What we measured

Tokens graduating 04–08 UTC doubled 63% of the time. At 16–20 UTC: 53%. The ordering holds in both halves of the sample.

Median move 60 seconds after a call: +0.5%. At two minutes: +0.0%. A coin flip — the audience does not arrive.

Win rate climbs with position size — but 75% of every position above 5 SOL comes from five wallets running industrial volume. Not something you can copy by sizing up.

Leader return and copier return are different numbers. Of the 89 wallets with at least ten graded round trips, 4 pay a copier anything at a three-second delay — 4 percent. Raise the bar to a hundred trips and only four wallets qualify at all, which is why we publish the whole distribution rather than a shortlist. Of the four, one has a confidence interval that excludes zero.

We had Photon at a flat 0.0035 SOL for three weeks and it was wrong. Four controlled trades read on chain put it at 1.00 percent per leg, and the accounts we had been calling Photon receive nothing on a real Photon trade. The correction is on the page.

Every tracked call priced every ten seconds for ten minutes, per channel, with the cost of trading it at your size.

Leader return and copier return are different numbers. Across every wallet we graded at usable sample sizes, copying at a realistic delay did not pay. Changing the delay barely moves it.

We split the calls in half by date and picked the top channels on the first half. On the second half they returned less than taking every call. Past channel returns do not carry.

The channels with the highest graduation rates get them almost entirely from calls that graduate within two minutes. Remove those and no channel beats what the token’s market cap at call time already predicted. It is detection speed, not foresight.

The best rule we could find on the first half of the data returned -3.28 percent on the second. Simple time exits survived the same test; tuned targets did not.

No edge that survived out of sample.

No edge. Nor did sizing ladders, hold-time rules, or market-cap bands.

Being faster did not pay at the resolution we can measure.

A cross-source artifact. Our own graduation prices and DexScreener’s differ by roughly half at the graduation instant, and once both sides are on the same basis the expected value disappears.

There is a spike. It belongs to snipers, not to anyone reading a Telegram message.

Graduation rates are 0.2 to 1 percent, distribution is the thing we do not have, and we would be detectable with our own clustering tools.

On corrected numbers no bot breaks even at any position size on a 30-second exit. The gross return is smaller than the fee.