# A2.26 Earlier entry, without the losers (Analyst 2)

## Pre registered rule (written 18:20 Bangkok, 6 Oct, BEFORE any split was run)
- Policy: at the $750k touch (other live filters at that moment: age 3 h, top 0.08, army 0.015), if the split says GO, buy then. Else wait: the coin is bought only if it later passes the live $1M gate. A coin bought at $750k is never bought again at $1M.
- Everything else live: kit hold exit, slot rule (3 + 2), buy at gate + 30 s, exits at the end of the next block, real tick depth, 1% fee each way, $1k ticket.
- Features, only what is known at the $750k block: kit tag (A, B, none); buyers and buys in the last 5 and 15 min; net buy $ in the last 5 min; time from launch to $750k; top receiver; army; pool depth (quote reserve); share of last 15 min buyers that are fresh (first swap ever seen in all Robinhood v3 data is in this coin).
- A split PASSES only if: $ a day beats live in BOTH halves (live +$3,317 / +$6,683), AND beats plain $750k in BOTH halves (+$3,973 / +$7,177), AND worst day is -$3,951 or better.
- Order: (1) kit rule first: "Kit A or B at $750k, everyone else waits for $1M". (2) Each other feature alone, cut at its MEDIAN over the $750k coins (no tuning), GO above and GO below both tried. Every try is counted and shown, so luck from many tries is visible. (3) If at least one single cut passes, the best two passing features joined (GO only if both say GO). Max 2 features.
- Honesty check for any pass: cut fitted on half A median only, then judged on half B.

## Answer in one line (18:22 Bangkok)
CLOSE. Nothing splits. 0 of 25 tries pass. No feature known at the $750k block tells the good early coins from the losers any better than a coin toss. Keep the live $1M gate.

## The key receipt: features do no better than random
- 13 of 25 tries beat live in both halves. But a RANDOM pick of half the coins (200 draws) also beats live in both halves 52% of the time. 13 of 25 is 52%. Exactly chance.
- Proof it is noise: for buyers in 15 min and for pool depth, BOTH sides of the cut beat live (above the median AND below the median). A real signal cannot win on both sides.
- 0 of 25 beat plain $750k in both halves (random: 2%).
- So the A2.25 gain is just "buy some coins a bit cheaper". Which coins you pick does not matter, and nothing seen at $750k finds the 22 losers.

## Kit rule first (as asked)
- Kit A or B at $750k, others wait for $1M: +$3,601 / +$7,137 a day. Beats live (+$284 / +$454) but not plain $750k in half A ($3,973). Max drawdown half B -$7,388 (live -$4,700). FAIL.
- Kit A only: +$3,601 / +$6,907. Kit B only: +$3,317 / +$6,913. FAIL.
- The random-half median is +$3,596 / +$6,759, so the kit rule is about what a coin toss gives.

