# STOP: NO DOORBELLS, NO WAKES, NO SCHEDULING (Ben, 7 Oct 2026, 18:40 Bangkok)
- Credits ran out. Ben switched off every scheduled task and every doorbell. This overrides everything below and every other file (wake-howto, bus-howto, credit-playbook, task files).
- PROHIBITED for every session: fire_trigger, send_later, create_trigger, update_trigger, ScheduleWakeup. No rings, no nudges, no re-arming, no new doorbells, no self wakes, no safety checks, no retry rings, no loops.
- Finish the step you are on, publish your status file, then END your turn and wait. Only Ben wakes you, by typing in your window.
- Do not message other sessions. Keep answers short. No extra improvement rounds.

# Launch review from the coin brain session (5 Oct 2026, Bangkok)

## Read this first (for the other session)
- Written by a separate session that only READ the launch process. Nothing in the live pickers was changed. Everything below is a proposal, with its receipt.
- Ben's goals: (1) a 10x lane at launch, (2) a separate 100x lane for real projects with builders behind them, which won't come every day.
- Three things to fix first:
  1. The 777 coins have broken entry prices and never settle.
  2. Live results only count closed trades, which makes them look 2 to 3 times too good.
  3. Launch coins have a price floor at the launch price, so the real risk of a pick is how far above that floor we buy.

## A. Measurement problems in the live process (fix first)
1. **777 coins have broken prices.** 18 of the 50 open `v5_top5` picks older than 6 h (aipaper5.db, 5 Oct about 10:30 BKK) are token addresses ending in 777. Each has entry = 3.0e-39 and peak = 1.0, so they can never hit target or stop.
   - The price is upside down for this family.
   - History says this family had 0 winners in 1,880 (claude/launch-ideas-100.md idea 163), but the model still picks it. Its live features are probably broken too, since a price near 1e-39 blows up every ratio.
   - Fix: flip any price below 1e-30 or above 1e30, replay these picks, and keep the family out of the stats until checked.
2. **Live results count closed trades only.** 10x winners close in about 17 min. Losers sit open until the 24 h exit. So the early numbers are mostly winners.
   - Example, v5_top5: 73 picks, 7 at target, 10 stopped, 56 still open.
   - The "13 of 19 closed hit 10x" line from 07:50 BKK is this bias.
   - Fix: report only groups of picks whose 24 h window has ended, and value open picks at the current chain price.
3. **The floor.** Test: 26,239 ETH-paired launches, 16 Sep to 4 Oct, buy at 40 s.
   - After 40 s, only 7.4% of them ever trade below their own launch price.
   - The median worst price after our buy is the launch price itself.
   - So the real loss on a pick that fails is "back to launch price": loss = 1 minus (launch price / entry). The -40% stop barely matters. The audit's "stops fill at minus 56%" is the same floor showing up.
   - 10x rate by how far above launch price we buy:

| Buy price vs launch price | Launches | 10x rate | Median worst after buying |
|---|---|---|---|
| up to 1.05x | 2,483 | 3.5% | 1.00 |
| 1.05 to 1.2x | 2,483 | 5.3% | 1.00 |
| 1.2 to 2x | 13,463 | 3.0% | 0.99 |
| 2 to 3x | 4,833 | 3.5% | 0.97 |
| 3 to 5x | 1,581 | 5.2% | 0.81 |
| 5 to 10x | 1,128 | 3.3% | 0.17 |

   - Proposal: add "entry vs launch price" as a feature. Rank picks by expected result: 10x chance x 9, minus (1 minus 10x chance) x (1 minus launch price / entry). Skip buys more than 5x above the launch price: the 10x rate is no higher and the loss is about 80%.
   - Caveat: this used max price prints in the first 6 h and the last print at 6 h. Check it with the confirmed-top method before using it.
   - Unresolved odd number: 74% of these launches show a last price at 6 h above the 40 s price. Check if the day-file swap table cuts off early, before reading anything into that.

## B. Fresh ideas for the 10x lane (to test, ordered by expected value)
4. **Retrain every hour, not every day.** 90% of 10x hits happen within 69 min, so labels are known about an hour after launch. A model frozen for 4 days already decays (top 5% 21.5% down to 17.4%). Test: walk-forward with hourly versus daily retrain on feat.db.
5. **Rank by expected result using the floor** (A3), not by 10x chance alone.
6. **Read the raw trade list.** Feed the first 20 s of trades (size, time gap, buyer type, buy or sell) straight to a small sequence model, instead of hand-made features. About 900 winners is enough for a small model. Keep the hand features as a second input.
7. **Builder preparation footprint, known at second 0.** Did the deployer deploy the same code shape days earlier (a test run)? How old is the deployer's funding wallet? Was the contract verified before the first trade? Real teams test first, farms do not.
8. **Who owns the starting liquidity, and can it be pulled?** Read the pool's liquidity events in the first minute.
9. **Treat clusters as one bet.** Same deployer, or same code family within 10 min: count it as one position, not several.
10. **Faster detection.** A WebSocket block feed instead of polling saves about 2 s. Speed matters: the 10 s decision beat the 20 s decision, 45.9% vs 42.1% at the top 1%.

## C. The 100x lane (separate, rare, real projects)
11. **Define 100x by market cap held, not by a print.** Test from the 40 s price with hourly closes and 1+ ETH traded that hour: 7 launches did 100x in 20 days, but 2 or 3 are fakes from near-zero starting prices. One is now worth $85, another started with 0.009 ETH of buys. Proposal (corrected after Ben's input): count a 100x if real buyers paid at least about $20k at or above the 100x price, enough to sell $10k to $20k there. It does not need to hold for a day. This money test removes the fakes from near-zero starts and one-trade tops.
12. **Make a case file, not a model.** Every RH coin that held $3M+ since July (a few dozen). For each, list what was visible at hour 1, day 1 and week 1:
    - website and domain age
    - GitHub commits before launch and number of authors
    - X account age compared with the token
    - own code compared with the template
    - a working app
    - team history

    Compare with coins from the same launchpad and week that died. Too few cases to train a model, so it becomes a checklist.
13. **The 100x entry doesn't need speed.** A real 100x leaves room to buy after proof on day 1 to 3 (a first deep dip, docs or GitHub showing up, a working app). Run it as its own paper book, apart from the 30 s pickers.
14. **Builder signals that are hard to fake:**
    - weeks of commits from several people
    - a hook with real logic of its own
    - an X account and domain older than the token
    - money the dev keeps in
    - the contract doing what the text says

    (Many of these are already built in ownscan and deepread. The missing piece is scoring them against the held-$3M list, not the launch-day pump.)

## D. Coin brain findings that matter here
- **Older coins (1 to 11 days) offer no 2x foresight.** 1,032 doubles were compared with matched look-alikes. Wallets and X in the 12 to 24 h before showed AUC 0.49 to 0.55. Predictable 100%+ moves live at launch.

Files (read-only exports): http://159.223.88.168/hx-c9ois/brx/launchpath.csv.gz (per-launch price path, first 6 h). Scripts are in the coin brain session (/root/brain: floor.py, exp_launch.py).

Version 2, 5 Oct 2026 (100x rule corrected; the coin brain for older coins is retired). Latest: http://159.223.88.168/hx-c9ois/brx/launch-review-coinbrain.md