---
name: trading-edge-hub-3
description: Continuation of trading-edge-hub-2 (full at ~50KB on 4 Oct 2026). Ledger of tested trading insights from 4 Oct 2026 02:30 BKK on; read with [[trading-edge-hub]] and [[trading-edge-hub-2]].
sources: [cowork]
aliases: [edge hub 3, trading hub continuation 3]
---
# Basket rotation: best buy ranking and jumps per cycle (4 Oct 2026 02:30 BKK, cowork; it100/rank_test.py)
- [stated] Ben: wants a ranking among baskets that are BUY now ("this one is the highest recommended"), a rotation chart (RRG), and how many basket-to-basket jumps per cycle if holding one basket at a time; he thinks the BTC low is already in and the bull started.
- 2026-10-04 | cowork: which BUY first | Rule v4.7 book, 14 ranking methods vs 20 random orders: strongest 30 day move first beats random in both halves 19/20 (2023-26) and 15/20 (2019-22) with 3 slots, 11/20 and 14/20 with 1 slot; weakest first is worst in both eras (beats 1 to 3 of 20); r90, clenow, mix win one era only | rank_new.txt, rank_old.txt | LIVE (#1 BEST BUY badge, headline, Telegram rank)
- 2026-10-04 | cowork: jumps per bull (one basket at a time, live swap rule) | Bull 2 (Jan 2023 to Oct 2025): 33 buys (1.0 a month), median hold 11 days, cash 54% of days; Bull 1 (Mar 2020 to Nov 2021): 28 buys (1.4 a month), hold 12 days, cash 32%; 3 slots: 66 / 53 buys, worst drop -24% vs -30 to -32% with 1 slot | rank_test.py jumps | PROVEN (descriptive)
- 2026-10-04 | cowork: BTC cycle low check (Binance daily) | Top 6 Oct 2025 $126,200; low 1 Jul 2026 $57,800 (-54%, 9 months after the top; past bear lows -77% to -84% after ~12 months); 3 Oct 2026 close $84,873 (+47% from the low), above its 100, 200 and 365 day averages | data-api.binance.vision klines | FACT (a low is only confirmed later)
- [stated] Ben: is there always a strong basket to be in, or times when everything falls? Wants a scenario (stablecoins or BTC) so money is always in the strongest option or out.
- 2026-10-04 | cowork: how often nothing is on | No alt basket risk on 58% of days 2023-26 (2025 61%, 2026 71%) and 42% in 2019-22 (2022 76%); every alt basket lower 30 days later on ~1 day in 6 (17% / 20%), majors also lower on almost all of those | it100/idle_test.py | PROVEN (descriptive)
- 2026-10-04 | cowork: where idle money waits (3 slots, best buy first) | Cash +2256% / +4427% (worst -24%); BTC whenever above 100d +2963% / +5591% but worst -37% / -31%; BTC only while the Bitcoin card rule is on (30d breakout + 100d + 15d low + 20% trail) +2625% / +5157%, worst -28% / -27%, 2022 -3% like cash; 2025-26 half weaker than cash (+53 vs +97%); majors as parking mixed | it100/idle2.py | LIVE (page line: park in BTC while the Bitcoin card is risk on, else stablecoins; rotation.html c6f068442b9ebac3d82e24d96acdf14a)
# Basket method, ranking 50, windows, slots, cost, cycle (4 Oct 2026 07:30 BKK, cowork; raw http://159.223.88.168/hx-c9ois/edge/rotation-plan-2026-10-04/)
- 2026-10-04 | cowork: basket construction | Rebuilt both eras from daily CMC snapshots: 15 biggest coins equal weight, $25k volume, weekly refresh is best or near best; 5/10 coins far worse (+251/+671% vs +1849%; old +669/+1214% vs +4382%); 15 to 20 is the plateau; size weights worse; $250k/$1M volume floors worse; new-coin waiting 7/14 days neutral, 30/90 worse; monthly/daily refresh worse | basket/varbuild.py | PROVEN (keep live method)
- 2026-10-04 | cowork: number of baskets | Random subsets: medians 5 baskets ~+170% / ~+815%, 8 ~+480% / ~+1300%, 11 ~+1000% / ~+2700%, all ~+2256% / +4427%: more baskets = more chances | baskcount_*.txt | LIVE (ETH memes, ETH AI, BNB memes, BNB AI added; breadth gate now the 13 bench baskets)
- 2026-10-04 | cowork: 50 more ways to pick the first basket | None beat strongest 30 day move in both eras (19/20 and 15/20 random orders); closest % above 50d avg (19/18), 30d move per worst drop (19/15); late join into running baskets worse in both eras; weekly relative-strength swaps worse | rank50.py, latejoin.py | KILLED (keep r30)
- 2026-10-04 | cowork: breakout windows | Live 30d (20 when hot) +2256/+4427%; 10d +464/+336%; 60d +520/+1233%; blends of 10/20/30/60 or live+60 +878 to +1163% / +1357 to +2657% | windows.py | KILLED (keep live)
- 2026-10-04 | cowork: slots and coins | 3 baskets at a time best in 2023-26 (+2256% vs 1:+1674, 2:+1636, 4:+1288, 6:+571), worst drop -24%; 2 coins per basket trail owning all by 42 pts in the worst tenth (6 coins 26 pts) | kslots.py | PROVEN (6 positions = 2 coins x 3 baskets)
- 2026-10-04 | cowork: trade cost at ~$12k per coin | Aggregator quotes: big coins 0.1 to 0.7%, mid 1 to 2%, small 4 to 30%+ to buy; TON coins 10%+ on chain (NOT 36% on STON.fi) so use CEX; fees fomo 0.5%, BasedBot 1%, Jupiter 0.1%, 1inch 0%, Binance/Bybit 0.1% | Jupiter/Kyber/STON quotes | LIVE (pop-up cost + 2 stars)
- 2026-10-04 | cowork: BTC stretched vs alt basket buys | Old era: buys while BTC Mayer/z20/r90 in top third did better (+70.6% vs +38.6% avg); new era better in 2023-24, worse in 2025-26 chop; ETH/BTC up in the 30 days before a buy = worse in both eras (top third +5% vs +19%; old +28% vs +81%) | btcext.py | TESTING (ETH/BTC lead worth a rule test)
# Buy window, ETH/BTC, coins per basket, re-pick, Giver (4 Oct 2026 09:20 BKK, cowork; it100/ethbtc2.py, ethbtc3.py, pick/repick.py)
- [stated] Ben: confused by Pump.fun SELL vs NEXT; asked how long the buy window is and whether rotation must be faster than daily (rotate within hours); wants ETH/BTC tested more granular and in combos (baskets topping + others breaking out + ETH/BTC up), ETH memes as leveraged ETH; asked for the best cross-chain Telegram bot (sell 30% buttons).
