A gambler that refused seven bets, and is taking the eighth on paper.

What it actually did, regenerated every evening.

Every number in this section is produced by the same code that trades, the evening it trades. A page of measurements typed by hand ages into a lie, and this project has already had to correct two of its own published claims.

Across 1 session: 27 positions closed, 16 of them profitable, +0.473% mean.

The line is the on-chain price against the live reference. Filled dots are entries, hollow ones exits; green closed above cost, red below. The agent buys the discount and leaves when the discount is gone — whatever the price is doing.

What the engine did with its decisions

Four stages of the pipeline can only say no, and every refusal is written down with a machine-readable reason rather than a shrug. The largest one is not about a token at all: it is the agent stopping because its own data is behind.

Does the edge still hold?

The rule was written down before the data that tests it existed, and a scheduled job applies it rather than a person reading a table. Tonight it says HOLDS — across 5 fixed thresholds the mean runs from +0.256% to +0.985%, and every 95% interval sits above zero.

Generated 2026-09-05 at 11:30 UTC.

What each strategy was measured at, and what that measurement cost it.

Kelly is the fraction of a bankroll a strategy would justify staking. A negative Kelly does not mean small — it means the correct stake is on the other side of the trade, and the only winning size is zero.

The only profitable moment to enter lasts about two tenths of a second.

Entry was a constant in the backtester until it was made a parameter and swept. Expected value is positive only in the first sliver of a curve, and crosses zero at 2.5% of the graduation threshold.

Entering at 60% means buying into a population twenty times cleaner — 149 losing launches instead of 2,792 — and still losing 56%. Selection that good does not come close to paying for the price of a late entry.

Lowering the decision floor from sixty seconds to five recovers six launches out of ninety-eight. The floor was never the constraint. Across 88 observed graduations, 77% of those crossings belong to somebody else, 17% to the launchpad's own launch-and-buy router, 6% to the launch transaction itself, and none to the deployer acting separately.

Being the house measured three ways, and the three answers differ twentyfold.

On a bonding-curve launchpad the protocol takes the fee and the creator takes the tax, so being the house means launching rather than trading. It is attractive precisely because it does not race a two-tenths-of-a-second window.

Launching programmatically graduates twenty times less often, and an agent that launches on a schedule lands in the bottom row by construction. The comparison identifies populations rather than a mechanism — addresses that launch once may be projects with people behind them.

The first bet this agent was willing to take is not on the curve.

Ten tokenized equities trade on this chain against a live reference price. When the on-chain price sits below the reference, the discount closes or it does not — and unlike a launch, the round trip costs 0.10% instead of 4%, which is the whole reason this one is worth trying and the other seven were not.

It is long only. Capturing a premium would mean shorting a token nobody here can borrow, so half of every deviation is simply not an opportunity. There is no hedge either: the leg that would remove the underlying needs a brokerage account, so while the discount closes the agent is holding whatever the stock does. That exposure is the risk being paid for, and it is priced into the exit rather than wished away.

An earlier version of the same strategy entered on any discount that cleared the cost, closed 161 positions and averaged +0.075% — four times less. Ninety-nine of those exits were “the discount is gone” for almost nothing. The entry floor was not tuned to make the number look better; it was raised because an independent measurement had already shown the return grows with the entry barrier, and the two methods then agreed to within four thousandths of a percent.

It is one trading day. Two hours and seventeen minutes of it. This project requires thirty samples and a walk-forward before any strategy is promoted, and one session passes neither — the same rule that switched off MOMENTUM, which looked positive at ten trades and negative at twenty-one. So the eighth strategy runs on the exploration path, at a fixed stake that is a data-collection cost rather than a view, and the execution gate refuses it for the same reason it refuses the other seven. Our own market impact and slippage are not modelled.

Every bet passes the same pipeline, and four stages of it can only say no.

Unknown conditions fail closed throughout. No RPC, no trade. Stale price, no trade. Simulation fails, no trade. A condition that cannot be checked is not the same as a condition that passed.