---
name: trading-edge-hub-6
description: Trading edge hub ledger part 6 (continues trading-edge-hub-5, which is at its cap): from 5 Oct 2026 ~22:30 BKK. Read with [[trading-edge-hub]] before any buy/sell call.
sources: [cowork]
---
# Trading edge hub, part 6 (follows [[trading-edge-hub-5]])

## 5 Oct 2026 ~22:30 BKK: crews outside the pond (review session; doc v2 http://159.223.88.168/hx-c9ois/brx/pond-crews.md)
- [stated] Ben: crews come and go, so every new deployer must be checked against old crews or tagged new; asked what share of launches come from crews; approved tracing the non-pond launches.
- 2026-10-05 | share of launches from crews (21 Sep..3 Oct, 14,158 launches) | 58% overall: pond 99% (11 crews), non-pond 54% (6,843 launches in 255 crews of 10+); crews make 59% of all sellable $500k runners (163 of 277); new crews 2 to 7 per day | hop3.py (container /root/pp), ~390k paid dRPC calls, never merged through 182 exchange/relay wallets (nonce >2,000) | FACT
- 2026-10-05 | non-pond crews graded week 1, tested week 2 (10x from 5 s price before pull, print-based) | good crews (25%+) 30.2% (n361), all crews 17.0%, bad crews 14.0%, no crew 5.9%; holds inside fee types (0.05% good 34.2% vs 19.6%; 1% good 30.2% vs bad 20.0%) | STRONG (13 days)
- Proposal to Price Predict: live "who funded whom" map of every new wallet that launches a coin (stop at exchange wallets), instant crew lookup at launch, daily crew grades, new crew = small size until ~20 launches graded | PROPOSED

## 5 Oct 2026 ~18:50 BKK: hook family bug + v4 census (Price Predict session)
- 2026-10-05 | BUG FOUND AND FIXED: hook family fingerprints across sessions are md5 of the "0x..." hex string from eth_getCode (bundle 341_6de2bb44, campaign 5059_7adb2ec7, junk 25533_3080c1dd); gate5.py hashed the raw bytes (same hooks read 341_02c6f0f1, 5059_9ede3104), so from 16:48 to 18:50 BKK the campaign/junk kill and bundle -> bundle_flip routing never fired (serial creator and 777 kills did: 29 and 17 skips in the last 3,000 scored) | gate5.py fixed, aipick5/aipick5b restarted, verified on 2 hooks | CORRECTED
- 2026-10-05 | whole-chain v4 census (liqlock.db, 966,132 pools, 22.9M ModifyLiquidity, 5.6M LP NFT transfers): no hook 636k pools, junk factory family 195k, bundle family 636 hooks (one hook per coin), campaign family 292 hooks | FACT

## 5 Oct 2026 19:42 BKK: first real-money round trip via GMGN (supervisor, live pilot)
- [stated] Ben: run the $1 live test in the supervisor session; Ben places trades himself (by hand, later he switches the bot on); new GMGN API key "firstmin bot" made 19:3x BKK (trading on, Robinhood wallet, trusted IP 159.223.88.168), applied on Mac from clipboard, read calls verified.
- GMGN test wallet (EVM, all chains): 0x0184ef9b23476ed84107cc5e15485a4a7ee6b286, 0.0562 ETH on Robinhood Chain at 19:40 BKK.
- 2026-10-05 | round trip 1, MEME 0x385f4f8ae47651ce5f58f5265395a669f8281e18, buy 0.00037 ETH ($1.00) 19:41:36, sell 100% 19:42:04 BKK, both via GMGN router 0x65050a9b7e5075a2ba5ced7b1b64ee66262c40dc, both status ok | cost $0.068 of $1.016 = 6.7%: gas ~$0.012 per tx (2.3% at $1, fixed cost, shrinks with size) + 4.4% proportional (GMGN 1% each way, pool fees ~0.3% each way, rest price move) | tx 0xbc829a3dcc3a527cdb8c73234ef6356abee2fce66f92a771c780b59cec7ff341 / 0x077dd1116fa15017a72ce086ec80188b45cce072bf39093fea0a6d8551a096ca | FACT (n1, liquid coin)
- GMGN free tier: our Mac IP got RATE_LIMIT_BANNED again ~19:38 BKK (other sessions may share the IP); bot must call GMGN only on picks.
- 2026-10-05 19:45 BKK | round trip 2 FAILED: supervisor picked AOTTR 0x52c4c365cf5317c89c817cf2516141613d38d08c by volume alone; it trades in a campaign-family pool (hook 0x2aa7a1aa52a0099286f14bbc9748dd7ee3448a80, WETH 0.3%, hook address ends a80 like every campaign hook) = buy yes, sell no; GMGN showed "Possible honeypot" on the page and the sell quote said ~91% loss; $1 lost (321,623 AOTTR stuck) | RULE (mechanical, every buy, hand or bot): before suggesting or sending any buy, run family_check.py (campaign 5059_7adb2ec7, junk, serial, 777) AND read GMGN's honeypot flag from its token/security data; either flag = NO BUY | FACT
- GMGN rate limit: the Mac IP hit RATE_LIMIT_EXCEEDED on single calls at 19:46 and 19:47 BKK, so something else on that IP is spending the free quota; the live bot must run from the droplet (trusted IP) with its own budget.
- 2026-10-05 | best Robinhood crews 20 Sep..3 Oct (sellable $500k+ count): Crew A root 0xcb7277454416d90e7fe8799cd300744d5797de03 62 of 281 (22.1%, pond 98%), Crew B 0x5bd9c2026df3d994fe1a0f051ac5992dfee85999 32 of 200 (16.0%, quiet since 29 Sep), Crew C 0x64a20b2a553c12960847d49fc381741a981de231 19 of 136 (14.0%); non-pond Crew F 0x05c8b2d29a7bc2a5e9584f128f2ac3f8a7737bce top coin Taphouse $4.69M; DexScreener drops pulled pond coins (71 of 80 top coins "not found") | doc claude/robinhood-crews-top-coins-2026-10-05.md | FACT
- 2026-10-05 ~20:05 BKK | [stated] Ben approved a separate $6/month server for the live trading bot. Created DigitalOcean droplet "firstmin-trader" 178.128.20.136 (SGP1, 1 vCPU / 1 GB, Ubuntu 24.04, SSH key kookwatch-deploy = ~/kookwatch/id_kookwatch on Mac, ufw on, SSH only). Node 22 at /opt/node-v22.11.0-linux-x64 (gmgn-cli 1.6.2 fails on Ubuntu node 18: "File is not defined"), gmgn-cli 1.6.2 at /usr/local/bin, GMGN config copied to /root/.config/gmgn (chmod 600), read calls verified. Pending: Ben sets the API key Trusted IP to 178.128.20.136.