- 2026-10-04 | cowork: buy window | Trade avg by entry delay after the BUY close: 0d +12.9%, 1d +10.8, 2d +10.3, 3d +7.3, 5d +5.9 (2023-26); 2019-22 flat to 3d then lower; book 1 day late +1246% vs +2256% (old +1660 vs +4427%, worst drop -46%); next close cheaper only ~41% of the time | ethbtc2.py | PROVEN (buy the morning of the BUY close; intraday entries untested, no hourly basket history)
- 2026-10-04 | cowork: ETH/BTC at BUY | ETH/BTC up 14/30/60d, above 50d, rising 50d or new 30d high = alt buys still positive but lower in both eras (e.g. up 30d +5.3% vs +18.3%; old +31.6 vs +55.4); cluster of BUYs + ETH/BTC above rising 50d +2.0% (old -10.5%) vs below 50d +18.7% (old +57.7%); ETH baskets: new era better when ETH/BTC up 14d (+15.2% n17 vs +1.4% n8), old era opposite; eth_memes daily beta to ETH 1.24 (old 0.71); book skips or half size on any ETH/BTC state cut returns in both eras | ethbtc2/3.py | KILLED as rule, LIVE as page info line (chains.json ethbtc)
- 2026-10-04 | cowork: others topping | 2+ other basket SELLs in the prior 10 days: buys +13.8% med +9.9% (old +91.6%, 14/14 wins) vs +12.7%; with a cluster of new BUYs +20.5% (old +73.5%); 1.5x size on it: old +7434% vs +4427%, new +2186% vs +2256% (B half worse) | ethbtc3.py | TESTING (not robust in 2025-26)
- 2026-10-04 | cowork: coins per basket (static, v4 spread pick) | 2 coins worst in both eras: new +10.6% avg, median -4.2%, 37% wins vs 4 coins +18.8%, 10 coins +27.3%; old 2 coins +49.9% vs 5 +57.9% | pick/repick.py | LIVE (page plan 3 baskets x 4 to 5 coins, 5 stars)
- 2026-10-04 | cowork: weekly re-pick while HOLD | Changes ~half the coins each week; better in the first half, worse in the second half of both eras (10 coins new +32.5/+27.2 vs +20.2/+34.5 static) | pick/repick.py | KILLED (keep the BUY day picks)
- 2026-10-04 | cowork: Hyperliquid basket | CMC hyperliquid-ecosystem has only HYPE plus 2 coins with real volume (KNTQ, PURR); no Telegram bot trades Hyperliquid spot | CMC listing | KILLED (too few coins)
- 2026-10-04 | cowork: trading bots | BasedBot 1% confirmed (docs.basedbot.app/welcome/platform-fees), 5 to 25% cashback, custom sell %, limit/TP/SL/trailing, no TON; Maestro 1% with 15 to 30% cashback covers SOL, Base, ETH, BSC, TON, HyperEVM, Robinhood, custom sell %, TP/SL/trailing, in-bot bridge (Oct 2023 router hack refunded); fomo ~0.5% but no ETH mainnet, no spot stops; Banana Gun 0.5% on ETH | subagent research | FACT (pick Maestro for cross-chain)
- 2026-10-04 | cowork: Giver substack review | 78 posts Sep 2024 to Sep 2026, ~23 with text (rest podcasts); macro/rates, special situations (HYPE, UNI, ASTER, BMNR/ETH), long BTC + ZEC from 1 Sep 2026, BTC over alts, stop ~74k, trigger ~82k; posts rarely, live calls in paid Telegram | Chrome read | LIVE (Giver card; Mac launchd com.kookwatch.giver hourly RSS to /giver.json, droplet IP is Cloudflare-blocked; stance in /giver_view.json written by hand)
# Real coin results, pick liquidity, RRG (4 Oct 2026 10:00 BKK, cowork; it100/coinlvl2.py, rrgbook.py)
- [stated] Ben: 4 of 5 Bonk.fun picks quoted 15 to 24% to buy at $12k; asked for real per-trade results with the coins (same morning, 1 day, 5 days late) and for 6 positions together; whether 07:00 BKK is the best decision time; what the RRG is for (his idea: lagging baskets catch up after a BTC move); corrected that fomo HAS Ethereum mainnet (App Store: Solana, Robinhood, BNB, Base, Ethereum, Monad).
- 2026-10-04 | cowork: pick costs | Quotes were real: AOL, DONT, MOMO, Ani were $0.2M to $0.5M caps with $1k to $9k daily volume (Jupiter impact 15 to 24% at $12k, 4 to 8% at $3k); only USELESS ($5.4M pool) was tradable | Jupiter + DexScreener | FACT
- 2026-10-04 | cowork: real coin results, live rule, 1% cost, biggest coins | Per trade 2023-26 / 2019-22: biggest 2 coins same morning +9.7% / +28.7% (median -4.8% / +4.3%), 1 day late +8.4% / +29.6%, 5 days late +5.4% / +28.7%; biggest 5 +6.0% / +31.5%; 3 basket book with 2 coins +2546% / +1637% (worst drop -27% / -28%), 1 day late +1853% / +672%, 5 days late +157% / +987%; index book +2256% / +4427%. Spread picks (incl. micro caps) were polluted by price glitches and untradable | coinlvl2.py | LIVE (picks v5)
- 2026-10-04 | cowork: correction | The 4 Oct "4 to 5 coins per basket" advice came from spread picks that include micro caps; among the biggest liquid coins 2 vs 5 makes little difference (2 better in 2023-26, 5 better in 2019-22) | coinlvl2.py | CORRECTED
- 2026-10-04 | cowork: RRG quadrant at BUY | Lagging basket + BTC up 14d: 2019-22 +64.8% (94% wins, n32), 2023-26 +21.3% but 2025-26 -1.7% (n8); leading + BTC up: 2023-26 +29.1% (74% wins) but 2019-22 weak; book: lagging-first ranking +1122% / +2346% vs +2256% / +4427%, skip leading BUYs when BTC flat +2348% / +3849% | rrgbook.py | KILLED (RRG is a picture only, folded on the page)
- 2026-10-04 | cowork: decision hour | 00:00 UTC (07:00 BKK) daily close is the only testable close; hourly basket moves since the close now logged to /opt/rotation/chains/hourly_levels.jsonl to test other hours later | chains.py | TESTING (needs ~2 months of data)
- 2026-10-04 | cowork: trading app | fomo ~0.5%, covers Solana, Base, Ethereum, BNB, Robinhood, Monad (not TON); no stop orders, not needed since exits are 07:00 close decisions; earlier "no ETH mainnet" (subagent) was wrong | App Store listing | FACT (fomo first, TON on Binance/Bybit)
# Pickers and winners vs crowd (4 Oct 2026 12:00 BKK, cowork; container /tmp/tgc, /tmp/cy, /tmp/ghost; doc claude/picker-study-2026-10-04.md)
- [stated] Ben: believes the key runner factors are not found yet; study chjokka (t.me/chjokkacalls), cryptic (t.me/crypticsden22), _0xghost_ subscriber posts, the 3 Unicorn wallets, then all big runners vs the crowd; stay flexible on the window. X API has 7.7M credits, use it carefully.
