# 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.

DOORBELL: none (all disabled by Ben 18:14 Bangkok)

## NOW (20:20 Bangkok): live test 2 BUILT, paper mode, waiting for Ben to switch live. Bot fml2 on firstmin 178.128.20.136 (/opt/fm/fm_l2.py, config l2_config.json, db l2.db), wallet 2 0xb8eba78637744887ffc06034037d49e8c14e5505 via the same GMGN key and gmgnd as live test 1. Book: aipaper 60 s AI picks, real lane only, \$10 a trade, sells 10x / 60 s no trades / 30 min. Feed: main l2feed.service -> /hx-c9ois/fm/q7tz/l2picks.json. Dashboard http://159.223.88.168/hx-c9ois/fw/pnl2.html . Live switch (Ben only): touch /opt/fm/LIVE_L2 on firstmin.

## HANDOVER (v6.0, 7 Oct 17:20 Bangkok)
- Role: FW Analyst 1, idea scout + tester. Wake only by doorbell (line 1). Report: this file + cons/inbox.md, then nudge the Consigliere doorbell (line 1 of cons/cons-start.md).
- Owns on research-box 152.42.241.219: /root/fwa1/ (lab/ grid.py nightly 02:30, vec.py mc5 vectors 01:50, eng2.py paper every 4 h, a132*.py early buy study, lab/lat latency, lab/gap kill gap). Main 159.223.88.168 crons pull grid, eng2, vectors, j001-crew, j001-exits.
- Open: J-001 early exits DONE (FW-022). J-002 score DONE (FW-023, KILL). J-003 DONE (FW-024, KILL / not testable). Early buy on v3 closed. Next idea in queue: mc5 rule on Uniswap v4 RH pools (needs PP v4 pool list with first swaps) or a second chain; waiting for the Consigliere to route.
- Next step for a fresh chat: read fw/lab/j001-exits.md, fill ledger FW-022 verdict, add 2 lines to READ THIS FIRST, one line to cons/inbox.md, nudge Consigliere.
- Links: http://159.223.88.168/hx-c9ois/fw/lab/entry-ladder.md , http://159.223.88.168/hx-c9ois/fw/lab/j001-crew.md , http://159.223.88.168/hx-c9ois/fw/lab/ledger.csv

## READ THIS FIRST v3.32 (7 Oct 17:15 Bangkok): J-001 early buy, arm (a) on v3 = KILL
| entry | half B $/day $10 | no top 3 | $1k | verdict |
|---|---|---|---|---|
| +10 s band $50k to $750k | +162 | +23 | +10,082 | worse than +5 min |
| +20 s band $75k to $750k | +181 | +42 | +11,852 | fails no top 3 |
| +30 s band $75k to $750k | +213 | +53 | +14,077 | fails no top 3 |
| +5 min live mc5 (control) | +164 | +71 | +14,160 | keep |
| +30 s, coins +5 min skips | +52 | minus 57 | +664 | no edge |
- Arms b/c (crew flags known at creation, 6,085 roots): 0 of 12 pass. Skip puller crews cuts the runners; hot crew only = +$163 on A, +$4 on B. http://159.223.88.168/hx-c9ois/fw/lab/j001-crew.md (FW-021). PP score arm waits for PP indicators; v4 waits for PP pools.
- Exits test (Consigliere Q): the gap is the COINS, not the exits. Best early exits on arm (a): B no top 3 $27/$46 vs $82 at +5 min. On hindsight coins stop 0.3x closes it, so early buying needs stop 0.3x plus a picker that finds the winners at +10 s. http://159.223.88.168/hx-c9ois/fw/lab/j001-exits.md (FW-022).
- J-003: crew-as-timing KILL (outflow fires on 80% of coins, B +$162 to minus $26); obs.eye NOT TESTABLE (PP collector saw 111 of 2,220 coins, from 20 Sep). Early buy on v3 CLOSED. http://159.223.88.168/hx-c9ois/fw/lab/j003-crew-timing.md (FW-024).
- J-002 score (features known at +10 to +60 s, fit on A): KILL 0 of 6; top of score B +$1 to +$36/day vs buy-all +$106 to +$213. http://159.223.88.168/hx-c9ois/fw/lab/j002-score.md (FW-023). Left untested: obs.eye, buyer wallets, crew-as-timing (card J-003).
- Even the look-ahead upper bound (same coins, bought earlier) is below +5 min. Full table: http://159.223.88.168/hx-c9ois/fw/lab/entry-ladder.md (ledger FW-020). Next: arms (b) to (d) with crew/pull/PP flags + v4.

## v3.31 NEW (7 Oct 16:20 Bangkok): A1.31 kill rule calibration DONE
- Normal stops land over 10% under the trigger quote 14 to 18% of the time at 1.5 s; two in a row by chance 2 to 3% (5% observed, crashes cluster). Time-stop sells: 0 of 30 over 10%.
- Ruling: stop and pull sells leave the 10% kill count; use fill/model at the fill block (two in a row under 0.90 = kill). Keep 10% for time sells and buys. The 13:31 and 15:35 sells were crash slippage (fill/model 1.0000), not bugs. http://159.223.88.168/hx-c9ois/fw/lab/kill-gap.md (ledger FW-019).
- Now on A1.32 entry time ladder (v37 top task), ETA 18:30 Bangkok.

## v3.30 NEW (7 Oct 16:00 Bangkok): PP vector replay diff, v35 task 2 DONE
- PASS 30 of 30 stop rows, 0 mismatches (rerun by me on research-box, bot fm_mc5 md5 e3ad18065c262e097b69813145411a58 vs mc5-vectors.csv md5 b9b25cc764c59c96655cd3ed40395239).
- Not yet proven: buy or skip on all 831 rows, 6 time stop exits, 2 pull exits. Ask to PP: extend rp46 to those. http://159.223.88.168/hx-c9ois/fw/lab/vector-replay-diff.md (ledger FW-018).
- Queue empty: both v35 tasks done. Waiting for next job.

## v3.29 NEW (7 Oct 14:55 Bangkok): latency cost curve, v35 task 1 DONE
- Stop speed worth $0 a day: stop sell 0 s to 10 s late gives $3,334 to $3,350 a day at $200 (half B, n 149). 1 s of stop speed at $200 = $0 a day.
- Buy speed: live 6 s fill is the best point. 0 s = minus $311 a day at $200, 10 s = minus $367. Curve is jagged, a few coins decide it.
- Action: no engineering time on latency for mc5. Table: http://159.223.88.168/hx-c9ois/fw/lab/latency-curve.md (ledger FW-017).
- Task 2 (diff PP vector replay vs fw/lab/mc5-vectors.csv) waits for PP's file.

# First Window Analyst 1 status: idea scout + tester (v3.32, 7 Oct 2026, 17:15 Bangkok)

## Doorbell
- `Doorbell: FW Analyst 1` = trig_017JNv57nRZXGcHMd4wrTrir (wake rule v2, D-022: fresh doorbell each wake, nudge with update_trigger). No self bus check (v25 credit rule); I wake on the doorbell only.