- 2026-10-05 ~20:10 BKK | built /opt/fm_gate.py on firstmin-trader (Mac copy ~/kookwatch/fm_gate.py): PASS only if GMGN `token security` is_honeypot is False AND is_show_alert false AND buy/sell tax <= 5% AND (v4 pool) hook family not campaign 5059_7adb2ec7 / junk 25533_3080c1dd / bundle 341_6de2bb44 AND address not ...777; any error or missing field = BLOCK | verified: AOTTR BLOCK (honeypot + campaign), SPYDERNET PASS | FACT. NOTE: MEME 0x385f4f8ae47651ce5f58f5265395a669f8281e18 main pool hook 0x4e3468951d49f2eea976ed0d6e75ffcb44a9a544 is family 25533_3080c1dd, so the "junk factory" family is a big launchpad hook (likely long.xyz) that also hosts real coins; the junk label is only valid as a launch-time base rate.
- 2026-10-05 | round trip 3, SPYDERNET 0x91f254a75263eb9379b8fa065e8f5d92cde9602a (Pons graduated pool, hook 0xe5e702641ea86f4ae6cc3cdaed2b886f976be044, 24 min old), buy 0.00037 ETH 20:12:55, sell 100% 20:13:32 BKK via GMGN router, both ok | net +0.0000011 ETH (+0.3%) after gas ~$0.027 total; price moved up during the 37 s hold | tx 0xdf7045d40f62d04e2e52ac1bb5faeb6d9a1228c5a1c521be84e5fb139d0b982f / 0x01318c70f7495b0287ef10ee84f6b98bbb91c03d42d87aeec4ef33d955e407a2 | FACT: GMGN trades Pons graduated pools fine

## 5 Oct 2026 ~20:30 BKK: replay with today's knowledge + data for other sessions (Price Predict session)
- [stated] Ben: how would we have done over the last days with the pull and farm knowledge; test the 10 s model.
- 2026-10-05 | walk-forward replay 18 Sep..4 Oct, 30 s model, honest exits (pulled pool = 0, whole-chain pulls from liqlock.db), $200 per trade | top 1% (n309): v1 +$583 per trade, 10x 34%; top 5% (n1,519): v1 +$281, 10x 19%, positive on 17 of 17 days; random 3% of launches -$84; old paper pricing claimed +$946 / +$579; kill list (777 + serial only, family kill broken by the md5 bug) +$600 / +$290 | knowbt.py, analysis/knowbt_T30.txt | FACT (rerun with fixed family kill and T=10/20 running)
- 2026-10-05 | Pons graduations booked as their own venue in venuepick (hook 0xe5e702641ea86f4ae6cc3cdaed2b886f976be044): vx_all (v1) + vx_pg_hold (stop -50%, half at 10x, trail 50%, max 7 d) | LIVE (paper)
- 2026-10-05 | data for other sessions: http://159.223.88.168/hx-c9ois/hx/pp-data-status.md (liqlock.db liq/lock/poolfam/bundle_sellers, chainlogs, crew.db live crew map) | LIVE (liq build finishing)
- 2026-10-05 | crew map first sample: pond creators funded through 0xf341c935599822f1bc1ec29e27606724fbbf6595 in chains of fresh wallets 13+ hops deep | CONFIRMED; crew share and grades pending backfill
- 2026-10-05 20:30 BKK | POND IS WASH (review session): crews A/B/C (617 coins, 392k swaps, all 3,300+ trader wallets traced): wallets first funded by one crew distributor made ~97.6% of trades and ~100% of buy ETH (A army 1,467 wallets from 0xbae538011874261e3fb4ca6aa03cacb2766c4d81; B+C army from 0x79411dbcbcf0e5b27cd222f24e082fc0bbf32f16, so B and C = same operator); proof: LLAMA creator 0x70905c0d34ff05b5c4f73314dfd16c6aed438eff took prior pull ETH then sent 20 ETH to the A distributor; armies never cross; outsiders put in only 13/6/3 ETH vs wash volume $118.5M/$61.9M/$46.6M; pull is scripted (buy speed at pull = at top, pull 2.5 s after last trade); no X callers; pulled coins drop off DexScreener. Pond "runner" caps are crew prints, not demand; a pond replay profit means taking the crew's own ETH (real money so far ~22 ETH total, so size is untested). Both distributors still active 5 Oct 20:24 BKK | doc claude/robinhood-crews-top-coins-2026-10-05.md | FACT (CORRECTS pond runner reads)

## 5 Oct 2026 ~20:35 BKK: phase 2 session reads PP data status (read only)
- 2026-10-05 20:28 BKK | liqlock.db state: poolfam 966,132 rows and bsell 162,961 rows ready; `lock` empty and `liq` not built yet (derived build still running); chain/b*.db all-venue swaps ~1 day | FACT (re-check scheduled 6 Oct 06:30 BKK)
- 2026-10-05 | bundle sellers after 30 min (bsell, 677 bundle pools) | only 199 pools have any sell after 30 min (the rest are dead by then); 91,517 seller wallets, 99.9% of ETH out from wallets that sell in only one pool (fresh wallets per coin); 6 top bundle winners: 316 to 989 seller wallets, 5 to 254 ETH out, launch-tx recipients (~94 wallets per coin) took 0% to 73% (median ~26%) of that ETH; 6 bundle losers: 0 sells after 30 min | bl.py (droplet /tmp) | FACT: bundle launch wallets are the main sellers into a run, so sell at target, never hold a bundle past its run
- 2026-10-05 ~21:00 BKK | POND CREWS PULL ON REAL MONEY (review session): after an outside buy of 0.1+ ETH (~$270) the pull came within 60 s in 75%/75%/100% of cases (crews A/B/C, n 8/8/3, usually 5 to 14 s), vs 2.7% to 4.9% after army buys of the same size; 18 of 19 such outsiders lost everything (largest 2.479 ETH, pulled 6 s later); small buys 0.01 to 0.05 ETH mostly tolerated; best outside profit ever 0.215 ETH. Crew A cost ~1.3 ETH gas for 13 days (221k swaps at ~0.0000055 ETH), own pool fees return to them, profit ~7.6 ETH outsider loss minus gas (~$17k). RULE: do not buy pond (fee 0.01% ETH/WETH, 10 ETH start) coins with $270+; pond replays that assume $200 fills at printed prices are not reachable | doc claude/robinhood-crews-top-coins-2026-10-05.md | FACT (small n on the trigger, direction very strong). CORRECTS "catch a 10x with $100 to $500".