- 2026-10-04 | cowork: chjokka TG calls Jul 14 to Oct 3 | 35 volume-backed calls, peak after call (clean candles): 2x+ 17 vs 17.5 power-law, 5x+ 9 vs 7, 10x+ 5 vs 3.5; -50% came before 2x on 20 of 35; wallet book (gmgn): $1.12M turnover, net +$14k, PRM +$22k is the profit, NET -$13k on $912k late Sep | score2.py, books | FACT (no edge outside one lane)
- 2026-10-04 | cowork: cryptic TG calls | 39 calls: 2x+ 24 vs 19.5, 5x+ 14 vs 7.8, 10x+ 9 vs 3.9; -50% before 2x on 22 of 39; Sol pump wallet net +$163k on $388k, STONK +$128k and Agency +$73k (3 Oct) carry it | score2.py | FACT
- 2026-10-04 | cowork: where both callers' edge lives | Their Stonk.fun calls: 16, of which 12 hit 5x+ and 9 hit 10x+ (expected ~3 and ~1.6); all other calls: 61, 10x+ 4 (expected ~6) | scored2.json by launchpad | PROVEN on their record (Aug to Sep only, one lane)
- 2026-10-04 | cowork: Unicorn wallets entry context | U1/U2 big wins were 10 to 40 day old coins bought 50 to 75% under the old ATH at $67k to $700k (HOOKR 90x after, MOO 91x, ZAPS 33x, DELTA 56x) plus AI at $6.2M 64% off ATH (74x); U3 wins were under 1 day old near ATH (PROLOGUE, ORBIO); fresh launch buys mostly lost (USD, ABTC, CHANG, PVE, HOOK, CUM, DRILL) | books + pk.json | FACT (record was picked after the wins, forward record since 18 Sep is 3 buys, best 1.7x)
- 2026-10-04 | cowork: ghost method (395 sub posts 28 Jul to 4 Oct) | Narrative + team track record + token value capture (revenue, buybacks, burns) + TAM; joins TG to judge community; buys bottomed charts of good projects on low volume; sizes only on conviction; biggest wins STONK, NET, STONKBROKER, TIBBIR; says ez mode was Aug to mid Sep and avoids new pairs since | /tmp/ghost/posts.json | FACT (his own account, no wallet)
- 2026-10-04 | cowork: Pons V2 census, 401 coins with a volume-backed $1M | Base 5x 20.7%, 20x 4.2% (matches power law). No lift in both halves for: twitter, website, DexScreener paid before $1M (0.88), boost (0.59), dev holds or sold, dev prior $1M hit (n13, 0.37), age at $1M. Creator with more than 5 launches 5x lift 1.31/1.43 (n58), likely a launch app wallet, unexplained. Image dup count and CTO flag look positive but are measured today (copycats follow winners) so they leak | feat.py | KILLED
- 2026-10-04 | cowork: NEW LANE effect (strongest found) | Pons coins in the first 10 days after their quote lane's first $1M coin: 20x 10% vs 4.2% (lift 2.3, positive in both halves where present); lanes older than 21 days 20x lift 0.5 to 0.8. Same pad, same days (29 Aug to 8 Sep): USDG lane 12 of 29 hit 5x and 6 of 29 hit 20x vs ETH lane 19 of 96 and 3 of 96 | lane.py | TESTING (strong but one pad; USDG may also mean serious devs)
- 2026-10-04 | cowork: regime | Pons V2 fair census, 2 week cohorts: 5x 32% (12 Aug, n19), 23% (26 Aug, n239), 19% (9 Sep), 6% (23 Sep); pump.fun random sample 5x 5% overall vs Pons 21%; simple regime gauges (count of new $1M coins in 7 days, recent cohort hit rate) did not separate within halves | regime.py, solout.py | TESTING (time matters more than any coin trait found)
# Bull-only trade stats and rejoining uptrends (4 Oct 2026 10:30 BKK, cowork; it100/bulls.py, bulls2.py)
- [stated] Ben: NEXT should not only list baskets that are down (an uptrending basket under its 30 day high could also be bought); per-trade numbers looked too low; measure bull markets, not 2025-26; does not accept that the average trade should lose.
- 2026-10-04 | cowork: per trade in bull years, live rule | Basket index 2020-21: avg +50.7%, median +17.6%, 74% wins, n110, hold median 31 d; 2023-24: avg +19.9%, median +11.3%, 59% wins, n61, hold 19 d. Biggest 2 coins (1% cost) 2020-21 avg +33.2% / median +8.4% / 62% wins; 2023-24 avg +14.6% / median -3.4% / 41% wins (2023-24 gains sat in smaller coins). Trades of +30% or more are 26 to 43% of trades and make 80 to 94% of all gains. 2025 index avg +4.4%, 2022 -13.4% | bulls.py | PROVEN (descriptive)
- 2026-10-04 | cowork: coin count in bull years | 2 / 3 / 5 / 8 biggest, or volume floors $1M / $3M: 2023-24 avg +9.7 to +16.7%, 2020-21 +33 to +37%; no pick rule clearly better | bulls2.py | PROVEN (count does not matter among big coins)
- 2026-10-04 | cowork: rejoin baskets already in an uptrend at a new 30 day high close, not extended | Book +1608% vs +2256% (2023-26), +2955% vs +4427% (2019-22); only the 2023-24 half slightly better | bulls.py | KILLED (fresh BUY only; NEXT stays = risk-off baskets)
- 2026-10-04 | cowork: Giver stance from TacFellas (Telegram Desktop, read 4 Oct) | 1 Oct: added alt short basket (HYPE 89, SOL 118, ZEC 1395), "debasement false start" theme, notionally net short under 1x; 3 Oct: took off 70 to 80% of remaining BTC long at 86.5k, short alts outright, added crypto shorts on payroll miss, small gold short at 4200, Brent long at 100; risk window into US midterms | Telegram | FACT (page lamps: BTC short Bearish, 1 to 2 months Unclear, Alts Bearish)
# Methodology audit (4 Oct 2026 11:00 BKK, cowork; it100/audit1.py, audit2.py, exitgrid.py, smallc.py)
- [stated] Ben: "+10% in a couple of weeks from a hand-picked strong basket in a bull is not it"; asked for a full methodology review; wants small but liquid coins if the big moves are there.
- 2026-10-04 | cowork: data check | Coin prices (CMC snapshots) match Binance closes within 1 to 2% (BONK, PEPE, SHIB, FLOKI, MEME over 20 trades); coin moves line up with basket lines at lag 0 (corr 0.79 to 0.96, base_memes 0.26 = weak history); historical test pool for bnb_memes held multi-chain Base coins (BRETT, TOSHI) in 2024, live basket uses real BNB contracts | audit1.py + Binance klines | FACT (numbers are real)
- 2026-10-04 | cowork: where the money goes | Median trade peaks +33% (2023-24) / +46% (2020-21) but exits keep +11% / +20%; system 2023-24 ~11x vs holding every basket ~8x (median +689%) vs BTC +462%; 2020-21 ~45x vs ~17x vs +543% | audit2.py | PROVEN
- 2026-10-04 | cowork: 9 exits to keep more of the peak | Tighter trails (15% after +20%, 12% after +15%), looser (25%), 10 or 20 day low, sell 50% at +40%, thirds at +30%/+60%, no profit taking: every one lower book in both eras (e.g. thirds +701% / +1755% vs +2256% / +4427%) | exitgrid.py | KILLED (live exit stays best)
- 2026-10-04 | cowork: small but liquid coins (bull years, tiered cost) | 3 smallest with $1M+ volume +36.7% per trade vs 3 biggest +30.8% (2020-21) but +5.2% vs +10.2% (2023-24) | smallc.py | KILLED (not reliable)
- [stated] 4 Oct 2026 11:15 BKK, Ben: doubts the entries; buying coins under their 4h MA100 and MA200 is a terrible place to buy (B, AKE, UAI of BNB AI look awful now).