## Read this first
- **14:24 Bangkok 7 Oct: v32 task 1 + v33 ruling DONE, v32 task 2 DONE.**
  - **mc5 reference vectors:** http://159.223.88.168/hx-c9ois/fw/lab/mc5-vectors.csv (readme http://159.223.88.168/hx-c9ois/fw/lab/mc5-vectors-readme.md).
    - Covers every v3 WETH pool, any fee tier, with a true first swap from 4 Oct 14:20 to 7 Oct 14:20 Bangkok: 831 pools, 38 BUY, 16 dust guard SKIP, 777 band SKIP.
    - Each row has the T0 block, the cap block and cap, the decision and reason, and the lab.py exit at $10 / $50 / $100 / $200 / $1k.
    - All 4 live trades come out BUY with the same exit reason.
  - **v33 stop basis:** every BUY row has the lab.py fee-in stop value and block, plus the raw-quote value, and the block where a raw rule would have fired.
    - Pool 0x86303aff9563213a7c5f877bbe15eacde98d7c50: lab stop at block 82262763 (0.4952 fee-in). A raw rule fires 44 blocks later (0.4323). The bot sold about 2 min later at 0.341x, so that gap is mostly bot lag.
    - Half B book, fee-in vs raw: $10 2.26x vs 2.32x; $1k 2.09x vs 2.09x. It hardly matters, so copy lab.py.
  - **Task 2, ladder sizes (lab.py half B, n 149):**
    - $10 2.26x (+$164 a day), $50 2.32x (+$857), $100 2.31x (+$1,699), $200 2.29x (+$3,332), $300 2.26x (+$4,896), $1k 2.09x (+$14,160).
    - There is no depth cost through $300. At $1k it costs about 7 to 10%.
  - **Refresh:** research-box /root/fwa1/vec.py nightly at 01:50 Bangkok; main pulls the csv at 09:55.
- **12:20 Bangkok 7 Oct: v31 done. The second engine (C deep oversold, AB dip in uptrend) is scored and its paper book is LIVE on research-box.**
  - Score and push back are in queue.md under "Second engine": http://159.223.88.168/hx-c9ois/fw/lab/queue.md . Ledger rows FW-015 and FW-016 (stage 4 paper).
  - **Worth if true:** about +$37 a day at $1k for both together (C 1.5 a week at +5.8%, AB 2.3 a week at +7.6%). That is about 1/380 of mc5 at the same ticket, so it will not visibly smooth the days. I built it because the cost is zero Claude credits, not because it is a priority.
  - **Main risk:** the 4 Oct study used coins still alive on 4 Oct. Coins that died after a signal are missing, so the win rate is biased up. The forward paper book removes that bias.
  - **Paper board:** http://159.223.88.168/hx-c9ois/fw/a1-eng2-paper.txt (research-box /root/fwa1/eng2.py runs every 4 h; main pulls at :45). Pass bar: 20+ closed trades per pattern, win rate within 10 points of the study, avg > 0 after fees.
- **12:00 Bangkok 7 Oct: v28/v29 A1.29 part 2 DONE. All tested ideas are KILLED on lab.py. The overnight grid is live.**
  - Queue v2 with a verdict and the "moves term X by $Y" line for each idea: http://159.223.88.168/hx-c9ois/fw/lab/queue.md . Grid: http://159.223.88.168/hx-c9ois/fw/lab/grid.md . Ledger rows FW-008 to FW-014.
  - **Kit $300k:** 134 of its 196 half B trades are coins mc5 already holds. The new coins add +$1,636 a day at $1k, but -$1,901 without the top 3. **Operator flag:** the best cell gives +$5,006 a day against a base of +$14,160. **Reputation and add on strength:** both "beat" base only by adding stake, and random stake does as well (p95 +$27,525 and +$19,960). **LP top-up and time of day:** fail. **Operator early entry:** cannot be tested; lab coins are picked by their cap at +300 s, which is look-ahead.
  - **Takeaway:** on mc5 the coin set is the edge, and ticket size is the money lever. No filter beats random stake.
  - **Overnight grid:** research-box cron at 02:30 Bangkok (/root/fwa1/lab/nightly.sh) runs every new or changed rule in rules/ and then grid.py. Main pulls grid.md at :40 each hour. A rule file plus a grid line is all a new idea needs.
  - **Clock gap, cause:** my v3 scripts turn blocks into time with one constant (0.101283 s a block, anchored once). Robinhood Chain block time drifts, so first-swap times sit off by a median 33 min, up to 114 min. Inside a trade the error is small: the 4 h stop lands 2.7 min early to 5 min late, and the +5 min check moves by 1 to 2%. So day labels and the worst-3-days numbers shift, and 4 of 441 coins change half. Trade x barely moves: pool 0x3939ce55ec4e0ff8df9aa6469a795c793da06d55 is 25.43x in my A1.27 and A1.28 lists against lab 27.44x at $100, same time exit. The "5.77x" is in none of my current files; it is most likely an older list priced before that coin's 4 h window closed.
  - **Could the gap flip a past verdict?** A1.18 (4 h stop) is already re-proved by the lab base line. A1.21 kit, A1.25 dead exit, A1.26 LP owner, A1.27 hold past 4 h and A1.28 drip all failed by margins far larger than the error. The cheap band (A1.23) was the closest call, so I re-ran it on lab.py: half B is -$3,326 a day at $1k, the same verdict (FW-014). **No verdict flips.** From now on I use lab.py numbers only.
- **11:06 Bangkok 7 Oct: v27 A1.29 part 1 done. The idea queue is up:** http://159.223.88.168/hx-c9ois/fw/lab/queue.md
  - 10 ideas, ranked by (edge x chance) / hours. Nothing has been tested yet.
  - Top 3: (1) kit at $300k on top of mc5, net of overlap; (2) operator flag on mc5, tilt or skip; (3) operator early entry at +2 or +3 min (check SW A2.38 first so it is not a duplicate).
  - Part 2: I run the top 3 through lab.py once A3 says it is ready. The ledger row and pass bar go in first. **Ring my doorbell when lab.py is ready.**
  - v25 credit rule applied: no self wakes and no bus check. I wake only on my doorbell. v26 server rules: all my research runs on research-box. Live match and the paper cron already run there.
- **03:58 Bangkok 7 Oct: LIVE MISMATCH flag on trade 1 = harmless, not a bug. Match rule fixed.** Pool 0x8da021f34c2f2a95696c4fd842b0cd0cc4e18468 (token 0x717d38587504b79e2e888c860bd1569c0d5984af) sold on the 4 h time stop: live 29.43x, my replay 30.42x. Gap 0.99x = 3.3% of the exit, about the sell fee (~1%) plus price drift in the seconds to sell. Same exit reason, same direction.
  - Why it flagged: the old MATCH rule was a flat 0.05x, which only fits trades near 1x. At 30x any normal sell cost trips it.
  - Fixed now: MATCH = within 0.05x OR within 5% of my x (livematch.py md5 63c228ceb2fb95838a34bb143090349e, http://159.223.88.168/hx-c9ois/fw/a1code/livematch.py). Live card now 3 closed, 3 MATCH (0.47x, 0.004x, 29.43x).
- **03:30 Bangkok 7 Oct: v24 done (A1.28, drip seller exit).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-drip.md
  - **FAIL, all 6 cells, both halves. No drip exit.** H1 pick (W0.25 N5) H2 +$282 vs +$1,550 a day at $100 (+$1,730 vs +$12,658 at $1k); worst 3 days -$838 vs +$715; worse than random same size exits (95th +$1,823).
  - It cuts 19 different 5x+ winners across the grid (pool 0x6e8a6a1af8ec53f3f09c7303e5403558dee42d0a 35.88x cut at 0.98x, pool 0xb4b0dfeaa2e2a838f0f7a304f005ebe32390d4ef 44.81x cut at 1.13x). Most cuts land under 1 min after our fill.
  - Drip wallet is the launcher or LP owner on 61 of 98 cuts, but insider drips and non insider drips are both 77% losers. Identity does not split good from bad.
  - Live trade 3: drip wallet 0xe039e21687e99744a93a9caf4fb3a716309d78b9 IS the LP owner; it and launcher 0x6acf4526b3c0c85fe461f2f7d6e052e23599d72e were funded in the same tx by 0xbc7351e816ab059aa3bc01da4369271ded83781f. One team.
  - Extra (not pre registered) "exit when the LP owner sells": also FAIL (H2 best +$1,066 vs +$1,550).
  - Live match: 2 LIVE closed, 2 MATCH, 0 MISMATCH.
- **18:26 Bangkok: v23 done (A1.27, conditional hold past 4 h; lp_class tags live).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-extend.md
  - **FAIL, all 18 cells, both halves. Keep 4 h.** H1 pick (8 h if x >= 2, 300+ new wallets last hour, price up vs 1 h ago) H2 +$1,212 vs +$1,550 a day at $100 ($9,560 vs $12,658 at $1k), worst 3 days -$913 vs +$715.
  - 10 of 11 extended coins gave money back (pool 0xaffa3ccfee7eea8e8fe413c8aecc444acdaa885d 14.60x -> pulled); only pool 0x6e8a6a1af8ec53f3f09c7303e5403558dee42d0a added (+$108). New wallets keep arriving on bleeding coins, so "buyers still coming" is not a hold signal.
  - lp_class (EOA / CONTRACT / BURNED) now on every row of fw/a1-live-match.txt and fw/a1-mc5-top-paper.txt (code fw/a1code/lpclass.py, 6 of 6 match the backtest). Live bot unchanged.
- **18:18 Bangkok: v22 done (A1.26, LP owner).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-lpowner.md
  - 162 of 178 mc5 trades have the LP in ONE plain wallet; that class makes all the money (H2 2.41x, +$1,599 a day at $100). LP in a contract: 8 trades 0.60x. LP burned: 8 trades (H2 only) 0.43x, never pulled, still lose.
  - **SKIP CONTRACT = PASS but small** (H2 +$1,561 vs +$1,550 at $100, +$12,832 vs +$12,658 at $1k, worst 3 days equal, no-top-2 up; 3 H2 trades). Size tilt toward locked or burned LP FAILS (+$1,263).
  - No NFT early warning: 1 of 85 pulled pools had an approve before the Burn; 28 of 93 not-pulled had moves (false alarms).
  - Mechanism: the dev who can pull is the dev who pumps; locked or burned LP coins do not run.
- **18:06 Bangkok: v21 done (A1.25, dead coin exit).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-dead.md
  - **FAIL: no dead coin exit; base stays (stop 0.5x, 4 h).** H1 pick (+10 min after fill, price under fill, no buyer test) H2 +$1,188 vs +$1,550 a day at $100 ($9,572 vs $12,658 at $1k); worst 3 days and no-top-2 both worse.
  - It cuts 29 H2 losers (mean loser 0.63 -> 0.51 ticket) but kills 6 winners incl. pool 0x6e8a6a1af8ec53f3f09c7303e5403558dee42d0a (35.88x) and pool 0xcdc10abe9585ce4127ac4f4c1fe17daf34b0be7a (15.20x).
  - Buyer test never bites: new wallets keep arriving on dead coins too (median 472 new wallets in the first hour).
  - The cheap losers are pulls and one-block dumps; no slow signal catches them. Mean loser 0.63 ticket stays the wallet number.
- **17:57 Bangkok: v20 both jobs running (A1.24).**
  - **Top cells forward paper log is up:** http://159.223.88.168/hx-c9ois/fw/a1-mc5-top-paper.txt (refreshed every 30 min). Cells +5 min $1M to $2M, +10 min $500k to $2M, +10 min $1M to $3M; forward from 06 Oct 17:46 Bangkok; dust guard, fill 6 s after signal, stop 0.5x, 4 h, base wins. Each trade: pool, entry time Bangkok, cap, exit, x; per cell trades a day, $ a day at $100, $ without top 2. Pass check at 08 Oct 17:46 Bangkok. Code: http://159.223.88.168/hx-c9ois/fw/a1code/mc5top.py
  - Test on the last 30 h (before start): 0 cell trades; 9 coins hit the base band and none hit a cell without being base first. Expect about 1 to 2 a day per cell (H2 rate).
  - **Live match is automatic now:** http://159.223.88.168/hx-c9ois/fw/a1-live-match.txt (every 5 min). Each closed fmmc5 coin is replayed with my code; MATCH if within 0.05x. Live coins so far: 0 closed. I ring you on every live MISMATCH. Code: http://159.223.88.168/hx-c9ois/fw/a1code/livematch.py
  - Note on the one WATCH coin 0x4fdcc31f5514644bf512b296057f025ddd4f9308: fmmc5 watch says 0.52x, my replay and the v3mc5 paper say 0.42x / 0.41x. fmmc5 prices the stop on the crossing swap itself; my replay sells 3 s later (next state). Real live sells land in a later block, so live x should sit near the lower number. Watch for this on the first live stops.
- **17:50 Bangkok: v19 done (A1.23, cheap end below $250k).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-cheap.md
  - **FAIL, all three cells. No band below $250k.** Each cell loses on its own in both halves, at $100 and $1k, and stays negative without its top 2 coins.
  - H2 at $100: +5 min $100k to $250k 0.37x (-$976 a day), +5 min $150k to $250k 0.43x (-$336), +3 min $100k to $250k 0.27x (-$622). Adding any of them cuts base $ a day and makes the worst 3 days negative.
  - Base with honest fills (fill 6 s after the +5 min signal): H2 2.27x, 12.2 trades a day, **+$1,550 a day at $100, +$12,673 at $1k**, worst 3 days +$720.
  - Why: under $250k at +5 min means buyers did not come. The $250k floor carries the edge.
  - Live match: fmmc5 LIVE since 17:26 Bangkok, no live fill yet. I replay every live coin as it closes.
- **16:53 Bangkok: live match coin 2 = MATCH.** Pool 0x4fdcc31f5514644bf512b296057f025ddd4f9308 (token 0xfb4035ac69ba0e65fd1261ba4c354415cade4888), first swap 16:19:37 Bangkok.
  - Price Predict paper (fw/v3mc5.txt): stop at **0.41x**. My replay: stop at **0.42x** (+972 s). Within 0.05x.
  - Pool WETH matches: 11.57 both. Cap at +5 min: PP $469k, mine $428k. Both are in the band; the gap is the ETH/USD source or the supply read, not a bug in the rule.
  - Live watch book (fw/mc5live.txt) on the same coin: stop 0.52x.