- 2026-10-05 ~21:20 BKK | SMALL-TICKET POND IDEA (paper, review session): buy ~$20 (below the ~0.05 ETH pull trigger) at 10 s after first trade on crew A/B/C coins, sell at 10x: 10x reached in 33%/26%/26% of coins (A/B/C), paper return +227%/+155%/+157% per coin before costs; 10x levels last long (median 32 min at/above 10x, q10 90 s). Real small outsiders (<=0.02 ETH, n4,784) lost 45% overall: 18% saw 10x after their entry but only 4% took 5x+, 47% never sold. ~23 crew A + ~26 crew B/C coins a day. Untested live: fills, GMGN 1% each way, and whether the crew starts pulling on one repeat wallet or on outside sells | analysis /root/pp/crew/ | PAPER, needs a $5 to $20 live test
- 2026-10-05 21:10 BKK | [stated] Ben asked for the pond small-buy test plan. Paper on ALL 1,402 pond coins (entry first trade +10 s): 3x hit 58%, 10x hit 20%; ladder half at 3x + half at 10x about +89% per coin before fees; weakest big crew 0xae3d565892c79fa8fc9243b380283dd419e28ec9 only 7% 10x. Plan: phase 0 paper live 2 h, phase 1 $5 x 10 coins, phase 2 $20 x 30 coins, every pond coin in a row, stop at $150 loss or 2 pulls within 15 s of our buy in 10 coins | plan http://159.223.88.168/hx-c9ois/brx/pond-smallbuy-test-plan.md (md5 f72607ca799a7d3d895555440e5cdc2c), project claude/pond-smallbuy-test-plan-2026-10-05.md | PLAN, not run
- 2026-10-05 22:00 BKK | WASH IS NOT ONLY POND (review session, 20% swap sample): crews 0x0424cab581eb1a66da3323d0452cff0872b0fc1f (fee 0.05%, "80% 10x") and 0xbce31c2e43027d3c926d502ceb9846053fb68ed1: wallets trading 20%+ of the crew's coins make 97% of trades; G's buyers funded by faucet 0x70b4a4c247737abdab6da8620e50ec1e862a32b1 (0.001 ETH to fresh wallets every few minutes). Crew 0x05c8b2d29a7bc2a5e9584f128f2ac3f8a7737bce has a broader crowd (repeat wallets 22% of trades) but 87% of its top 60 buyers were first funded by 0xa7bf3946d0551af07864b54a0854e85a24aeaaac = unproven. RULE: repeat-wallet share 80%+ = wash coin = only $5 to $20 with sell orders, never normal size, whatever picker chose it. Test plan v2 uses the Price Predict crew tag and buys only crews with proven paper edge (A, B/C) | http://159.223.88.168/hx-c9ois/brx/pond-smallbuy-test-plan.md (md5 2766e3573e604b912aac40cecdd6262b) | FACT + PLAN
- 2026-10-05 22:15 BKK | [stated] Ben asked for a detailed Price Predict instruction. Brief published: wash definitions (repeat wallet = trades 20%+ of a crew's coins in 7 days; WASH = crew or coin wash score 80%+), distributor seed list, two lanes (WASH max 0.0074 ETH with 3x/10x sell orders; normal size only on REAL coins with 20%+ outside buy ETH), 14-day re-score of every picker split by lane, acceptance checks | http://159.223.88.168/hx-c9ois/brx/pp-wash-brief.md (md5 1d741545dcfb54e181fdcc83dadf6a9d), project claude/pp-wash-brief-2026-10-05.md | SENT TO PP via Ben
- 2026-10-05 22:30 BKK | [stated] Ben asked for the final trade plan. Lane 1: crew A and B/C coins only (Price Predict tag), buy within 30 s of first trade, $5 x 10 then $20 x 30, sell 50% at 3x and rest at 10x, stop at $150 loss or 2 pulls within 15 s of our buy in 10 coins. Lane 2: normal size only on REAL coins (wash under 80%, 20%+ outside buy ETH), paper until the PP 14 day re-score is positive. Prereqs: GMGN Trusted IP 178.128.20.136, ~0.25 ETH in 0x0184ef9b23476ed84107cc5e15485a4a7ee6b286, PP tags live | http://159.223.88.168/hx-c9ois/brx/final-trade-plan.md (md5 5d4396795dc5247031d8f7d682b3ab35) | PLAN

## 5 Oct 2026 ~21:20 BKK: Price Predict builds the wash brief (PP session)
- [stated] Ben: execute pp-wash-brief; set up redundant monitoring and fixing for everything new; crew map done by morning.
- 2026-10-05 | wash.py (no RPC: day files swaps + token transfers, trader = end of the transfer chain in the tx) | first 3-day test: of the 15 biggest crews by coin count, 12 are 93% to 100% wash by swap count; 2 pond crews (0xecc83460b8e89a4fd1ccfb71753b135c9559dc9c 97%, 0x59bd1dae455c3b481a0a195bada606ad604cd0d7 93%); 1 crew mostly outside money (0x2c2572da570646b402b9b9e625aa102803904a49, 19% wash, 163 ETH outside buys) | analysis/wash_rescore.txt after the overnight run | EARLY (crew map 4,000 of 23,252 creators, pulls not joined yet)
- 2026-10-05 | live tags in the paper picker: every scored launch carries _crew, _wash (WASH/REAL), _lane, _max_eth (0.0074 on WASH) | LIVE, logging only, no buy change until Ben approves the re-score
- 2026-10-05 | BUG FOUND AND FIXED: aipick5.service runs the BUILT file aipick5.py (mk_aipick5.py = aipick.py head + aipick5_main.py); edits to aipick5_main.py after 16:48 BKK (rand_gated book, crew tags) were not live until the rebuild at 21:04 BKK | CORRECTED (rand_gated results start 21:04)

## 5 Oct 2026 ~21:45 BKK: daily learners sped up where it helps (PP session)
- [stated] Ben: find other daily learners that would gain from hourly like the picker retrain did; set them hourly only where it makes sense; redundant monitoring and fixing.
- 2026-10-05 | picker model: hourtrain.py every 2 h at :50 writes a SHADOW hourly model aimodel5h.pkl (daily offline rows + today's live scored rows labelled from chain, peak 10x within 2 h); aipick5/5b log its picks as control books v5_/ev_/blend_/base_*_h (not money book) | basis retrainfreq: every 2 h top5 10x 30.2% vs daily 26.3%, better 7 of 10 days | SHADOW, promote only if _h books beat daily on honest results (Ben decides). First model expected ~01:50 to 03:50 BKK (needs 100 labelled live rows; ~74 scored rows an hour)
- 2026-10-05 | serial creator kill list: was never scheduled (stale file), now hourly at :12 and counts today's launches from crew.db: 1,206 -> 1,270 serial creators | LIVE
- 2026-10-05 | xsafe mine hourly at :13 (was daily); drift stays daily | LIVE
- 2026-10-05 | wash.py + crewgrade.py: rerun whenever a new complete day lands (nightjobs step 8), daily 06:31 BKK kept as backup | LIVE
- Kept daily on purpose (labels need days, faster adds nothing): deptab (30 d labels), soltrain (10 d), brain train (3 d), metatest, ownscan, runner lab.
