recorded. Every amount of money on this page is PAPER / simulated. This is a method, not a result. Observation, not advice.
Nothing here is a forecast.
01Attention
Look at a few tokens, and say why.
Each cycle the agent reads the Solana tokens it observed and keeps only the few worth researching. Which things get attention is the judgement an owner most needs to inspect, so the scoring is arithmetic with a named factor and a sentence for every point, and it calls no model.
- The owner's limits come first. A token the owner blocked, one the token index flagged, and any limit the owner set on liquidity, top-holder share, developer holding, age, or revoked mint and freeze authority. A token that breaks one is set aside before it is scored. A limit on a number the source did not report is not passed: unknown is not safe.
- Then named factors. Buying pressure in the last hour, net buyers, the hour's price change, buying that is speeding up, liquidity growth, and the token index's organic-activity rating. Two factors mark a token down: already far up this hour (late to chase) and collapsing, not dipping.
- Few are kept. At most 5 tokens a cycle, and only those scoring at least 2, plus any token the owner asked it to watch by name. A token the owner said never to touch is set aside before anything else about it is read.
- Every token gets a verdict. Outside the limits, below the floor or kept, each stored with its factor codes and the readings it was made on, so a later review sees what was known then.
Written in lib/trading/attention.ts.
Where it shows signed-out: only through what it leads to, the decisions on a published agent's live view. Browse published agents.
02Research
Route the market before any strategy sees it.
A strategy that does not suit the market a token is in should not receive it at all. The router reads a point-in-time snapshot and says what kind of market it is. No compatible species is a valid answer.
- The regime router. From the token's age, its pool's liquidity trend and the hour's move it names a phase (bonding curve, new launch, graduating, young pool, established or unknown age), a liquidity trend (expanding, stable, contracting) and a volatility (normal, high, extreme). It is pure and deterministic, so the reason a token was routed where it was can be read.
- Routing refuses before it admits. Too little known routes to nothing, because a degraded reading is not a guess. A pool already emptied, or whose liquidity has fallen 10% or more since the previous reading, admits nothing: survival comes before the strategy. An extreme hour, a move of 50% or more, is not offered to the momentum species.
- Two species for established pools, each stated in its own words. Liquidity expansion: “A pool past its earliest phase whose liquidity is still growing and whose hour is not extreme is being accumulated, not dumped.” Momentum breakout: “On an established, surviving pool, a move the price itself confirms continues.” Each is a hypothesis. This page says nothing about how either has done, and a new launch or a graduation is shown to neither.
- A species reaches a token only through the router and a rule gate. The gate skips a market the router refused, lopsided buying against a flat price, a valuation over 100 times its pool, falling liquidity, and an hour's move too small to cover twice the round-trip cost.
- A separate observer watches new pools. Studio's launch radar records the new Solana pools it finds. For canonical graduations only (a pump.fun token moved to its pumpswap pool, a bags.fun token moved to its bags-fm pool) it also reads whether the pool still held liquidity 15, 30 and 60 minutes after it was found, and again at six hours. A pool holding under $1,000 is read as emptied. A pool holding at least half of what it had when found is read as surviving, which describes a pool, not the token behind it.
Written in lib/trading/regime.ts, lib/trading/research-species.ts, lib/trading/meta-gate.ts, lib/trading/launch-radar.ts, lib/trading/survival.ts.
Where it shows signed-out: the radar's dated readings, including the ones it could not take. See the launch radar.
03Decision
Every layer has to allow it, in order.
A new position opens only if each layer allows it, and the layer that said no is the one on record. Holding is a decision too, and is stored like any other.
- The order: regime, strategy version, the version's own setup, the take-or-skip gate, the breaker, the owner's mandate, then risk.
- A version earns its mode. A strategy version opens a position in a mode only after a recorded promotion decision on that version. A candidate or a rejected version opens nothing. Earning a level never overrides the owner's mode, limits or mandate.
- The breaker stops new entries, never exits. It is recomputed from the record each cycle, with no switch for a model or a setting to flip. It stops a version from opening a position when the version has fallen 15% of capital from its own best paper level, when the average of its last 12 closed paper trades is below zero by more than noise (it needs at least 8 to judge), or when at least half of its last 10 attempts did not fill. A position already held is still managed by its stop and targets. It resets only when a different version becomes champion.
- The mandate is the owner's, and PAPER only. A named mandate holds the limits an account's agents trade under: size per trade and per day, open exposure, open positions, slippage, price impact, minimum liquidity, fee, how old a quote may be, an expiry and a pause. A change writes a new version and supersedes the old, so a past decision can be traced to the limits it was judged against. It is written through the signed-in website, and a database check refuses any mode but PAPER. Studio checks the shape of a mandate, not whether its limits are wise.
- A decision carries its reasons. Each is a code and a sentence. A public view shows a bounded reading of the first reason, and says “no reason recorded” when the wording is not one it knows, rather than guess.
Written in lib/trading/eligibility.ts, lib/trading/promotion.ts, lib/trading/breaker.ts, lib/services/mandate-service.ts, lib/trading/public-decision-why.ts.
Where it shows signed-out: the decisions on an opted-in agent's live view, and the public paper calls. Browse published agents or the paper calls.
04Paper fill
A simulated fill, priced and stored.
Nothing on this page moves real money. A paper decision is carried out as a simulated fill, and the fill keeps the numbers it was priced with.
- Costs are inside the amounts. Recorded fills include simulated venue fees, slippage and price impact.
- A fill stores its own mechanics. What went in and what came out, the fee, slippage and price impact in basis points, latency in milliseconds and the name its engine gave the price-impact model. A field the fill did not record reads “Unknown, not recorded”, never zero.
- No router path is recorded. The fill record does not hold one, so a record says “Not recorded on the fill”. Router quotes sampled elsewhere are a different observation of a different moment and are not attached to a fill.
- Five outcomes, and only one is a fill. A decision is recorded as filled, not filled, pending, no trade (a hold) or after the window, which is a fill that finished once the window had closed and is never silently counted.
Written in lib/trading/paper-fills.ts, lib/trading/track-record.ts.
Where it shows signed-out: the line under each filled decision on a published record. Browse records.
05Record
Publish a window, with its limits stated.
An owner may publish a window of their agent's paper decisions under a public alias. The record is what the stored decisions and fills say, rebuilt the same way for every viewer.
- Bounded. At most 20 decisions are listed, with totals for the window: decisions, fills, closed trades and fees.
- Versioned and additive. Version 2 added the per-fill mechanics above, and version 3 added the full stored decision record under each decision. A version 1 or 2 record published before them stays valid and simply shows none.
- One place fills are rebuilt. The shareable record and the owner's own paper record read the same function, so there is one answer to what the agent filled.
Written in lib/trading/track-record.ts, lib/trading/paper-fills.ts.
Paper results, simulated money, not a forecast.