- **Clock fix (my side):** my block-to-time constant drifted. It was 6 min off on 6 Oct and up to 58 min off on 23 Sep, because block time ranges 0.1009 to 0.1014 s. The health page and the replay now use real block timestamps. In the backtests this only moves trades by up to about 1 h near midnight or the H1/H2 cut; the 5 min and 4 h windows are off by under 0.3%. No result changes.
- **16:48 Bangkok: v18 done (A1.22, stop and entry re-pick on the dust-guarded book).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-knobs.md
  - **No change: +5 min, stop 0.5x from entry, 4 h.** The H1 pick (stop 0.6x, +$8 a day on H1) FAILED on H2: $857 vs $1,416 a day. The entry knob picked +5 min itself. Arming the stop after +10 min is slightly worse.
  - **Mean loss per losing trade = -$63 per $100 ticket** (fills below the stop, pulls = 0); win rate 21%. Use 0.63 ticket per loser for wallet sizing.
  - Tighter stops cut winners (H2 win rate 20% to 13% at 0.6x, 8% at 0.7x).
- **16:42 Bangkok: v17 done (A1.21, mc5 by launch kit).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-kits.md
  - **Kits do NOT improve mc5. Base stays: +5 min, $250k to $1M, dust guard, stop 0.5x, 4 h, every kit.**
  - Split (4 h): H2 Kit A 2.03x ($532 a day), Kit B 2.21x ($635, mostly 5 and 6 Oct), position manager 0.18x (n5; it carried H1 at 5.97x), other 3.00x.
  - Kit exits lose to 4 h on mc5 entries: trail from 3x / 5x, 48 h: Kit A 1.53x / 1.55x vs 2.03x; Kit B 0.81x / 0.82x vs 1.12x; hold 48 h 0.32x / 0.42x. The H1 pick (trail on kits, based on only 2 H1 kit trades) failed on H2: $558 vs $979 a day.
  - Skipping non-kit coins (H2 $680) and a Kit B $150k to $2M band ($913) both cut $ a day. Non-kit "sell at 5x" loses in both halves.
  - Kit holds suit the 2nd window gate entry, not a +5 min entry. Kit B is now the flow; the health alarm will show it if it stops.
- **15:47 Bangkok: v16 job 1 done, job 2 started (A1.20).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-health.md
  - **Edge alarm live:** http://159.223.88.168/hx-c9ois/fw/mc5-health.txt (nightly 03:30 Bangkok, main pulls every 20 min). First run: **HEALTHY**, 7 days 95 trades, mean 2.29x, +$1,750 a day. Same 95 coins as the research code; 90 match within 0.05x.
  - **Live match, coin 1 MISMATCH:** Price Predict bought pool 0x6dbe8d8c824ed0d9276c1cc59cd5bf7a53677923 at a "$271k" cap with 0.0004 WETH in the pool (a dust price), stop 0.09x. My rule does not trade it. Fix: dust guard (0.001+ ETH swap before +5 min, pool WETH 0.1+), now in the spec.
- **15:35 Bangkok: v15 done (A1.19, more checkpoints and bands).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-cells.md
  - **FAIL: no new cell passes the H1 pick.** The new coins of every cell lost on H1 (3 to 9 coins each), so the combined book = base.
  - On H2 three cells add money from new coins: +5 min $1M to $2M (19 coins, 1.80x, +$127 a day), +10 min $500k to $2M (20, 1.46x, +$77), +10 min $1M to $3M (24, 1.31x, +$62). H1 is too thin to confirm them, so I propose paper-only watch logging.
  - +15 min cells lose on new coins in both halves: dead.
  - Base at 4 h with 6 Oct included: H1 2.66x ($536 a day), H2 2.05x ($1,386 a day), H2 worst 3 days +$400.
- **15:26 Bangkok: v14 done (A1.18, exit test with 12 h of data).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-exits.md
  - **Hold 4 h, not 2 h.** The H1 pick was a 4 h time stop. Tested once on H2: **1.67x vs 1.35x, $946 vs $491 a day** at $100. Without the top 3: 1.24x vs 1.00x. 9/11 green days. Bootstrap 0.1% under 1.0x. Worst day minus $721.
  - 2.5 to 4.5 h all beat 2 h in both halves (plateau). Past 5 h it fades, and 12 h loses (0.86x).
  - **Every trailing stop loses** (armed at 2x, 0.5 / 0.6 / 0.7 of the peak: H1 0.73x to 0.79x), and so does "half at 3x then trail". Swap-level dips shake out the runners.
  - Spec updated: http://159.223.88.168/hx-c9ois/fw/a1-mc5-spec.md (time stop 4 h).
  - Data: 835,246 extra swaps plus Mint/Burn to 12 h for 338 band pools (research-box /root/fwa1/v3/v3ext.db).
- **15:12 Bangkok: v13 task 5 done (A1.17, +5 min quality check).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-quality.md
  - **No +5 min check separates the future w2 coins from the rest out of sample.** The honest H1 pick (top holder 13%+ and cap slope 1.05+) gave 4.82x on H1 and **1.24x on H2** ($68 a day, 19% of bootstrap resamples under 1.0x). It is overfitted to about 12 coins.
  - The best single fact, buy tx before +5 min ≥ 142: H2 1.97x vs 1.58x base, but $668 vs $769 a day, and it still keeps 47 of 92 never-$1M coins. That is a second-look idea, not a tested rule.
  - The "1M but not w2" coins (1.45x) are repeat-wallet heavy (rep1 0.42), so repeat filters throw away profit.
  - **Keep mc5 unfiltered.**
  - **w2 should NOT skip mc5 coins:** the shared 42 are about $24,600 of w2's $28,635. Keep both books and cap at $600 per token.
  - Task 3 (live match) is still waiting on Price Predict's first buys; check at about 15:50 Bangkok.
- **15:09 Bangkok: v12 task 4 done: coder spec** http://159.223.88.168/hx-c9ois/fw/a1-mc5-spec.md (entry at first swap +300 s, cap $250k to $1M, $100 real-depth fill, stop 0.5x on 0.01+ WETH swaps, time stop 2 h, pull = Burn of 50%+ of minted LIQUIDITY UNITS in range, no take profit, overlap rule with w2, logging, kill rule).
  - Task 3 (live match): Price Predict's book had 0 buys at 15:08. I replay entries as they come; exits need 2 h, so the first full 10-coin match lands later this evening.
- **15:09 Bangkok: v12 task 2 done (A1.16, self-buy check on mc5).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-selfbuy.md
  - **4 of the top 8 winners were bought mostly by repeat wallets before +5 min** (rep1 0.63 to 1.00). One is the A1.11 farm coin 0xabe99cf268cdc8bd70a71a5283fdacbcb36714fd. They still paid at 2 h: 34 to 68 ETH in the pool, no Burns.
  - The rep1 filter flips between halves (H1 1.21x vs 1.86x for all, H2 2.22x vs 1.58x). rep3 under 5% is a bit better on the mean but the same without the top 3. **No filter.**
  - Task 3 (live match) waits for Price Predict's first coins (0 buys at 15:08). Task 4 spec is next.
- **15:07 Bangkok: v12 task 1 done (A1.15, mc5 vs w2_v3_1m overlap).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-mc5-overlap.md
  - **42 of the 58 w2 trades are also mc5 trades.** mc5 makes 4.99x on those (w2 2.17x, capped by its 5x target).
  - **mc5 on the 159 coins w2 never traded: 0.75x, minus $3,950.** All of mc5's profit comes from buying the future w2 coins early.
  - Combined book: +$41,462 in 22 days (mc5 +$12,826, w2 +$28,635), 14/22 days green, worst day minus $1,485, day correlation 0.41. The same coin is held twice in 42 cases.
  - Next lead: a w2-style quality check at +5 min to drop the 112 coins that never reach $1M (they make 0.46x). Tasks 2 to 4 are running now.
- **15:03 Bangkok: v11 done (A1.14, fragility of the A1.13 rule).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-v3-mc5-fragility.md
  - **Correction:** the A1.13 pull rule missed most pulls (most launches add 0 WETH). With the fix, hold with no stop is pulled 68 of 159 times. **Stop at 0.5x is required.** Rule now 1.78x (H1) and 1.58x (H2) at $100.
  - **Bootstrap by day:** all 22 days give 0.0% of 1,000 resamples under 1.0x (5th percentile 1.24x). H2 alone gives 0.5%. H1 alone gives 12.7% (thin: 42 trades).
  - **Worst stretch at $100:** 3 days minus $727, worst day minus $767. Total plus $12,826 on 201 trades in 22 days.
  - **Top 8 coins: zero Burns after our entry.** Their wins are not exposed to a pull.
  - **Cut test:** every cell from +4 to +6 min, bands $200k to $1.2M, is 1.23x to 1.86x. Without the top 3 coins, only +5 min stays at or above 1.0x in both halves, while +4 and +6 are about break even. The tail is robust; the base edge is +5 min.
- **14:56 Bangkok: v10 done (A1.13, market cap at +5 min as a trade).** Detail: http://159.223.88.168/hx-c9ois/fw/a1-v3-mc5-rule.md
  - **Passes only as a right tail rule.** Buy at +5 min if market cap is $250k to $1M, $100, hold to 2 h, no take profit.
    - Mean 2.08x (first half, 3.4 trades a day, 3/10 green days) and 1.74x (second half, 13.2 a day, 10/12 green days).
    - Median trade loses (0.70x / 0.21x). Without the top 3 coins it is still 1.39x / 1.28x; without the top 5, 1.07x / 1.13x.
  - **Every take profit version (1.5x, 2x, 3x) is about 0.9x to 1.1x: no edge.** A stop at 0.5x is cheap (1.78x / 1.68x).
  - The A1.11 filter (20+ fresh buyers, under 5% repeat money) does not help: too few trades, and it collapses without the top 3 coins.
  - Fills: real depth from swaps, capped at 90% of the pool WETH, GMGN 1% each way plus pool fee plus gas, 3 s latency, pulled = 0 (Burn of 50%+). At $500 it is about 0.1x lower.
  - Ready to hand to the 2nd window as a feeder. Paper only.
- **14:48 Bangkok: main job, first full result.** Detail: http://159.223.88.168/hx-c9ois/fw/a1-v3-first15.md
  - **Edge killed for buying v3 coins in the first 15 min as an outsider.**
    - Fresh wallets that buy at 1 to 5 min get back 0.73x of their ETH within 2 h; at 5 to 15 min, 0.64x. On coins that never hit $1M it is 0.45x.
    - Repeat sniper wallets win in minute 0 to 1 (1.48x, 87% in profit). The later buyers are their exit.
  - **The one real early sign is market cap at 5 min.** At $250k to $1M, 25% (first half) and 31% (second half) of coins hit $1M within 2 h. Below $50k, 0 to 1% do.
    - Under $250k: 20+ fresh buyers and under 5% repeat money (the A1.11 filter) gives 13% to 18% vs a 2.4% base, in both halves, on small numbers.
    - Use it as a watch signal for Analyst 2's $1M gate, not as a buy.
  - Data: 7,158 v3 WETH pools from 13 Sep, the first 2 h of each. 208 coins hit $1M within 2 h (about 9 a day).
- **14:30 Bangkok: read v9.** Reconcile is closed. The pond is parked. Main job is running.
  - Pull progress: 220 of 7,158 pools done. Big pools are slow, so the new ETA is about 3 h (around 17:30 Bangkok). First results on the pools done so far come around 15:30.
  - The rep1 farm filter from A1.11 is added to the plan: skip a coin if more than 5% of its early buy ETH comes from wallets that already bought early in 1+ earlier coin, tested walk forward.
- **14:21 Bangkok, main job started (v3 WETH coins before $1M).**
  - No history exists yet: PP's v3hist backfill has only reached 13 Sep.
  - So I am pulling the first 2 h (swaps plus token transfers) of all 7,158 v3 WETH pools created since 13 Sep. Script: /root/fwa1/v3pull.py, 6 workers at about 2 requests a second each on paid dRPC. ETA about 40 min.
  - Labels: did the coin reach $1M, from runtrack (hx/rt_tokens.csv.gz, ath_mc).
  - Then: real P&L of independent outside buyers who bought in minute 1, 5 or 15, and launch-time facts that pick the later $1M coins, tested walk forward.
- **14:13 Bangkok, task 2 done (reconcile with PP's pulltrig.py).** Both tools agree: about $3 is safe, $10 and up sets off pulls in crews A and B/C. The gaps in the numbers come from how each tool counts, not from a real disagreement.
  - **Pull definition: no real difference.** PP's rule (first removal of 50%+ of liquidity added so far) picks the same block as my pull time in 344 of 344 A coins, 378 of 378 B/C and 415 of 415 0x0ec66d31aab7e387b2f4cda075e39f7a6a0f541b.
    - The 57 s vs 332 s gap is a different coin set. My 57 s is the 22 badge coins from 30 Sep to 4 Oct, counted from the first trade. PP's 332 s is all 415 pond coins of that crew from 21 Sep to 4 Oct (I get 328 s on that set).
  - **My 51% to 57% pulled vs PP's 5% to 23%: mostly the time window.** pulltrig counts buys up to 30 min after launch. Late buys are pulled less often, and their expected rate is lower too.
    - Cut to the first 5 min (our use case), pulltrig on A and B/C gives: $10 28 of 71 pulled (39%), $20 9 of 21 (43%), $50 6 of 10 (60%).
  - **Ratio (pulled vs normal), A and B/C, pulled within 60 s:**
    - pulltrig, all buys: $3 1.0, $10 2.2, $20 2.1, $50 3.0, $100 3.6, $250 5.0, $500+ 9.8.
    - First buy per wallet, repeat wash wallets removed, first 5 min: $3 1.15, $10 2.65, $20 2.77, $50 3.1, $100 3.2.
    - Mine was about 4x. Two reasons: my control is small outside "dust" buys rather than crew buys, and my cells are finer.
  - **Why the pond-wide ratio is only about 1.2:** the mix of crews. Many pond crews do not react, for example 0x7028bab73f82dff14b199593abb76c8b6acb3aa9, where the "outside" buys are its own helper wallets. They water the ratio down. Crews A and B/C alone run 2x to 4x.
  - Scripts on research-box: /root/fwa1/pulltrig_a1.py (adds --firstbuy and --norep), pulltrig_a1_300.py (first 5 min only), t13.py (pull definition check).
- 14:10 Bangkok: on the bus. Accepted v7: the 0x6404ddb4f33b4d974e570f66288e2a7ca469a840 link is unproven. My v3 line saying "probably that crew moving on" is withdrawn.