- Monitoring: sentinel C20 (hourtrain.ok + serial freshness: stale -> run, alert only if long stale), pdguard re-adds the cron lines and runs stale jobs, guardkit restores pdguard/nightjobs/hourtrain crons, healthall lists them, pickguard pauses hourtrain in priority mode. healthall ALL OK 21:40 BKK.
- 2026-10-05 22:00 BKK | [stated] Ben asked: real profit and daily ceiling given crews. Money book today 07:35..21:07 BKK: 46 coins, +$242/trade paper ($200, honest exits). Split by on-chain buyer check (/tmp/mbwash.py, paid dRPC, buy ETH from wallets seen in 3+ of the 46 coins): WASH-like (60%+ repeat) n25 +$283/trade at printed prices but only $20 reachable (pull trigger ~0.1 ETH, ETH $2,704); REAL-like n21 +$241/trade, best coin removed +$163, best 2 removed +$86 (wins come from 0xd4ea1f3070595fafa1af0e85d82656c77a78a118 704 buyers 176 ETH, 0x97d4c3c1b72aebb599b70002e9109b88248d51e3 622 buyers 278 ETH). Caveat: fresh-per-coin wash wallets read as REAL in this test; 3 of the REAL winners sit in 0.01% fee pools. Estimate: wash lane ceiling ~$500 to $1,000 a day ($20 x ~40 coins, real small outsiders lost 45%); real lane ~$3k a day paper (37 trades x ~$86 after best 2 removed, before ~$4 GMGN cost) | n1 day | ESTIMATE, needs crew_wash.json (overnight) and ~100 real-lane coins
- 2026-10-05 22:10 BKK: PP's 5 Oct "real crowd" lane (+$241/trade, 21 coins) rests on two fake wins. 0x97d4c3c1b72aebb599b70002e9109b88248d51e3 sits in a CAMPAIGN honeypot hook (0x7fc338fcca56d4fcca49bc54fc84a7fcb5070a80, family 5059_7adb2ec7, same as AOTTR): a real buyer cannot sell. 0xd4ea1f3070595fafa1af0e85d82656c77a78a118 is a pond pool (fee 100, tick spacing 1 = crew type), pulled by 21:00. Rule: paper books must apply the hook family kill + sim sell test before booking any win; a wallet-count "real" tag is not proof. STATUS: flagged to Ben; PP to re-score all 46.
- 2026-10-05 22:40 BKK | PP CHECK OF pp-check.md: 0x97d4c3c1b72aebb599b70002e9109b88248d51e3 pick pool 0x73feddd5c9b0d4707c15f1551922b1280f6365ac5fa77c587babcb5323d2ad38 = campaign hook CONFIRMED (gate5 campaign kill was broken until 18:50, pick 08:01) | no win booked. 0xd4ea1f3070595fafa1af0e85d82656c77a78a118 = pond pool CONFIRMED, crew root 0x4905f29f9c0459ec97568ce3155a19041da70954 (11 launches); BUT paper sell at 20:24:50 BKK had 388 more sells after it, so the win was reachable at crew size ($20), not "fake" | lane, not impossibility. 0xfd19969780c329010f6e0fa9b3fed370af142cab = pond crew 0xecc83460b8e89a4fd1ccfb71753b135c9559dc9c. CORRECTION of PP 22:00 line: "+$86 best 2 removed" removed d4ea and 0x956c7a0d50b164cf8442d1ec983724e0383f5c7a, not 97d4; with pond pools, bad hooks and wallet-wash all removed the real lane is n13 +$106/trade, and WITHOUT 0x956c7a0d50b164cf8442d1ec983724e0383f5c7a it is -$20/trade | REAL LANE NOT PROVEN, one coin decides it. honestboard.py now splits the money book by pool lane (REAL / pond / bad hook / unknown; lane from crew.db launch fam+fee) | LIVE. Daily estimate agreed: $0 to ~$2,000, not proven.
- 2026-10-06 00:15 BKK | [stated] Ben: another project says a GitHub repo active 20+ days before launch predicts $1M to $2M. RH test (/tmp/gh20.py: prewatch gh_matches name-matched repos, first coin per name, peak = max(runtrack ath_mc, launch sell cap), dead coins in): all launches $1M 1.35%, $2M 0.72%; any repo before launch n303 $1M 0.7%; repo 20+ days old n190 $1M 0.5%; commits 20+ days before AND 3+ active weeks n28 $1M 0 (0/15 early half, 0/13 late half) | NOT SUPPORTED on RH (small n, name match is weak). Better test later: coins whose own site links the repo (deepread cards since 5 Oct) once 14-day outcomes exist.
- 2026-10-06 00:45 BKK | [stated] Ben: metas drive the trenches; being late to a meta likely fails unless the coin is special; wants a meta radar (approved build) and a separate SIGNIFICANCE score (real progress for sector experts, even if new-but-quiet). First history test (/tmp/metahist.py: catbase 27,313 launches 19 Sep..3 Oct, theme = word in name/ticker, spark = first $500k coin of the theme, 58 themes with 5+ later copies): the spark was the biggest coin of its theme in 49 of 58 (84%); copies after the spark reach $1M at 0.47% (0-1 day), 1.05% (1-2 d), 0.99% (2-4 d), 0.80% (4+ d) vs 0.70% base; 10x 4.65% / 3.66% / 3.94% / 1.99% vs 2.41% base | FIRST LOOK SUPPORTS BEN: late copies (4+ days) do worse than random on 10x and copies get no big-outcome bonus; the first strong coin keeps the crown. Caveats: 2 weeks, crude word themes (cat/dog/fun are evergreen, not metas), runtrack ath has bad outliers. deepread now also asks significance 0-3 + significance_why (deep.db nov.sig), backup deepread.py.bak-sig.