- 2026-10-04 | cowork: coin trend at the BUY day (daily 17 and 33 day averages as the 4h MA100/MA200 proxy, 5 biggest coins, 1% cost) | 2023-24 bull: above both +12.3% per trade, 45% wins (n171) vs below +(-6.9)%, 17% wins (n52); all 2023-26 above +4.7% vs +4.6% (wins 40% vs 28%); 2019-22 no difference (+36.2% vs +36.9% bull) | it100/coinma.py | LIVE (picks v5.1 skip coins under their 4h MA100 or MA200 at the BUY close, GeckoTerminal 4h candles of the deepest DexScreener pair; unknown history = not blocked)
- 2026-10-04 | cowork: BNB AI check | BUY close 12 Sep 07:00 BKK: B was +23% over its 4h MA100 and +26% over MA200 (AKE, UAI also above both); B then peaked +24% (15 Sep) and is -8% now, under both MAs; page says TOO LATE for new money | GeckoTerminal 4h | FACT
# ATH breakout: hunting 2x in 14 days (4 Oct 2026 12:00 BKK, cowork athbreak deep session; /tmp/ab ab_2x.py, ab_2xm.py; droplet /opt/abstudy flow3.py)
- [stated] Ben: a ~30% 2x chance is not good enough; wants a breakout pattern that leads to 2x within ~2 weeks.
- 2026-10-04 | cowork: 14d base | 1,469 breakouts with 14d data: 2x 43%, 2x before -50% 36% (micro 62/48%, small 51/40, medium 30/24, $3M+ 30 to 39%) | ab_2x.py | PROVEN (descriptive)
- 2026-10-04 | cowork: price model for 2x | Gradient boosting on 40 price/basket features, out of sample by time: top 10% hit 52% (late) / 63% (early) vs base 41/44%, AUC 0.58/0.63; fair Pons list late half top 20% 40% vs 36% (AUC 0.51); 2x before -50% top tenth 34 to 49% | ab_2xm.py | KILLED (no 60% setup from price)
- 2026-10-04 | cowork: wait for confirmation | Odds of 2x before -50% measured from the price at T+6/12/24/48/72h in 10 states (above/below old ATH, up 30/50/100%, at new high, never lost ATH, 30% off peak): all 30 to 45%, none above ~45% in both halves; a coin at any point doubles before halving about 1 in 3 | ab_2x.py states | PROVEN (price path alone is near a coin flip)
- 2026-10-04 | cowork: wallet flow at breakouts | rh_flow has only 15 usable breakouts (4 coins), snaps start 2 Oct; untestable today. RH full chain rebuild (every transfer) runs backward ~1.5h per day toward 3 Jul (at 21 Sep on 4 Oct) | flow3.py | OPEN (test buyers, first-time buyers, whales, sellers quitting when the rebuild reaches early Aug; follow-up scheduled 7 Oct)
# RH launch winners vs duds, walk-forward model (4 Oct 2026 12:30 BKK, cowork; /opt/hist feat.py, mine.py, pnlcheck.py, wincheck2.py)
- [stated] Ben: no live money yet; analyze the incoming data far and deep; focus on what the winners have in common vs coins that go nowhere.
- 2026-10-04 | cowork: single traits, 10x from a 60 s buy, both halves (12 days 22 Sep to 3 Oct, 19,369 coins, base 1.9%) | early buyer with a past 10x hit rate >= 12.5% in the first 30 s: 5.0x / 3.9x lift; hookless 0.01% fee pools (factory 0x3194e32622c5d860a0572c74edda99dcfd8fc827) 9.2% 10x (4.8x); pools.trade instant (factory 0x0000ffffbe8efe702c8703ae3477ff5de3d319c0) 9.2%; token minted before its pool opened 4.6% vs 1.0%; no single wallet dominating early ETH 2.4x/2.9x; WETH pair 2.3x/3.7x; creator with ANY earlier launch 0.2x/0.1x; longxyz hook pools 0.2x | mine.py | TESTING
- 2026-10-04 | cowork: walk-forward model (gradient boosting, trained only on earlier days, chain-only features known at 60 s, metadata removed because it leaks) | 7 test days: top 5% per day 577 buys, 10x 15.9% vs 1.9% (8.3x); top 3% 19.7%; every test day 12 to 25%. $200 each: paper +$144k, worst case (every non-10x trade loses 100%) +$63k on $115k in; positive every single day | picks60.json, pnlcheck.py | TESTING (best edge found; forward paper next)
- 2026-10-04 | cowork: are the 10x prints sellable | Of 112 model-picked 10x coins, 94 had real sellers take 0.5+ ETH (~$2,200) at 10x+ of our entry; median 236 trades and ~30 min spent above 10x; median 118 ETH bought after our entry | wincheck2.py | PROVEN (the earlier "real early buyers lose" check only saw a 30 min tape; tops come ~1 h in)
# Winners deep pass, 10 iterations (4 Oct 2026 12:30 BKK, cowork; /opt/hist mine2.py, featlib.py, aimodel.py; doc claude/launch-winners-study.md)
- [stated] Ben: defining the smart wallet is very good; go deeper on what winners have in common; 10 iterations; run the AI picker live, paper only.
- 2026-10-04 | cowork: SMART WALLET definition (grid of 54, walk-forward) | Wallet with 2+ earlier early buys (first 30 s) and 10%+ of them became 10x. Coin with 2+ such wallets buying in its first 60 s: 6.0x / 6.2x lift (n 1,233); 3+: 6.0x / 6.7x; only 1: 2.6x / 1.2x; none: 0.7x / 0.5x. Smart wallets arriving after 10 s 5.0x / 5.5x beat those in the first 5 s 3.8x / 3.2x; smart money 0.5+ ETH 5.3x / 6.1x vs under 0.1 ETH 3.4x / 1.4x | mine2.py | TESTING (strongest single trait)
- 2026-10-04 | cowork: crews | Pairs of early wallets that buy together a lot = bots (0.8x); pairs that were early TOGETHER in 2+ past 10x coins 2.5x / 2.2x | mine2.py | TESTING
- 2026-10-04 | cowork: first-minute flow | Best: 10 to 60 different buyers (2.0 to 2.9x), biggest single buy 0.3 to 3 ETH (2.4 to 2.9x), top buyer under 60% of ETH (2.6 to 3.7x), nobody sold yet (1.5x / 2.3x). Worst: first sell within 5 s 0.2x (n 8,539), 3 or fewer buyers 0.5x, one wallet 90%+ of ETH 0.5x, whale buy over 3 ETH 0.3 to 0.6x, 1% fee hookless pools 0.4x / 0.1x (biggest group, 5,962 coins); time of day no edge | mine2.py | TESTING
- 2026-10-04 | cowork: model v2 (adds smart wallets, crews, flow shape), walk-forward 7 days | top 5% per day: 577 buys, 10x 18.0% (9.4x), every day 16 to 20%; top 2%: 24.6%; worst case (every non-10x a total loss) +$86k on $115k | picks60v2.json | TESTING
- 2026-10-04 | cowork: AI picker LIVE PAPER since 4 Oct 12:24 BKK | aipick.service scores every RH launch 60 s after its first trade (same featlib code), paper buys above the model line (top 5% of the last unseen day, 0.814), $200, sell at 10x, stop -40%, 24 h; dRPC 100-block getLogs (free plan limit); retrained daily in the 08:00 report; watched by systemd, livewatch2 (120 s), super | /opt/hist/analysis/aipaper.db | LIVE (paper)
# Winners round 3, 10+ passes (4 Oct 2026 14:00 BKK, cowork; /opt/hist deep3.py, deep3_an.py, symfetch.py)
- [stated] Ben: 10 more iterations, every little aspect of the winners, outside the box.