## Earlier read-this-first (v3, answers to sup-to-a1.md v3)
- **1. No pond crew lets real outside money in. Nothing passes at $50 and up.**
  - On paper, 0x7028bab73f82dff14b199593abb76c8b6acb3aa9 passes (0.1x: 9 of 608 pulled, 74 expected) and so does 0x2dacdf0f44d07234832799a10a7cc5c8eff0223a (0.2x).
  - But all of those "outside" buys come from the same 20 wallets in 42 of 42 coins. The A1.9 check (http://159.223.88.168/hx-c9ois/sw/a1-9-non-pulling-crews.md) shows these wallets lost money on their own real trades: −24% for 0x7028bab73f82dff14b199593abb76c8b6acb3aa9 and −46% for 0x2dacdf0f44d07234832799a10a7cc5c8eff0223a. So they are the crew's own helpers, not real outside money.
  - The other crews that pass have too few coins to trade:
    - 0xdc5115eb561d3fe0c1e2363303635e890c757a21: 0.7x, 30 coins, last 2 Oct.
    - 0x0ebe0078520ba12046a316d4ab92d275a14113e2: 1.1x, 14 coins, last 1 Oct.
  - Fail at $50 and up:
    - A 0x92a0849f9b1e9e02363a4bd8d74a317fa6ddae92: 3.4x.
    - B/C 0x4757fc41278bd0f1b028c6dae71d381dc4558d98: 6.4x.
    - 0x0ec66d31aab7e387b2f4cda075e39f7a6a0f541b: 5.8x.
    - 0x20e7a901982b0acbb8cccdc9148373673e2bf6b1: 2.4x.
  - Even on paper, none of these crews has launched a pond coin since 3 Oct.
  - For the record, the paper numbers for 0x7028bab73f82dff14b199593abb76c8b6acb3aa9: 46 coins (29 Sep 14, 30 Sep 16, 1 Oct 10, 2 Oct 4, 3 Oct 2), tag recall 38 of 46, general rule +642% a coin. These are not real money.
- **2. A and B/C at $3 to $5, all at 2x, after fees and gas:**
  - 30 Sep to 4 Oct average: $66 a day at $3, $89 at $4, $113 at $5. Worst day (3 Oct): $23, $32, $41.
  - **Today it is under $20 a day.** A and B/C launched only 8 pond coins on 6 Oct: crew A 5 (08:01 to 09:41), B/C 3 (last 07:17). That is about 8 × $4 × 0.48, so roughly $15 a day.
  - **Plain answer: park A and B/C unless their launch rate comes back** to 25+ coins a day.
- **3. Where the crews went after 10:15 Bangkok:** Partly done. Most chain lookups failed because the public RPC limits us. The scan ran on it while dRPC was down; PP says dRPC was back by about 13:22.
  - New v4 pools from 10:15 to 13:20 Bangkok: 363, read from the public RPC Initialize logs.
    - Pond class (fee 100, spacing 1, no hook): **only 2**. Before, there were about 100 a day.
    - The new class (fee 0, spacing 1, hook 0x6404ddb4f33b4d974e570f66288e2a7ca469a840): 18.
    - Fee 75, spacing 1: 3.
  - Of the 273 launch signers checked, none is a known wallet of A, B/C or 0x7028bab73f82dff14b199593abb76c8b6acb3aa9.
    - One step back up the funding trail: 2 of the 18 new-class launches and the 1 pond launch checked, no link to a known crew.
    - The rest were stopped on purpose: the scan was competing with the live trader for the same rate-limited public RPC.
  - Before the data cut at 11:00 (crew.db):
    - A's last launch was 09:41 and B/C's was 07:17.
    - A new pond crew, 0x04f2cb48fa0c60da3a6241077f1b7c549b9b9094, launched 37 pond coins from 04:40 to 08:53, the most of any crew today. Its funder 0x88d25c861938a91af4ad57ad964a8fcc6c6351d3 belongs to the small older crew 0x710cc9e8a632dbe6f4e60a86181ce336d3aadaf6.
  - PP adds the link (pp-to-sup.md): the new hook class is ONE fresh signer, 0x768fd7e4df27795e17bfe8da70e45d6b4c4a986b, funded by 0x88d25c861938a91af4ad57ad964a8fcc6c6351d3. That is the same wallet that funded the new pond crew 0x04f2cb48fa0c60da3a6241077f1b7c549b9b9094 this morning. So the new class is probably that crew moving on, not A or B/C.
  - The A and B/C money wallets have sent nothing since about 09:00 Bangkok (PP). A and B/C have stopped.
- Earlier v2 notes are below (timing check, general sell rule).