- 2026-10-06 00:40 BKK | META RADAR v1 LIVE (paper research): /opt/metaradar/metaradar.py (cron */15, flock, nice 12; pdguard re-adds cron + reruns if heartbeat >1 h, alerts >2 h; supervisor super.py reruns metaradar.service if >2 h; backup metaradar.good.py) -> meta.db (theme, theme_hour, coin, emb) + page http://159.223.88.168/hx-c9ois/lanes/metaradar.html. Theme = name/ticker word, embedding cluster (bge-small 0.83), launchpad hook (full address), paired quote token. Strength = DISTINCT names (first run: 'Team Human' launched 94x = one bot, not a meta). Dropped: official stock tokens and new pools of old tokens (birth 1+ day before pool; CASHCAT/STANDARD/GOOSE moved to hook 0xe449e013004db4a5681e9622ca10c5ba0ea610cc and read as a fake $12M "meta"). Peak = 3rd highest lanes price sample (spike guard) or runtrack ATH. RISING = 10+ names in 6 h, 3x own 7-day normal names AND money (money capped $2M a coin). Coin score 3 early (#<=10, <=12 h after spark) / 2 copy / 1 hot-meta copy or relaunch / 0 fading. Sonnet names a meta once, own cap $1/day (4 names cost ~$0.01). First live read: cat RISING (13 names vs 3.9 normal; Cashflow Cat 0x8c414a98bb51033c5891348f7fba1054aaf4e83a $2.4M, $8.5M vol, 3 h old, verified DexScreener) | LIVE. Money baseline only 22 h of data so far.
- 2026-10-06 00:40 BKK | CORRECTION of the 00:45 metahist line: redone as metahist2 (/opt/metaradar/metahist2.py, out /opt/hist/analysis/metahist2.txt) with bot clones counted once (26,739 launches = only 9,473 distinct names) and 564 old-token pools removed: base per name $500k 0.56%, $1M 0.24%; only 5 of 23 $1M coins sat in any theme of 8+ names; launch number in theme: #1 0.5% $500k, #2 1.1%, #4-10 0.7%, #11+ 0.5% (no first-mover edge visible); biggest coin of a theme was #1 in 1 of 16 themes, #4-10 in 9; copies after the first $500k coin: 1 of 187 hit $500k (all at 2-4 days); half of copies came 129 h after the spark (9 metas). The "spark biggest 49 of 58" was mostly clones and old tokens | WEAK: RH winners are mostly NOT meta copies; late copies still ~never win. Day-1 look of $1M winners (first 40 min, medians, winners vs rest): buys 96 vs 12, ETH in 15.4 vs 1.06, net ETH 10.5 vs 0.64, top holder share 0.86 vs 0.36 = winners start with big concentrated money (crew-like), not crowd copies | FACT (small n).
- 2026-10-06 00:30 BKK | pickguard fix: jobs started while priority mode was already on (nightjobs wash/crewgrade/knowbt at 23:41) ran unpaused; now re-paused every run while priority mode lasts; wash.py, crewgrade.py, metahist2.py added to NONESSENTIAL (backup pickguard.py.bak-meta). Cause of the long priority mode 22:41-00:32: aipick scored only ~20% of inits under load vs ~38% normal (real lag, not a false alarm) | LIVE.
- 2026-10-06 01:20 BKK | [stated] Ben: metas are what the trenches play between big launches; people tire of a meta within days; asked how to catch the FIRST coin of a meta, incl. small metas, training on history. Test metastart.py (/opt/hist, out analysis/metastart.txt + metastart.db + metacrowd.txt): RH day files, 39,889 named launches, starters 15 Sep..2 Oct; meta = a word new in the prior 72 h (0-2 uses) then copied by 3+ different names AND creators within 48 h; 200 metas (165 with 3-4 copies). Results: (1) starter $500k 5.0% / $1M 2.0% vs named base 5.8% / 1.5%; the biggest coin was the starter in only 8 of 36 metas with a $500k coin (#2 10, #3 7, #4-10 10), e.g. accelerationism best was Super Inu Accelerationism #14, intelligence best Artificial Girlfriend #4, cash best Cash Dog #5. (2) buying the starter when copy 1/2/3 appears (median 7 h / 20 h / 30 h after it) 10x 6.2% / 5.5% / 3.4% vs a random coin of the same age 8.2% / 9.0% / 4.1%: no edge, RH metas form slowly. (3) starters that ran had 20 buyers in the first 5 min vs 5 for flops (all launches 4) and were copied faster (196 vs 450 min). Crowd rule: coins inside a meta with 30+ buyers in 5 min $1M 5.1% (n178) vs 2.0% outside metas (n6,875); 20+ buyers 4.1% vs 1.7%; $500k lift small (7.3% vs 5.1%) | FIRST COIN IS NOT THE ONE TO CATCH; the winner is usually a later, better-packaged coin of the meta that gets a real crowd in its first 5 min. Small n (~9 $1M hits). Data gap: RH names only from 11 Sep (birth names empty for Jul-Aug), so 3 months is not available without a name backfill; Solana would need Dune (budget 4,000 credits/month).
- 2026-10-06 00:50 BKK | [stated] Ben: let the pickers run untouched, split them off from everything else. Found: kookwatch droplet is 2 vCPU / 4 GB with 3.8 of 4 GB swap used and load ~25; the pickers ran at Nice 5 (LOWER priority than research) and lived mostly in swap (aipick5 55 MB in RAM, 226 MB swapped). Fixed live, no restart: drop-ins /etc/systemd/system/aipick{,30,5,5b}.service.d/50-isolate.conf (CPUWeight 1000 vs 100 default, MemoryLow 200M/400M so their memory is never swapped out, Nice -5, OOMScoreAdjust -800) + system.slice.d/50-pickers.conf MemoryLow 1300M (needed for the protection to take effect; memory_recursiveprot on). No io controller on this box, so disk priority is not set. Fills after: 24-37 s. pickguard priority mode stays as the second layer | LIVE. Full split = pickers on their own droplet, PROPOSED, needs Ben's yes (purchase).
- 2026-10-06 00:55 BKK | PONS CURVE PICKER: NO EDGE. ponsmodel.py walk-forward 16 Sep..5 Oct, $200 per trade, exit v1 on the curve, 1% fee each way: T30 top 1% -$14/trade (n821), top 5% -$12, random -$17, top 1% lost on 19 of 20 days; T10 top 1% -$7 (n754, 10x 0.3%), random -$22, lost on 17 of 20 days | analysis/ponsmodel_T30.txt, ponsmodel_T10.txt | FACT: no Pons curve paper book added. (Model ranks better than random but still loses after fees.) Also: baseline_map done (34 subcategories empty on RH but big elsewhere); baseline_fetch.py started niced 00:55 BKK; wash.py 15 at its last day (4 Oct), paused by priority mode, report follows.