- 2026-10-04 | cowork: INSIDER DISTRIBUTION (strongest new trait) | Minter (wallet that got the minted tokens) sent tokens to 3+ wallets before 60 s: 4 winners in 6,325 coins (0.0x both halves); 4+ wallet-to-wallet transfers 0.5x / 0.6x; none 2.7x / 2.2x. Now the model's top driver | deep3_an.py | TESTING
- 2026-10-04 | cowork: smart wallets selling | 2+ smart wallets and none sold by 60 s: 7.6x / 7.2x; 2+ smart but one already sold 4.9x / 4.7x | deep3_an.py | TESTING
- 2026-10-04 | cowork: HOT LAUNCH TYPE | Launch type (pool hook + fee) that had 5%+ 10x yesterday: 4.6x / 5.6x; under 1% yesterday: 0.3x / 0.2x; whole-market heat yesterday: no edge (agrees with the NEW LANE effect from the picker study) | deep3_an.py | TESTING
- 2026-10-04 | cowork: first buy and start price | First buy 0.01 to 0.1 ETH 2.1x / 3.2x; first buy over 2 ETH 0.3x / 0.7x; start cap 13.6 to 22.7 ETH (one launchpad default) 0.1x / 0.2x; round-ETH (human) buys, weekday, launches in the last 5 min: no stable edge | deep3_an.py | TESTING
- 2026-10-04 | cowork: winner path | Of 337 winners only 2.1% dipped 60%+ (below 0.4x) and 3.9% below 0.6x before 10x; 20.5% dipped below 0.9x; minutes from a 60 s buy to 10x: median 17, 75% 38, 90% 69 | deep3_an.py | PROVEN (descriptive: a -40% stop costs almost no winners)
- 2026-10-04 | cowork: exits on model picks (price only, 3% cost) | Trail 50% after 3x (stop 0.6) +256% per trade, trail 40% after 5x +253%, all out at 10x +236 to +247%, half at 3x/5x worse (+167 to +191%) | deep3_an.py | TESTING (trail after 3x slightly best)
- 2026-10-04 | cowork: model v3 (adds insiders, first buy, start price, smart selling, hot launch type), walk-forward 7 days | top 5% 553 buys, 10x 18.4% (9.3x base), avg trade +333% (trail exit), worst case +$87k on $111k; every day 14.5 to 24%; top 2% 24.8% | deep3_an.py | LIVE (paper picker upgraded to v3 4 Oct ~14:15 BKK)
# RH launches: entry cap and $1M to $5M targets (4 Oct 2026 14:45 BKK, cowork; /opt/hist mcap.py, mcan.py, mcan2.py, mcan3.py)
- [stated] Ben: at what market cap do we buy, how high do they go; wants targets of $1M, 2M, 3M, 4M, 5M and how good we are at finding those.
- 2026-10-04 | cowork: entry cap at the 60 s buy (ETH/WETH quoted, 12,890 coins, 22 Sep to 3 Oct) | median $44k (10% $10k, 90% $101k); 10x winners entered median $32k; their highest SELLABLE cap (price right after a real sell of 0.25+ ETH) median $412k, 75% $814k, 90% $1.17M | mcan.py | PROVEN (descriptive)
- 2026-10-04 | cowork: how rare big caps are | Coins bought under $100k that later hit sellable $0.5M: 1 in 63; $1M: 1 in 360 (32 coins in 12 days); $2M and $3M: 3 coins (1 in 3,841); $5M: 0. Most coins reaching $2M+ launched big ($0.6M to $1M already at 60 s), so the multiple is only ~2x | mcan3.py | PROVEN (descriptive; 14 day window, newest days shorter)
- 2026-10-04 | cowork: picking $1M coins from 10x below (walk-forward) | v3 10x model top 5% per day caught 19 of the 25 test-day coins that went from at most $100k to $1M sellable (4.5% of picks, 15x base, median 17x to target); own $1M model 13 of 25; $0.5M from at most $50k: 15 of 31 (3.5%, 10x base). $2M+ from 10x below: too few coins to train or test | mcan2.py | TESTING
- 2026-10-04 | cowork: exits for big targets on v3 top 5% (424 picks, 7 days, $200 each) | sell at 10x +279%; trail after 3x +268%; hold for $1M else zero +6%; half trail + half hold for $1M +137%; hold for $2M or $3M: 0 hits (-100%) | mcan3.py | KILLED (holding for $1M+ loses to selling at 10x; keep 10x exit)
# Winners round 4, 5 iterations (4 Oct 2026 16:00 BKK, cowork; /opt/hist deep4.py, deep4_an.py, deep4b.py)
- [stated] Ben: one more round, five iterations, patterns and leading signals of winners, outside the box, turn every stone.