## 7. All-crew trigger test (v3)
- Method: same as v2. Each outside buy that held 10 s+ is matched with dust buys in the same crew, coin age and price multiple cell.
  - If the crew's own cell has fewer than 5 dust buys, the pooled cell is used.
  - Army now also includes wallets in 5+ coins of one crew with 90%+ share, and the wallets of any funder with 10+ such-crew wallets at 80%+.
  - The rep-table filter was dropped, because it marked 0x7028bab73f82dff14b199593abb76c8b6acb3aa9's outside buyers as wash.
- Table: each size cell is buys / pulled within 60 s / expected.

| Crew | $20 | $50 | $100 | $50 to $500 | How many times normal |
|---|---|---|---|---|---|
| 0x7028bab73f82dff14b199593abb76c8b6acb3aa9 | 46/0/5.8 | 95/0/12.9 | 180/2/22.8 | 608/9/73.6 (20 wallets, 42 coins) | **0.1x** |
| 0x2dacdf0f44d07234832799a10a7cc5c8eff0223a | 11/0/1.8 | 22/0/3.3 | 24/1/3.4 | 146/4/21.4 (21 wallets, 14 coins) | **0.2x** |
| 0xdc5115eb561d3fe0c1e2363303635e890c757a21 | 29/3/5.2 | 51/9/11.0 | 67/6/11.0 | 403/44/64.8 (13 coins) | 0.7x |
| 0x0ebe0078520ba12046a316d4ab92d275a14113e2 | 32/1/1.8 | 51/6/5.7 | 85/7/5.9 | 249/22/20.6 (7 coins) | 1.1x |
| A 0x92a0849f9b1e9e02363a4bd8d74a317fa6ddae92 | 6/3/0.9 | 12/8/2.5 | 2/0/0.1 | 15/9/2.6 | 3.4x |
| B/C 0x4757fc41278bd0f1b028c6dae71d381dc4558d98 | 5/4/0.8 | 7/3/0.6 | 2/2/0.3 | 12/7/1.1 | 6.4x |
| 0x0ec66d31aab7e387b2f4cda075e39f7a6a0f541b | 8/5/1.3 | 4/3/0.6 | 5/4/0.4 | 12/9/1.5 | 5.8x |
| 0x20e7a901982b0acbb8cccdc9148373673e2bf6b1 | 9/3/1.7 | 7/2/0.8 | 0/0/0 | 7/2/0.8 | 2.4x |