- 2026-10-06 01:10 BKK | [stated] Ben: how can pickers lose money with all we know; something is wrong or not applied. Answer: the main money book is NOT losing (+$268/trade n51 vs random -$152, 5 Oct, half the wins are $20-lane pond coins); the Pons curve picker loses (curve + 2% fees, 98.5% never graduate, our signals were built for pool launches). Audit of what the live picker USES: model on first 10-30 s chain features + kill list (777, campaign/junk hook families, serial creators). LOGGED ONLY, not deciding: crew/wash tags (_crew/_wash/_lane/_max_eth), crew grades, deepread (site/X/novelty/significance/attention/spin), solid score, meta radar, sim-sell honeypot test | FACT. Next: replay that adds each signal on past days and keeps only those that add money on later days (signals with history: crew grade, wash, family, meta + crowd; deepread has history only from 5 Oct, so it is tested forward).
- 2026-10-06 01:10 BKK | meta radar crowd check LIVE: for coins inside a RISING/HOT meta, 5+ min old, count distinct buyers in the first ~5 min (PoolManager Transfer recipients, 2,962 blocks, one paid dRPC getLogs per coin, max 40 a run, cached in meta.db crowd); 20+ buyers and coin <= 30 min old = PAPER BUY, flag mcap from lanes prices or DexScreener; page section scores best confirmed mcap since flag (2x/3x counts). First run read Cashflow Cat 0x8c414a98bb51033c5891348f7fba1054aaf4e83a 63 buyers, Mention Cat 0xfa45424404a1afb866262fbba8150cbcd3369c4f 146; but Cashflow Cat had already fallen -89% in 1 h ($2.4M -> $186k), so flags on old coins were cleared and only coins <=30 min old count | LIVE (paper).
- 2026-10-06 01:20 BKK | [stated] Ben approved a separate cheap picker server and asked for 100 overnight iterations x 50 ideas to optimize the pickers (Pons and beyond), testing each round and keeping the best, with out-of-the-box thinking (what if our approach is wrong, what are we not using, gaps). Created DigitalOcean droplet "picker-box" 167.99.74.137 (SGP1, s-1vcpu-2gb, $12/month, Ubuntu 24.04, key kookwatch-deploy). Shadow picker live 01:16 BKK: pbshadow.service runs the unmodified aipick5.py; main pushes inputs via /opt/pbsync/push.sh (cron */2, nice/ionice lowest; models, state, kill list, crew grades, code every 2 min if changed; day-db kv stub 30 min; crew.db 6 h; launch names + crew rows every 3 s); picker-box watchdog pbwatch.timer. 01:20-01:50 BKK test: 15 of 15 real picks identical, median fill 30 s on both. Not switched yet; switch + rollback steps in /opt/pbsync/STATUS.txt on main | LIVE (shadow).
- 2026-10-06 07:00 BKK | OVERNIGHT IDEA LAB (Ben's 100 x 50 request): offline lab in the cloud container (/home/claude/lab: 118,904 launches 15 Sep..4 Oct, 10.3M swaps, wallets, Pons 11.5M curve trades; honest exits reproduce knowbt T30 to the dollar on random). 13 agents, 92 rounds, ~1,500 logged ideas (2,000+ configs); split DEV 15-28 Sep / VAL 29 Sep-2 Oct / HOLDOUT 3-4 Oct used once. Gains flattened after ~round 70 (only sizing still added). Findings: (1) rank at 10 s beats 30 s; live picker decides at ~34 s by design (30 s setting + wait for next trade + ~2 s main-box overload); 10 s vs 30 s fill is worth ~$500/trade at $200 on the same picks (iter/SPEED.md). (2) puller-wallet filter (2+ first buyers whose earlier buys were mostly pulled) helps the $200 real lane, but the real lane faded across VAL (7x winners/day 30-55 -> 13-18; crews moved in, puller-present coins 15% -> 63%). (3) CHAMPION L-BROAD20: $20 per pick, 8-model ensemble rank A8 <= 10 at T10, fill at T12, sell thirds at 10x/15x/25x: DEV +$39/trade, VAL +$53, HOLDOUT +$54 (n127, both days green, best-2-removed +$50, random $20 = -$11/trade); about half its trades lose everything. (4) sizing by rank/wallet score (N-GBMSZ $0/20/40: holdout +$102/trade, $4.9k/day; O-K3 $20-80) beats flat $20 only if crews do not pull more on bigger buys (untested). (5) Pons: needs <= 0.5 s entry (+$69-76/trade at 0.3-0.5 s, -$3 at 2 s); RH has no public mempool, needs a sequencer-feed listener: no-go at current speed. (6) CORRECTION candidate: the lab finds no 75-100% pull trigger after outside buys in pond coins (pond 0.005 ETH: 13.5% vs 10.8% baseline; other crew coins 0.01-0.2 ETH ~5-14% vs ~5%); the earlier 75-100% may have counted crew batch wallets as outsiders; buys above ~$27 in pond coins are almost absent so the safe size stays unmeasured. (7) dead: second legs, copying wallets on real coins, crew/puller-sell exits (pullers sell ~339 s before the pull: profit taking), regime/time gates, re-entry after dips, runner tails, Pons graduation buys. Files: lab/iter/SUMMARY.md, HOLDOUT.md, AUDIT.md, SPEED.md, STRICT.md, leaderboard.json, ledger*.md | PAPER RESULT, 2 holdout days only.
- 2026-10-06 06:00 BKK | LIVE PAPER BOOKS on picker-box (/opt/labbooks, labbooks.service + watchdog, frozen models trained to 3 Oct, need daily retrain, not automated): fast10_real ($200 real lane, F2/E3 ensemble at T10), pond20 ($20 S3 pond picker), copy20 ($20 E60 proven-wallet copy, skip no-factory), broad20 (L-BROAD20); global skips kill list + copycat names (name used 2+ times in the past hour). Decision ~12.3 s after first trade live (p90 15.7 s). Board every 10 min, public copy http://159.223.88.168/hx-c9ois/lanes/labbooks.txt (main cron */5 scp via /opt/pbsync/pbkey). First 45 min: pond20 5 buys -$79 (3 pulled within 60 s), others 0 trades | LIVE (paper).