- 2026-10-04 | cowork: NEW BUYERS IN THE SECOND HALF MINUTE | At least 2 wallets that first buy between 30 and 60 s: 2.9x / 3.0x lift (n 3,875); 0 to 1: 0.4x / 0.5x. Buyers arriving from 30 to 60 s at a slower count than 0 to 30 s (accel 0.7 to 1.0): 0.2 to 0.4x | deep4_an.py | TESTING
- 2026-10-04 | cowork: HUMAN vs BOT buy sizes | Buy sizes very uneven (coefficient of variation over 1.13): 1.9x / 2.5x; near-uniform sizes (under 0.49): 0.1 to 0.4x; under 60% unique sizes 0.4x / 0.4x; wallets buying twice in the first minute 0.7x / 0.5x vs none 1.6x / 2.8x | deep4_an.py | TESTING
- 2026-10-04 | cowork: CROWD ARRIVAL beats crowd quality | Coins where 10+ known losing traders (3+ past coins, got back under half) buy in the first minute: 3.8x / 4.2x (n 1,674); 10 to 20 veteran wallets (20+ past coins) 3.9x / 3.4x; 4 to 8 wallets with 2+ past doubles 2.7x / 2.9x; 40+ never-seen wallets 0 of 329 (sybil); profitable-trader count (ROI 1.3x+) only 1.8x / 3.2x for exactly one | deep4_an.py | TESTING (attention from active traders is the signal, not their skill)
- 2026-10-04 | cowork: DEV BUYS OWN COIN | Minter (wallet that got the supply) buys in the first minute: 0.1x / 0.3x (27 winners in 6,344); minter not buying 1.3x / 1.6x. First buyer holding 5 to 50% of first-minute ETH 1.8 to 4.3x; over 50% 0.7x / 1.1x | deep4_an.py | TESTING
- 2026-10-04 | cowork: WAIT AND CONFIRM is worse | v4 top 5% picks: 10x from a 120 s buy 12.4%, 180 s 10.2%, 300 s 6.1% vs 20 to 24% from 60 s; no price-confirmation band rescues it | deep4_an.py | KILLED (speed is part of the edge)
- 2026-10-04 | cowork: model v4 (v3 + 23 round-4 signals) walk-forward | top 1% 34.3% hit 10x (17x base), top 2% 27.5% (v3 24.8%), top 5% 17.4% (v3 18.6%); all signals at once overfit the wider list | deep4_an.py | TESTING (pruned and steadier variants in deep4b.py)
- 2026-10-04 | cowork: model variants (walk-forward, 10x from 60 s, sell at 10x) | v3 top 1% 32.4% / top 2% 24.8% / top 5% 18.6%; v4 pruned (v3 + 11 strong round-4 signals) 36.1% / 26.6% / 17.9%; steadier v4 25.9% / 24.3% / 17.4%; blend v3+v4 30.6% / 22.9% / 19.5% (+237% per trade). Ceiling at 60 s looks like ~18 to 20% for the top 5% | deep4b.py | TESTING (no big jump from more features)
- 2026-10-04 | cowork: SPEED IS THE EDGE (biggest round-4 finding) | Same features known at 30 s instead of 60 s, walk-forward, 10x from the fill price, sell at 10x stop -40%, 3% cost: top 5% fill at 30 s 28.1% hit (+352% per trade), 35 s 25.3%, 40 s 20.7% (+315%), 45 s 19.4%, 60 s 16.6% (+263%); top 1% fill at 40 s 38.0% (+450%); worst case top 5% at 40 s +$171k on $141k ($200 each, 707 picks, 7 days) | lat30.py | LIVE (paper)
- 2026-10-04 | cowork: second live paper picker | aipick30.service (same aipick.py, AIP_T=30, model aimodel30.pkl retrained daily by daily_analysis.py, line 0.798, results aipaper30.db, log aipick30.log); both pickers now book the HONEST fill = last price seen when the decision is made (not the 30/60 s print) and only count targets after that block; watched by systemd, livewatch2 and super | aipick.py | LIVE (paper, from 4 Oct 15:32 BKK)
# Winners round 5, 10 iterations, 100 ideas (4 Oct 2026 16:30 BKK, cowork; /opt/hist deep5.py, deep5_an.py, routers.py; doc claude/launch-ideas-100.md)
- [stated] Ben: 10 more iterations, 100 new out-of-the-box ideas, look at every dashboard's data (Bubblemaps, GMGN, Solscan, RH explorer, rug check sites) for massive edges.
- 2026-10-04 | cowork: WALLET IDENTITY BUG in rounds 1 to 4 | 55% of RH buys go PoolManager -> router or bot contract -> wallet inside one tx (1 Oct: 51,811 of 94,118); top "buyers" 0x1cbaf24d53fe930fce8eff149fa797d2611da149, 0x540ee09a131bbffe1224ff5d85b3e9650f3d6089, 0xe5e702641ea86f4ae6cc3cdaed2b886f976be044, 0xe08e4041458338bf3a0798eeeb86bbab32f727c0, 0x6f02324d20cc679d0e585290caa6b16bacbc0f77 are all contracts, so earlier smart-wallet / veteran counts partly counted routers. deep5 follows the token to its FINAL holder | routers.py | FACT (fixed in deep5)
- 2026-10-04 | cowork: APP MIX at 30 s (fill at 40 s, 10x base 2.4%) | Coins where 50%+ of buys go through router 0x1cbaf24d53fe930fce8eff149fa797d2611da149: 2 winners in 7,345 (0.0x both halves); 0x6f02324d20cc679d0e585290caa6b16bacbc0f77 50%+: 0.0x / 0.1x; 0x9982538f41f2ae29ddb9d3d9307010052984fdbb 50%+: 0.1x / 0.2x; 0x6aa80dbbed9ae5ab45fbf61f9644fada3b29326e 50%+: 2.0x / 2.1x; no known router at all 2.2x / 2.5x; 2 to 3 router apps 0.2x / 0.4x; 75%+ of buys held by bot contracts 0.3x / 0.2x | deep5_an.py | TESTING (strongest new filter: dead-router coins)
- 2026-10-04 | cowork: TRUE-IDENTITY wallets at 30 s | 5+ hot hands (wallets with a 10x hit in the last day): 6.8x / 3.8x; 5+ smart wallets 6.5x / 3.5x; 2+ smart and 1+ hot 5.0x / 3.4x; a smart wallet already sold by 30 s 5.4x / 5.1x (active traders, not a warning); 3+ wallets with past net profit 3.3x / 4.8x; 20 to 40 human (non-bot) buyers 2.9x / 2.2x; first buyer not a bot contract 1.6x / 2.4x | deep5_an.py | TESTING
- 2026-10-04 | cowork: first sell timing at 30 s | First sell inside 3 s 0.2x / 0.4x; first sell after 20 s 2.1x / 2.6x | deep5_an.py | TESTING
- 2026-10-04 | cowork: model v5 at 30 s (live 30 s features + app mix, bot share, true-identity smart/hot wallets, insiders, flow), walk-forward 8 days, 10x from a 40 s fill (base 2.5%) | live 30 s model top 1% 32.3% / top 2% 23.8% / top 5% 22.3%; v5 33.1% / 30.5% / 21.5%; blend 35.3% / 30.1% / 22.4%; top 2% worst case +$105k on $54k; every day 18 to 25% at top 5%; leans on minter_sends, app0 router share, start cap, top10 | deep5_an.py | TESTING (sharpens the top 2%, not the top 5%)
- 2026-10-04 | cowork: decay | v5 trained only on the first 4 days and never retrained: top 5% 17.4% vs 21.5% retrained daily | deep5_an.py | PROVEN (keep daily retraining)
- 2026-10-04 | cowork: hand rule, no model, at 30 s | 2+ true smart wallets, no minter sends, dev did not buy: 8.0% hit 10x (4.0x / 3.0x), 2,364 coins; plus hot launch type 8.2% (4.5x / 3.1x) | deep5_an.py | TESTING (simple, transparent fallback)
# Lane radar: landscape map + blind backtest (4 Oct 2026 17:30 BKK, cowork; container /tmp/lanes; docs claude/lane-radar-2026-10-04.md, claude/lane-map-v1.md)
- [stated] Ben: a lane is a new meta (PAID = paying X users from trading revenue was one); get a full overview of what exists, what each does and does NOT do, the direction of experimentation ("tokenize everything" on RH), so new or better things can be spotted within seconds; simulate seeing only up to 2 months ago, replay the thousands of new projects per day, measure speed and accuracy of new-lane detection and performance, test on unseen data, forward test after.