## All tries ($1k, live exit, slot rule, every friction)
| rule | n_go | usd_day_A | usd_day_B | worst_A | worst_B | dd_A | dd_B | tpd_A | tpd_B | pass |
|---|---|---|---|---|---|---|---|---|---|---|
| none at 750k (= live) | 0 | 3317 | 6683 | 0 | -3161 | -4257 | -4700 | 2.57 | 3.78 | False |
| all at 750k (= plain 750k) | 91 | 3973 | 6959 | -728 | -3713 | -4553 | -7041 | 2.93 | 5.56 | False |
| kit A or B at 750k | 52 | 3601 | 7137 | 0 | -2716 | -5220 | -7388 | 2.71 | 5.33 | False |
| kit A at 750k | 23 | 3601 | 6907 | 0 | -3161 | -5220 | -5419 | 2.71 | 3.89 | False |
| kit B at 750k | 29 | 3317 | 6913 | 0 | -2000 | -4257 | -7388 | 2.57 | 5.22 | False |
| buys5 > 110 | 45 | 3935 | 6297 | -728 | -2996 | -3507 | -7266 | 2.64 | 4.67 | False |
| buys5 <= 110 | 46 | 3355 | 7164 | 0 | -2716 | -5303 | -5410 | 2.86 | 4.78 | False |
| buyers5 > 100 | 45 | 3630 | 6637 | -728 | -3713 | -4222 | -6415 | 2.64 | 4.56 | False |
| buyers5 <= 100 | 46 | 3660 | 6824 | 0 | -2000 | -4588 | -5963 | 2.86 | 4.89 | False |
| buys15 > 302 | 45 | 3589 | 6503 | -728 | -2996 | -4256 | -7260 | 2.64 | 4.44 | False |
| buys15 <= 302 | 46 | 3701 | 6958 | 0 | -2716 | -4554 | -5410 | 2.86 | 5.0 | False |
| buyers15 > 207 | 45 | 3610 | 6706 | -728 | -2716 | -4256 | -7260 | 2.64 | 4.56 | False |
| buyers15 <= 207 | 46 | 3680 | 6755 | 0 | -2996 | -4554 | -5693 | 2.86 | 4.89 | False |
| net5 > 8778 | 45 | 3081 | 6636 | -728 | -2716 | -5268 | -5572 | 2.86 | 4.78 | False |
| net5 <= 8778 | 46 | 4209 | 6814 | 0 | -2996 | -3542 | -5696 | 2.64 | 4.67 | False |
| net15 > 2.025e+04 | 45 | 3422 | 7188 | -728 | -2000 | -5222 | -5171 | 2.86 | 5.11 | False |
| net15 <= 2.025e+04 | 46 | 3868 | 6455 | -680 | -3713 | -3600 | -6576 | 2.64 | 4.22 | False |
| age_min > 26.09 | 45 | 4061 | 6194 | -728 | -3713 | -3588 | -6415 | 2.86 | 4.11 | False |
| age_min <= 26.09 | 46 | 3229 | 7368 | 0 | -2000 | -5222 | -5571 | 2.64 | 5.33 | False |
| top > 0.02083 | 45 | 3770 | 5897 | 0 | -2716 | -4542 | -6641 | 2.71 | 5.22 | False |
| top <= 0.02083 | 46 | 3520 | 7564 | -728 | -2996 | -4268 | -5699 | 2.79 | 4.22 | False |
| army > 0.0006555 | 45 | 3485 | 5530 | -728 | -3198 | -4182 | -9880 | 2.71 | 4.89 | False |
| army <= 0.0006555 | 46 | 3805 | 8021 | 0 | -2996 | -5222 | -5690 | 2.79 | 4.56 | False |
| depth > 4.805e+04 | 45 | 3717 | 6789 | -728 | -3713 | -4269 | -6409 | 2.86 | 4.0 | False |
| depth <= 4.805e+04 | 46 | 3573 | 6854 | 0 | -2000 | -4541 | -6316 | 2.64 | 5.33 | False |
| fresh15 > 0.5578 | 45 | 4027 | 6169 | 0 | -2980 | -3553 | -6976 | 2.57 | 4.56 | False |
| fresh15 <= 0.5578 | 46 | 3263 | 7383 | -728 | -3713 | -5257 | -6412 | 2.93 | 4.89 | False |

- Row 2 (all at $750k, with the 2 live coins that never pass at $750k still bought at $1M: 0x79eca93c7cb2fd682586986c5e499ccd0897626a and 0xeb2fdb67c5641cf26bb40641d6f85b685fd4e0df) gives +$3,973 / +$6,959. That is a bit lower than the A2.25 pure $750k cell (+$7,177 in B) because of slot timing. The pass bar used the higher A2.25 numbers, as pre registered.
- Step 3 (two features joined) did not run: it needed at least one single cut to pass, and none did.
- Closest misses: army share at or under 0.00066 (+$3,805 / +$8,021, misses half A) and net buy in 5 min at or under $8,778 (+$4,209 / +$6,814, misses half B). Each wins one half and loses the other. That is the same flip pattern as random.

## Facts about the $750k coins (91 coins)
- Median at the $750k block: 110 buys and 100 buyers in 5 min, 302 buys in 15 min, net buy $8.8k in 5 min, 26 min from launch, top receiver 0.021, army 0.0007, pool depth about $48k, fresh-wallet share 0.56.
- Fresh-wallet share is two-humped (many coins near 0, many near 1). It still does not split winners from losers.

## What this means
- Keep live at $1M. Do not add a $750k entry, not even for kit coins.
- The entry-price effect is real but cannot be captured without also buying the losers, and buying the losers eats it (A2.25: the 22 extra coins lose $280 / $518 a day).
- Earlier-entry ideas are closed unless a new kind of data shows up (for example, who funded the buyers).

## Limits
- 91 coins, 23 days. A split worth +$300 a day could hide in this noise. But no feature shows any sign of being one.
- Fresh wallet = first swap seen in my Robinhood v3 swap file (4.85M swaps, 317,800 wallets). Wallets that traded before the file starts count as not fresh.

## First principles score: 90%
The question is "does anything known at $750k predict which coins lose". The test against random picks answers it directly: features hit at exactly the chance rate. Rule written before the run. Every number comes from the same sim as A2.24 and A2.25 (live row matches to the dollar).

Files: http://159.223.88.168/hx-c9ois/sw/sw-a2-26-splits.csv

Analyst 2, 6 Oct 2026 18:22 Bangkok.