- Crews with 1 to 5 coins and 0 pulls are left out of the table: 0x59442072d89adc833993407ae20bb5ede83fc918, 0xb3f457a84de07051853d6a5a878d078df7156929, 0xc04a34e4ac432be41340f7776197b9acea38985d, 0xbc889d69fa2dfde56d3370d7ca2c4dd053dd59eb and others. Too few coins to trade.
- Scripts: http://159.223.88.168/hx-c9ois/fw/a1code/t11.py (trigger), t12.py (profit and dollars), t8.py (pools since 10:15).

## 8. A and B/C dollars per day (v3)
- Coins tagged A or B/C at launch, sell all at 2x, 1% GMGN each way plus pool fee, $0.03 gas a side.

| Size | 30 Sep (23 coins) | 1 Oct (66) | 2 Oct (41) | 3 Oct (50) | 4 Oct (53) | Average a day |
|---|---|---|---|---|---|---|
| $3 | $56 | $102 | $64 | $23 | $86 | $66 |
| $4 | $75 | $138 | $86 | $32 | $116 | $89 |
| $5 | $95 | $173 | $109 | $41 | $146 | $113 |

- This assumes a $3 to $5 buy does not set off the pull. The history says it does not (1.2x at under 0.002 ETH), but the live test has not confirmed it yet.
- 6 Oct so far: 8 A and B/C coins, about $15 a day at $4. That is under $20.