- 2026-10-04 | cowork: lane map v1 | 103 lanes (72 RH, 31 Sol/Base/BSC) with pioneer, first date, followers, limits, status | mem_rh.json, mem_other.json | LIVE (reference)
- 2026-10-04 | cowork: mechanical lane detector (new v4 hook or new quote token getting pools) | ~3 alerts a day; caught only pair.fund (hook day 1) and HOOD10; most projects use plain v4 or the Pons/Long hooks | mech.py | KILLED as standalone (needs text layer)
- 2026-10-04 | cowork: blind lane radar replay 4 Aug to 2 Oct (top 20 new RH coins a day by 6 h swaps, blind LLM judge with only that day's data + lanes known before that day + its own earlier finds) | caught 20 of 42 later-recognised lanes, median lag 0 days (11 early or same day); 118 alerts (~2 a day): 20 known lanes, 27 real new lanes outside the labels, 71 weak or noise; missed lanes mostly never reached top 20 by trades (PRM, Snowball, Nest, Abyss) or are social/hook lanes | match_out.json | TESTING
- 2026-10-04 | cowork: radar alerts as buys (entry at 6 h, clean candles, 14 days) | all alerts: 2x 38%, 5x 20%, 10x 11%, -50% before 2x 67%, median entry $217k, 23 of 117 later peaked at $5M+; the 27 "real new" alerts 10x 19%; random other top-20 coins: 2x 24%, 5x 9%, 10x 5%, 14 of 199 peaked $5M+, median entry $61k | outc.py | TESTING (alerts find coins that get big, but from a 6 h entry the odds are near power law; next: buy lane leaders inside the lane's first 10 days)
# Launch picker round 6: 50 new ideas + leftovers, live multi-model (4 Oct 2026 18:00 BKK, cowork; /opt/hist deep6.py, deep6_an.py, deep6_rules.py, aipick5.py, train5.py, retrain.sh)
- [stated] Ben: run all the good models in parallel (paper), test for a couple of days, keep adding models; another round with 50 new ideas; test the remaining ideas from the 100 list.
- 2026-10-04 | cowork: INDEPENDENT SMART MONEY (strongest single trait so far) | Smart wallets arriving through 3+ different apps/routers by 30 s: 7.2x / 5.3x (n 1,258); 2: 4.3x / 3.5x; none 0.5x / 0.4x. Followers (non-smart wallets buying within 5 s after a smart wallet) 1 to 5: 3.6x / 2.9x; smart wallet buying at 1 to 2x its usual size 4.5x / 3.0x | deep6_an.py | TESTING
- 2026-10-04 | cowork: price shape and pool at 30 s | 2/3+ of trades up-ticks 2.6x / 2.7x vs under 2/3: 0.1 to 0.3x; thin pool (under 2.4 ETH depth) 2.0x / 2.4x, 2.4 to 4.6 ETH 0.2 to 0.5x; a single sell of 0.16 to 0.23 ETH (bot exit size) 0.1x / 0.1x; buys per wallet over 1.7 0.3x / 0.3x; token with 2 pools 2.3x / 2.4x | deep6_an.py | TESTING
- 2026-10-04 | cowork: launch-type heat intraday | Same launch type 6%+ 10x rate in the last 1 to 4 h: 2.3x / 3.8x (10%+: 2.5x / 4.8x), under 1%: 0.4x / 0.2x; launch type quiet (5 or fewer launches last hour) 2.0x / 2.1x, busy (24+) 0.4x / 0.3x; under 2% false starts last hours 0.2x / 0.2x | deep6_an.py | TESTING
- 2026-10-04 | cowork: burn at launch | Tokens sent to a burn address before 30 s: 0.1% to 5% of supply 0 of 1,696; 5%+ 0.1x / 0.3x | deep6_an.py | TESTING (strong avoid)
- 2026-10-04 | cowork: no edge | fomo router share, ETH move 1 h / 24 h, US market hours, vanity contract, chain-wide 10x count last hour, launches last hour, second-buyer ratio, same-block share, pressure vs depth | deep6_an.py | KILLED
- 2026-10-04 | cowork: DECIDE AT 20 s (biggest round-6 finding) | feat5 features only, same coins: decided at 20 s with fill at 30 s top 1% 42.1% / top 2% 33.1% / top 5% 23.9% vs decided at 30 s fill 40 s 25.6% / 24.2% / 17.7% | deep6_an.py | LIVE (paper models f5_20s_top1/2/5 inside aipick5)
- 2026-10-04 | cowork: model v6 (v5 + 27 round-6 ideas), walk-forward 40 s fill | v5 36.8% / 28.3% / 21.5% (top 1/2/5%), v6 34.6% / 30.9% / 22.1%, blend 37.6% / 31.2% / 22.4%; leans on minter_sends, n_pools_token, app0, depth_eth | deep6_an.py | TESTING (small gain)
- 2026-10-04 | cowork: hand rules at 30 s | clean (no dead router, no burn, no insider, no dev buy) 5.2%; + smart via 2+ apps + launch type hot 3 h: 13.5% both halves (53 coins a day); clean + up-ticks + smart via 2+ apps 11.5% (101 a day) | deep6_rules.py | LIVE (paper rule_r6_clean_up_2apps)
- 2026-10-04 | cowork: killed leftovers | one model per launch type (no gain: 4.1 vs 4.1%, 11.6 vs 13.4%); scale-in half at 40 s + half at 60 s (+259% vs +280% all in at 40 s); max 1 to 2 trades per hour (total $ drops +$97k to +$137k vs +$181k); daily stop after 5 losers (no effect) | deep6_an.py | KILLED
- 2026-10-04 | cowork: LIVE PAPER ROUND 5 (aipick5.service, one fetcher, many models, aipaper5.db, decides at 20 s and 30 s) | models f5_20s_top1/top2/top5, v5_top1/top2/top5, blend_top2, base_top5_nodead, rule_smart2_clean, rule_r6_clean_up_2apps; picks filled later than 20 s past the stage are skipped ("late skip"); replayed Init logs can no longer reset an open paper trade (fixed in aipick.py too); retrain.sh daily via hist-retrain.timer 13:40 BKK (backups 16:40, 22:40), skips when no new day, heartbeat retrain.ok watched by super; 08:00 report shows every model | aipick5.py | LIVE (paper)
# Lane radar v2: first-minute judge of every RH coin (4 Oct 2026 20:20 BKK, cowork; container /tmp/lanes; droplet /opt/laneradar; doc claude/lane-radar-v2-2026-10-04.md)
- [stated] Ben: must work live in the first minute after launch; read the X bio and website for what the project does and how it is different; go beyond the day's top 20; build a full map of every crypto business model, service and product; say how a project gets categorised fast; rerun as v2; ignore Dune; focus on Robinhood.