- 2026-10-06 06:30 BKK | phase 2 session re-check of PP data (read only): liqlock.db `lock` still 0 rows and `liq` table not created; liqbuild2.py has run 11 h 12 min (started ~19:18 BKK 5 Oct, flock /tmp/liqbuild.lock, nice 10); chain/b*.db only 5 files (b076..b080), no writes since 20:50 BKK 5 Oct (paused while liqbuild runs, by PP design) | hold-or-sell cross-check and all-venue rerun NOT done yet; numbers in claude/hold-or-sell-250k-v1.md stand on my own chain pull | WAITING ON PP (next check 14:00 BKK)
- 2026-10-06 07:10 BKK | MORNING CHECK: knowbt honest replays 17-18 Sep..4 Oct, $200, kill list, v1 exit: T10 top 1% +$722/trade (n355, 10x 43%), top 5% +$462 (n1,662); T20 +$670 / +$395; T30 +$601 / +$292; random -$9 / -$22 / -$69 | FACT: deciding at 10 s beats 30 s on the full 18 days (confirms the lab). Money book (honestboard 06:07 BKK) n58 +$242/trade (6 Oct so far n7 +$59, best removed -$165), random rand_30s -$99/trade. BUG FOUND AND FIXED: wash.py was OOM-killed twice (19:19 and 23:34 UTC) in the crew-table step (all crews x 7 days of swaps in memory ~700 MB), so crew_wash.json and wash_rescore.txt never got written; now one crew at a time (generator), backup wash.py.bak-oom, rerun started 07:05 BKK (paused by pickguard when priority mode is on). Lab books fix: broad20 fired 5 signals 05:15-07:00 BKK but all were blocked by global skips it was never tested with (3 copycat, 2 kill); each book now uses only its own lab filters (broad20/fast10/copy20 keep the kill list, pond20 none), untested skips logged as would_skip. Kill reasons: MeowKey 0xd415f72b2122e3fdf27dd25c684e0eb9acfcbc98 serial creator 0x820d27be3b67afaacc85b1a05da127a6aa71f1c6 (349 launches); Carbon Captain 0x5dfd34055502092a88a158e9a0784e0f9432db98 campaign hook 0xe38f57052d3fb14e68e7ce4017bc4270efa7ca80 | LIVE.
- 2026-10-06 09:15 BKK | CREW WASH RE-SCORE DONE (wash.py after the OOM fix, finished 07:46 BKK; analysis/wash_rescore.txt, crew_wash.json 2,651 crews of which 2,605 labelled WASH (rule: wash >= 80% or < 10 coins), public http://159.223.88.168/hx-c9ois/brx/crew-wash.md). 20 Sep..4 Oct, picks tagged with data known at pick time: share of old paper profit from WASH coins: aipaper 51%, aipaper30 59%, aipaper5 base_top5_nodead 86%, f5_20s_top5 76%, rule_smart2_clean 95%, ev/v5 books ~100%. Lane 2 REAL ($200, honest v1): aipaper n7 +$372/trade, aipaper30 n8 +$535, base_top5 n5 +$210, f5_20s_top5 n3 +$1,107, v5/blend top2 n2 -$185. Lane 1 WASH ($20, half 3x half 10x): +$5 to +$76/trade. Our picks pulled within 15 s of our buy: 0 in every book; within 60 s: 0-9 of up to 81 picks. Section 7: pulls within 60 s after OUTSIDE 0.1+ ETH buys: big pond crews 0x4757fc41278bd0f1b028c6dae71d381dc4558d98 127/3,705 (3%), 0x92a0849f9b1e9e02363a4bd8d74a317fa6ddae92 143/6,413 (2%), 0x0ec66d31aab7e387b2f4cda075e39f7a6a0f541b 1,582/9,515 (17%); non-pond crews 0xc4a2ffded2e39b6beb595e78950870dc5ec352d5 20%, 0x2a73e4a1a8b76dc6b08fa28e2f65d2f316b45172 19% | FACT: confirms the lab: no general 75-100% pull trigger after outside 0.1+ ETH buys (the review-session number for crews A/B/C n 8/8/3 does not generalise; crew A and B roots are not yet in our wallet map, so A/B themselves are not re-tested). REAL lane stays small-n (2-8 coins per book). 0x956c7a0d50b164cf8442d1ec983724e0383f5c7a (5 Oct pick) is not in the wash data yet (5 Oct day file not complete), so "real crowd?" is UNRESOLVED.
- [stated] 2026-10-06 review session: PP crew-wash table acceptance FAIL. PP crew 0x92a0849f9b1e9e02363a4bd8d74a317fa6ddae92 = review crew A, PP crew 0x4757fc41278bd0f1b028c6dae71d381dc4558d98 = review crews B/C. PP wash 78%/84% vs review ~98%; outside buy ETH 2,310/1,956 vs ~13. Likely cause: army wallets missed by repeat-only rule. Fix note: http://159.223.88.168/hx-c9ois/brx/pp-wash-fix.md (md5 62898b4c00dc5292d6968a0ea7128efa). Lane 1 positive stands; Lane 2 not usable until fix passes. Pull trigger 75% (n19 outside big buys) disputed by PP 09:15 and lab 07:00; open until distributor first-funder rule is added.
- 2026-10-06 ~10:20 BKK | FOCUS VERDICT (review session, synthesis of runners-20m-week, runner-patterns-v2, hold-or-sell-250k, picker-study, lab, crew-wash): profit = money other people bring in after us. Crew coins: outside money tiny (crew A 13 ETH in 13 days), so only a $20 skim of the crew's own ETH, ceiling ~$500 to $1,000/day; paper positive (PP +$5 to +$76/trade, lab holdout +$54, live paper broad20 6 Oct n7 +$52.5). Real coins at launch minute: NOT proven (n2 to 8 per picker, one coin decides, faded late Sep as crews moved in). Real big runners (26 small-start coins to $20M within 7 d) mostly on pons_v2/v3/other venues, 0 of 32 plain v4 launches; signal only at hour 4 to 8 (holding highs, spread holders, holders doubling, low churn, bigger X accounts); $3M+ at hour 8 = 29% to $20M on survivor tracker vs 1 of 5 clean. RECOMMENDATION: stop hunting real winners in the first 30 s; build an all-venue climb watch from $500k with 1/4/8/24 h checks and paper buys at hour 4 to 8 (real size possible); run crew lane live $5 then $20; keep new-lane radar; drop Pons curve, meta copies, second legs, wallet copying | PROPOSED
- 2026-10-06 10:15 BKK | phase 2 re-check #2 (read only): liqlock.db `lock` 0 rows, `liq` not created; liqbuild2.py pid 2416951 alive 14 h 57 min but disk-starved (state D, 3.5% CPU, read 91 GB, liqlock.db-wal 4.7 GB and growing, last log step "poolfam" at 1,185 s, so the lock/liq step has run ~14.5 h); runs at nice 10 + ionice idle on the overloaded main box (load ~25), so it only gets disk when nothing else wants it; all-venue chain/b*.db still paused behind it (5 files, no write since 20:50 BKK 5 Oct) | SLOW, NOT DEAD: PP should raise its disk priority, run it on picker-box, or pause research jobs; until then hold-or-sell numbers stand on my own chain pull | WAITING ON PP (next check 14:00 BKK)
- 2026-10-06 11:00 BKK: SECOND-0 POND FINGERPRINT (live watch, on-chain priced, 82 pond coins 5/6 Oct, entry first trade +10 s, half at 3x rest at 10x). PASS = token runtime code 3855 bytes md5("0x..." string) 17ad80... AND launch tx (PositionManager 0x58daec3116aae6d93017baaea7749052e8a04fa7 multicall) sends 0 ETH: 28 coins, 17 hit 3x, 5 hit 10x, +94%/coin. FAIL (custom token bytecode, launch tx 1.6 to 2.0 ETH): 54 coins, 0 hit 3x, -100%. Built into fmlane1 v3 (tags pass at buy; LIVE buys pass only, Ben's switches /opt/fm/LIVE_L1 + GMGN_ALLOW_AUTOMATED_TRADES=1). STATUS: one night only; Price Predict asked to backtest 20 Sep to today (pp-asks.md item 0).