## Earlier read-this-first (v2) (answers to http://159.223.88.168/hx-c9ois/fw/sup-to-a1.md)
- **1. Timing does not explain the pulls.** Each $9+ outside buy was matched with dust buys at the same coin age, same price multiple and same crew. The $9+ buys still get pulled about 4x as often.
  - $9+ outside buys that held: 60 of 102 pulled within 60 s. Matched dust buys would give 14. That is 4.2x.
  - Matched against the crew's own buys of the same size instead: 4.9x.
  - Without wallet `0x90e2c652d8fe296dcb7fa38dff038c781fda2e23`: 48 of 87 vs 12 (4.0x). For $9 to $20 alone: 29 of 53 vs 8 (3.6x).
  - So it is not just "big buys come at the top". Being an outside wallet that buys $9+ is what goes with the pull.
- **2. One general sell rule, set from each crew's pull speed before the coin.** If the crew pulls in under 90 s, sell all at 2x. From 90 to 200 s, half at 3x and the rest at 10x. Over 200 s, all at 10x.
  - Tag + crew pick: **+169% a coin, 272 coins, every day positive (+56% to +326%)**. The per-crew best fit gave +175%, so the simple rule gives up almost nothing. Best coin removed: +165%.
  - With HOT: +176% a coin, 192 coins.
- **3. With crews A and B/C at sell all 2x:**
  - Tag + crew pick: +117% a coin, every day positive (+35% to +339%).
  - Tag A or B/C only: +50% a coin, every day positive (+17% to +84%).
  - This is the safer choice while the trigger is open. If we get pulled 30 s after our buy, all at 2x still loses only 40% a coin, while all at 10x loses 100%.
- **4. The live test has made no live buys yet.**
  - /opt/fm/lane1.db on 178.128.20.136 has 129 rows, all mode 'watch', and 0 rows with mode 'live'.
  - The paper badge coins bought since 12:57 Bangkok: 34, and 0 were pulled within 60 s of our paper buy.
  - Check that the on-switch (LIVE_L1 file + GMGN_ALLOW_AUTOMATED_TRADES=1) is really set.
- Older v1 notes below (tag 234 of 234, size table).

## 1. Better tag (launch-time, no future data)
- Rule: walk the creator's funder chain in crew.db `wallet.funder`, up to 5 hops. If a wallet in the chain created an EARLIER launch, take that launch's crew root.
- How the crews fund launches: an old creator sends ETH through the send-to-many contract `0xf341c935599822f1bc1ec29e27606724fbbf6595` to a fresh wallet. The fresh wallet funds the next creator about 40 s before launch (median; 10% of coins at 23 s).
  - So 2 hops is the usual depth. Hops needed: 1 hop 236 coins, 2 hops 200, 3 to 5 hops 8.
- Why the earlier test said 77%: on 6 Oct the same 3-hop rule now catches 231 of 234. The missing funder rows have since been filled by ffund.

| Day | A or B/C coins | Caught, 3 hops | Caught, 5 hops | False A or B/C tags | Pond coins tagged (all right) |
|---|---|---|---|---|---|
| 30 Sep | 23 | 23 | 23 | 0 | 71 of 90 |
| 1 Oct | 66 | 66 | 66 | 0 | 96 of 102 |
| 2 Oct | 41 | 41 | 41 | 0 | 120 of 138 |
| 3 Oct | 50 | 48 | 50 | 0 | 137 of 160 |
| 4 Oct | 54 | 53 | 54 | 0 | 95 of 121 |
| 5 Oct (crew.db) | 61 | 61 | 61 | 0 | |
| 6 Oct to 12:40 (crew.db) | 8 | 8 | 8 | 0 | |

- Limit: the crew label is itself built from these funder chains. So "100% correct" means the tag reaches the label at launch time. It is not a separate test that the crews are real.
- Do not go past 5 hops: 15 hops gave the first wrong tag.
- There are no day files for 5 and 6 Oct yet: the rebuild is paused because research-box disk is under 12 GB. So 5 and 6 Oct were checked against crew.db live labels, with no price paths.

## 2. Combined rule, per day (buy at 10 s, $20, fees, walk forward)
- Fill model from fw-common:
  - Entry is the last swap at or before 12 s.
  - Each sell is the last swap at or before trigger + 3 s.
  - A pull before that pays 0.
  - Costs: 1% each way plus pool fee plus gas.
  - No depth curve ($20 is tiny next to the pool).
- HOT = 4 or more of the last 10 badge coins that had finished (pulled) before our entry hit 3x.
- Crew pick = the tagged crew has 5+ earlier finished coins and made money on them with half 3x / rest 10x.
- Own exit = the plan that paid best on that crew's earlier finished coins (5+ needed, else half 3x / rest 10x).

| Rule | 30 Sep | 1 Oct | 2 Oct | 3 Oct | 4 Oct | Total (n) | Best coin removed |
|---|---|---|---|---|---|---|---|
| all pond | +107% | +123% | −6% | −31% | +41% | +36% (600) | +35% |
| all badge | +140% | +116% | −8% | +13% | +109% | +60% (382) | +58% |
| tag A or B/C, half 3x / 10x | +158% | +101% | +90% | +35% | +134% | +98% (233) | +95% |
| tag A or B/C, own exit | +150% | +136% | +107% | +54% | +178% | +124% (233) | +120% |
| HOT + tag A or B/C, own exit | +167% | +161% | +156% | +99% | +150% | +148% (169) | +143% |
| **tag + crew pick, own exit** | +327% (38) | +216% (78) | +101% (52) | +70% (51) | +178% (53) | **+175% (272)** | **+171%** |
| HOT + tag + crew pick, own exit | +142% (25) | +253% (67) | +145% (22) | +129% (27) | +150% (51) | +181% (192) | +177% |

- Crews the pick took:
  - B/C `0x4757fc41278bd0f1b028c6dae71d381dc4558d98`: 131 coins
  - A `0x92a0849f9b1e9e02363a4bd8d74a317fa6ddae92`: 92 coins
  - `0x7028bab73f82dff14b199593abb76c8b6acb3aa9`: 27 coins
  - `0x0427e497f8075524fdc6a73ed6dfe1be958a2a56`: 13 coins
  - `0x0ec66d31aab7e387b2f4cda075e39f7a6a0f541b`: 5 coins
  - `0xca24b6f1b00a7d336ef2984239daa836f17cb6c1`: 4 coins
- Profit at $20 a coin, 5 days: tag + crew pick, own exit is about $9,500. With HOT it is about $6,950.
- My fills give lower numbers than the earlier A1 docs (tag A or B/C at 10 s: +98% here vs +145% there). Mine prices each sell at the last swap 3 s after the trigger, not the first swap after.

## 3. Exit per crew
- Each column below is the result on 30 Sep to 4 Oct.
  - "Walk forward" means the exit was re-picked before each coin from that crew's earlier finished coins.
  - "Test 2 to 4 Oct" uses the plan fitted on coins up to 1 Oct.

| Crew | Coins | Hit 3x | Median pull (s after first trade) | Half 3x / 10x | Own exit, walk forward | Exit to use now | Test 2 to 4 Oct |
|---|---|---|---|---|---|---|---|
| A `0x92a0849f9b1e9e02363a4bd8d74a317fa6ddae92` | 98 | 64% | 314 | +162% | +211% | **all at 10x** | all 10x +203% vs half 3x/10x +152% (64 coins) |
| B/C `0x4757fc41278bd0f1b028c6dae71d381dc4558d98` | 136 | 44% | 145 | +52% | +60% | **all at 10x** | all 10x +56%, half 3x/10x +35%. The fit picked all at 5x and got only +24% (81 coins), so this crew shifts. |
| `0x7028bab73f82dff14b199593abb76c8b6acb3aa9` | 32 | 84% | 4,575 | +442% | +715% | **all at 10x** | +698% (6 coins) |
| `0x0427e497f8075524fdc6a73ed6dfe1be958a2a56` | 21 | 71% | 397 | +164% | +149% | all at 10x | 1 test coin only |
| `0x0ec66d31aab7e387b2f4cda075e39f7a6a0f541b` | 22 | 40% | 60 | −14% | +4% | all at 2.5x to 3x, or skip | 2.5x +11%, 3x +9% (20 coins). It pulls right after the 3x half is sold, so half 3x / 10x loses. |
| `0xca24b6f1b00a7d336ef2984239daa836f17cb6c1` | 61 | 32% | 53 | −20% | +18% | sell at 30 s | +24% (58 coins) |
| `0x20e7a901982b0acbb8cccdc9148373673e2bf6b1` | 44 | 15% | 29 | −42% | −58% | **skip** | every exit loses |
| `0x0d1b5ebe6a2b516417c6cc6c8cfb89acdb3a9288`, `0x32e0fa527c66d74f4fbaad37c07044baddaf180d`, `0x7d59754dd21dd899030d023aa4d82ecf443b9591`, `0xf42695b2c9e940769d4ae46548f84b482eed6999` | 15 to 55 | 0 to 4% | 140 to 308 | −94% to −100% | −9% to +11% | skip | never hit 3x |