- 2026-10-04 | cowork: crypto map | 459 types on 5 axes (product 289, mechanism 64, RWA 36, venue 32, meme 38) with definition, business model, token role, how to tell apart, limits, keywords | taxonomy.json | LIVE (reference)
- 2026-10-04 | cowork: what is knowable in minute one (23,490 RH coins with 300+ swaps in 30 min, 5 Jul to 2 Oct) | only 33% carry any project text (on-chain description 5,071, X bio 3,284, website 4,239, linked post 2,908); Pons V2 description()/socials() and Doppler tokenURI() are readable on chain at launch | texts2.json | FACT
- 2026-10-04 | cowork: math-only novelty (embedding distance to all earlier RH projects, all-but-top fix, taxonomy cell counts, new words) | best AUC 0.71; at 10 flags a day it finds no more lanes than random; texts all share "on Robinhood Chain / tokenized stocks" boilerplate; 30 min trade count alone AUC 0.83 | nov_eval*.py | KILLED (math cannot say "new")
- 2026-10-04 | cowork: embedding category (bge-small, 459 types) | product type top-1 ~37/110, category top-1 ~45 to 55% vs gold | evalfc.py | KILLED as the main categoriser (AI judge does category in the same call)
- 2026-10-04 | cowork: AI judge on the FULL first-minute stream (5,754 coins with text, 4 Aug to 2 Oct, blind, own memory, 2 nearest earlier projects) | 360 NEW/BETTER = 305 distinct ideas (~5 a day): 145 real, 108 weak, 52 lane-type; caught 24 of 42 known lanes: 8 on time (within 3 days), 10 idea seen 9 to 40 days BEFORE the lane took off (first try flopped, e.g. Snowball idea 6 Aug, BLAST 27 Aug), 6 late; 18 missed, mostly coins with no text at launch (pair.fund, PAIREX, Nest, Abyss, Pons hook flags) | match3.json | TESTING
- 2026-10-04 | cowork: first-minute NEW flags as buys (entry = first hourly close, 14 days, GeckoTerminal) | flags n247: 2x 42%, 5x 19%, 10x 11%, peak $1M+ 12%, $5M+ 5%, -50% before 2x 50%, median entry $43k; same-population random coins n260: 2x 28%, 5x 10%, 10x 3%, $1M+ 5%, $5M+ 0%; flags later matching a lane 10x 25%, $5M+ 16% (hindsight label) | outc3.py | CORRECTED in [[trading-edge-hub-4]] (full data + same-day baseline: 10x lift 1.6, not 3.7)
- 2026-10-04 | cowork: LIVE first-minute radar | laneradar.timer every minute: new pool -> metadata, tokenURI, X bio (cap 300 a day), website, code and hook template, first-use quote -> 3 nearest earlier RH projects -> AI judge (needs /opt/common/anthropic_key, Ben runs ~/kookwatch/set_anthropic_key.sh) -> public feed http://159.223.88.168/hx-c9ois/lanes/live.json; watched by super, stale.py cron, livewatch2, guardkit (drill: cron and livewatch2 each fixed a disabled timer in under 1 min) | radar.py | LIVE (paper, no alerts; judge off until key)
# Launch picker round 7 (4 Oct 2026 21:00 BKK, cowork; /opt/hist deep7.py, deep7_an.py, tokcode2.py, settle_open.py)
- [stated] Ben: check the remaining ideas, then 50 new ideas and test them.
- 2026-10-04 | cowork: INTRADAY HOT HANDS (strongest round-7 trait) | 8+ first-30 s buyers who were early in a coin that already hit 10x EARLIER TODAY: 8.0x / 5.4x (n 866); 4 to 8: 4.4x / 1.8x; none 0.5x / 0.6x | deep7_an.py | TESTING (needs intraday tracking to run live)
- 2026-10-04 | cowork: BOT FINGERPRINTS (avoid) | 3+ early buyers who were early in a dead-router coin earlier today: 1 winner in 6,290; token address ending 777: 0 of 1,880; median early buyer with 148+ buys today 0.2x / 0.1x; first 3 buyers all bots 0.3x; biggest buyer sold by 30 s 0.4x / 0.3x; 2+ loss sells by 30 s 0.4x / 0.2x; 39%+ identical ETH sizes 0.5x / 0.4x; more sells than buys but price up 0.1x; 2+ same-type launches in the last 60 s 0.2x | deep7_an.py | TESTING
- 2026-10-04 | cowork: good signs at 30 s | median human buy 0.01 to 0.054 ETH 2.6x / 2.3x; first buyer past 10x hit rate 3%+ 2.4x / 2.1x; early buyers who usually hold 2+ min 1.9x / 4.0x; later buys 1.7x+ bigger 1.7x / 2.1x; pool opened 0.4 to 4 s before first trade 2.0x / 1.7x | deep7_an.py | TESTING
- 2026-10-04 | cowork: no edge | bundle tx share, whale share, retail ticket share, preset sizes, vanity buyers, rotation buyers, spray buyers, curvature, time to 10th buyer, hook popularity | deep7_an.py | KILLED
- 2026-10-04 | cowork: leak caught | first deep7 model run showed 98.5% top 1% because x5_pk30 / x5_x30 slipped into the feature list; fixed and rerun; round 5/6 analyses and train5.py excluded them | deep7_an.py | CORRECTED
- 2026-10-04 | cowork: live picker ops | aipick5 fell behind (73 to 174 s late) under dRPC 429s; fixed with parallel incremental transfer pulls, heartbeat during restore (watchdogs restarted it in a loop), lastb so restarts refetch only new blocks, lost books settled from the day DB (settle_open.py, *_settled). Alchemy RH refuses eth_getLogs ranges (400) | aipick.py, aipick5_main.py | LIVE
- 2026-10-04 | cowork: DECIDE AT 10 s (round 7) | same feat5 features, same coins: 10 s decision / 20 s fill top 1% 45.9%, top 2% 37.2%, top 5% 27.4% vs 20 s 42.1 / 33.1 / 23.9% vs 30 s 25.6 / 24.2 / 17.7% | deep7_an.py | LIVE (paper f5_10s_top1/2/5 in aipick5 from 4 Oct 20:32 BKK, first fill 11 s after first trade)
- 2026-10-04 | cowork: RANK BY EXPECTED RESULT (round 7) | regression on the trade result (sell at 10x, stop -40%, else 6 h price) instead of 10x probability, live v5 features, walk-forward: top 5% +396% per trade (median +364%) vs +277% (median +103%), every day +305% to +544%; top 1% +549% vs +388%; with all round 5 to 7 features top 5% +409% | evtest5.py, deep7_an.py | LIVE (paper ev_top1/2/5 in aipick5; median result needs live confirmation because non-exit trades are valued at the last 6 h print)
- 2026-10-04 | cowork: model v7 (all rounds), 40 s fill | top 1% 39.1%, top 2% 32.7%, top 5% 22.9% (v5 36.8 / 28.3 / 21.5); blend top 1% 41.4%; 5x-trained model fewer 10x (13.1%) but more money per trade (+397%); money per trade falls steadily with rank (top 1% +426%, 1 to 2% +309%, 2 to 5% +212%, 5 to 10% +169%) | deep7_an.py | TESTING
- 2026-10-04 | cowork: bug | after the lastb column was added, positional INSERTs failed in all three pickers for ~25 min (4 Oct ~20:16 to 20:40 BKK, picks in that window lost); fixed with explicit column lists | aipick.py | FIXED