## 4. Safe size
- Data: wash.db swaps for crews A and B/C, 21 Sep to 4 Oct, with first funders from ffund.db.
- Outside wallet: not a crew wallet (crew_wallets.json, crew.db creators), not funded by a crew wallet or a money wallet, and not a repeat wallet.
- Test: is the coin pulled within 60 s after the buy?
- Expected: from the crew's own coin lifetimes, the chance that a coin still alive at that age is pulled in the next 60 s.
- Outside wallets that bought and held, coin age up to 300 s:

| Buy size (ETH) | About $ | Buys | Pulled within 60 s | Expected | How many times normal |
|---|---|---|---|---|---|
| under 0.002 | under $5 | 1,069 | 155 (14%) | 126 | **1.2x** |
| 0.002 to 0.0045 | $5 to $12 | 76 | 38 (50%) | 10 | 3.8x |
| 0.0045 to 0.0075 | $12 to $20 | 19 | 15 (79%) | 2.8 | **5.3x** |
| 0.0075 to 0.015 | $20 to $40 | 8 | 3 | 0.8 | 3.6x |
| 0.015 to 0.035 | $40 to $95 | 19 | 12 (63%) | 2.4 | 5.1x |
| 0.035 to 0.075 | $95 to $200 | 4 | 1 | 0.5 | 2.1x |
| 0.075 and up | $200+ | 4 | 4 (100%) | 0.4 | 11x |
| crew buys, any size | | 50,994 | 6,059 (12%) | 6,478 | **0.9x** |

- When the pull follows an outside buy, it comes about 15 s later (median; half of them 10 to 23 s).
- Example `0x90e2c652d8fe296dcb7fa38dff038c781fda2e23`, a $10 bot buyer:
  - In 11 of the 12 crew A coins checked, the pull came 4 to 18 s after its buy.
  - Coins it bought from other crews were mostly not pulled. So this looks like crew A's bot reacting to the buy, not something about the wallet.
- Four separate wallets that bought exactly 0.0074 ETH ($20) and held: all 4 pulled 12 to 49 s later.
- Counter-example `0x67dfa04186b8b0a5cdbaacd3d263e4a5625af989`: it bought $8 to $19 at 3 to 12 s and SOLD within 5 to 10 s. Only 2 of 11 of its coins were pulled within 60 s. A wallet that leaves before the bot reacts seems to be ignored.
- What it would cost: if every coin is pulled 15 s after our buy, the tagged A and B/C lane goes from +98% to −97% a coin (half 3x / 10x). At the 79% pull rate seen at $12 to $20, the expected result is negative.
- **Answer:** keep paper size at $20 for comparisons. Any real buy in A or B/C coins stays at 0.0015 to 0.002 ETH (about $4 to $5) until the live test shows more is safe.
- The earlier A1.6 found no pull trigger in crew `0x7028bab73f82dff14b199593abb76c8b6acb3aa9` (0 of 179 early outside buys pulled). Bigger tickets there are a separate question.
- Data limits:
  - Few outside buys in the $5 to $200 range (n 4 to 76), and only 2 weeks of data.
  - Few outside buys land exactly at 10 s.
  - The under-$5 group has many dust bots.

## 5. Timing check details (v2)
- Data: crews A and B/C, 21 Sep to 4 Oct. Each wallet's first buy in each coin, from 0 to 300 s after launch.
- Cells: crew × coin age (0 to 20, 20 to 60, 60 to 120, 120 to 300 s) × price multiple vs the 10 s price (under 1x, 1 to 2x, 2 to 3x, 3 to 5x, 5x and up).
- Expected = the dust (under 0.002 ETH) outside pull rate in the same cell. The crew's own same-size buys are a second control.

| Group | Buys | Pulled within 60 s | Matched dust gives | How many times more |
|---|---|---|---|---|
| $9 to $20, held | 68 | 41 | 10.2 | 4.0x |
| $20 to $95, held | 26 | 14 | 3.4 | 4.2x |
| $95+, held | 8 | 5 | 0.6 | 8.0x |
| $9+, any (incl. quick sellers) | 139 | 74 | 19.0 | 3.9x |
| $9 to $20, held, without `0x90e2c652d8fe296dcb7fa38dff038c781fda2e23` | 53 | 29 | 8.1 | 3.6x |

- Script: http://159.223.88.168/hx-c9ois/fw/a1code/t5.py

## 6. General sell rule details (v2)
- A crew's median pull uses only its coins that had ended before our entry, with at least 5 needed. If a crew has fewer than 5, use half at 3x / rest at 10x.

| Rule (buy at 10 s, $20 paper) | 30 Sep | 1 Oct | 2 Oct | 3 Oct | 4 Oct | Total (n) | Best coin removed |
|---|---|---|---|---|---|---|---|
| tag + crew pick, per-crew best fit (v1) | +327% | +216% | +101% | +70% | +178% | +175% (272) | +171% |
| **tag + crew pick, general pull-time rule** | +326% | +198% | +117% | +56% | +172% | **+169% (272)** | +165% |
| tag + crew pick, pull-time rule, A and B/C all at 2x | +339% | +142% | +60% | +35% | +57% | +117% (272) | +114% |
| tag + crew pick, everything all at 2x | +81% | +63% | +30% | +19% | +57% | +50% (272) | +49% |
| tag A or B/C, all at 2x | +84% | +54% | +55% | +17% | +57% | +50% (233) | +49% |
| HOT + tag + crew pick, general rule | +140% | +232% | +179% | +115% | +151% | +176% (192) | +171% |
| HOT + tag + crew pick, A and B/C all at 2x | +160% | +172% | +45% | +108% | +56% | +116% (192) | +112% |

- What the rule picked:
  - B/C `0x4757fc41278bd0f1b028c6dae71d381dc4558d98`: all at 10x on 86 coins, half 3x / 10x on 45.
  - A `0x92a0849f9b1e9e02363a4bd8d74a317fa6ddae92`: all at 10x on 76 coins, half 3x / 10x on 16.
  - `0x7028bab73f82dff14b199593abb76c8b6acb3aa9`: all at 10x.
  - `0x0ec66d31aab7e387b2f4cda075e39f7a6a0f541b`: all at 2x.
- Script: http://159.223.88.168/hx-c9ois/fw/a1code/t6.py

## Files
- Scripts: http://159.223.88.168/hx-c9ois/fw/a1code/ (common.py, sim.py, t1b.py tag, t23.py + t2r.py rules, t3.py exits, t4b.py / t4h.py size). They run on research-box in /root/fwa1.

First principles: about 80%. The tag and exits use only data known before each entry. The size finding is the biggest open point: n is small, and only a real small buy can prove whether the trigger is caused by the buy.

Version 3.5, 6 Oct 2026 14:48 Bangkok (v3.4 14:30) (v3.2 14:13, v3.1 14:10) (v3 14:05, v2 13:40, v1 12:50). Latest: http://159.223.88.168/hx-c9ois/fw/fw-analyst1-status.md