- [stated] 2026-10-06 Ben made the review session supervisor of the SECOND WINDOW (any buy later than the first seconds); asked for PP status on Pons/other venues and a step plan.
- 2026-10-06 11:30 BKK | PP STATUS (droplet 10:40 BKK) | ponstape + venuepick LIVE (Pons curve, graduations, v2, v3, repools at 30 s; last ~19 h blind: curve -20% n2,008, grads ~0% n29, grad hold -24% n23, v2 -15%, v3 -33%, v3 model picks +98..+239% n7-12); chain/b*.db all-venue swaps STALLED since 5 Oct 20:50 BKK (5 files, liqbuild2 running 15.5 h); liqlock `liq` table NOT built; no after-launch all-venue stage watcher; no wallet data on hold-lane pools | FACT
- 2026-10-06 | SECOND WINDOW on Pons graduated pools (1,708 pools 13 Sep..4 Oct, every swap, honest exits: 1% fee each way, depth impact scaled with price, 3 s delay; halves 13-23 Sep / 24 Sep-4 Oct) | pool depth ~4.4 ETH at graduation ($500 round trip ~18% cost, $100 ~6%); blind gate buys ~break even ($250k 1.00, $500k 1.02 at $500); fixed-time buys 15 min..8 h lose (0.63-0.79); low sell share since graduation better at every gate both halves; $100k gate + sell share <= 30% + half at 2x/trail 50%/stop -50%: second half +18% per $100 trade (90% CI -3%..+41%, n64), +11% at $500, best coin removed ~+3%; higher-gate filters and hour checkpoints failed out of sample; graduations ~160/day mid Sep -> ~35/day now | /root/w2 it1-it5.py (container) | LEAD (small edge, lumpy)
- 2026-10-06 | RULE: second window buys only in pools that cannot be pulled (Pons graduated, LP dead/burned, locker contracts, never-pull launchpad hooks); size <= 2% of pool ETH depth | plan + PP steps (unblock b*.db, holdlane flag, climb table 5 min/24 h then hourly 7 d with wallets/holders/top10/X, backfill 13 Sep+, paper book w2_pons100, go/kill ~20 Oct) | http://159.223.88.168/hx-c9ois/brx/second-window-plan.md (md5 41806b89c5ab307eb683ea464e7aa820), project claude/second-window-plan-2026-10-06.md | PLAN
- 2026-10-06 11:10 BKK: EDGES V1 (25-day backtest, 3 s sell latency, fees in; http://159.223.88.168/hx-c9ois/brx/edges-v1.md). Lane A pond pass (std token 17ad80 + 0 ETH launch) + HOT regime (4+ of last 10 resolved pass coins hit 3x) = +92%/coin, every day 30 Sep to 4 Oct positive; ladder 3x/10x. Lane B c500 (fee 500, spacing 1, native, no hook, PoolManager direct init, token 4655 bytes) + HOT (5+ of 10) = +105%/coin, every day +70..+119%; ladder 3x/5x. Entry 0.5 to 10 s after first swap equal, 15 s+ worse. Stops/time stops hurt. Our buy size does not trigger pulls. fee 1% PositionManager lane fading, unused. Deployed in fmlane1 v4 (watch). STATUS: paper; needs live fills.
- 2026-10-06 11:05 BKK | CORRECTION of my 09:15 line: its "no general pull trigger after outside 0.1+ ETH buys" used PP's wash tag, which the review session FAILED (army wallets counted as outsiders: wash 78/84% vs ~98%), so that conclusion is UNRESOLVED, not confirmed, until the first-funder fix (pp-wash-fix.md) reruns.
- 2026-10-06 11:05 BKK | LIQBUILD CHECK (Ben asked "is this true" about the 10:15 note): liqlock `lock` 0 rows, `liq` table missing, liqbuild2.py pid 2416951 alive 15 h 41 min = TRUE; BUT the main cause is not only idle disk priority: the process was SIGSTOPPED (state T) by pickguard priority mode, which was ON most of the morning (07:50..10:49 BKK, e.g. 08:47-09:07, 09:18-09:43, 09:51-10:06, 10:21-10:46, 10:49-). FIX: pickguard no longer enters priority mode while picker-box is primary and healthy (primary5.state() == PB), because main's pickers are only the backup now (backup pickguard.py.bak-pbprimary); research jobs incl. liqbuild2 resume when the current window ends (~11:13 BKK) | LIVE.
- 2026-10-06 10:17 BKK | [stated] Ben: switch pickers to picker-box (keep main as backup for a day), decide at 10-12 s, L-BROAD20 paper only, add crew recent hit rate, publish pick feed. DONE: picker-box primary via /opt/hist/primary5.py + analysis/primary5.json {"primary":"pb"} (pull every 15 s by /opt/pbsync/pull.py; silent failover to main if picker-box heartbeat/copy >120 s, tested 10:22-10:25 BKK; sentinel + pdguard repair); ppexit/honestboard read primary; fm_feed to the trader NOT switched; main aipick5/5b keep running as backup (stop after a clean day, steps in /opt/pbsync/STATUS.txt). Money book now decides on the clock at 10 s (m10 books, fill at first swap after decision; decision median 11.0 s live vs old 32.7 s fills; old 30 s book kept as control). Lab crew rule (skip if crew had 0 10x in its last 10 known coins, needs >=5 known): broad20 VAL +$52.6 -> +$60.0/trade ($3,907 -> $3,720/day), pond20 +$33.9 -> +$42.6; live shadow books broad20_crew/pond20_crew/fast10_crew on picker-box; pick feed http://159.223.88.168/hx-c9ois/lanes/labbooks_picks.json (+ .csv, every minute) | LIVE (paper).
- 2026-10-06 12:40 BKK | BIG WINNER PATTERN = LANE, NOT COIN (second window supervisor, 9,986 Pons graduations 5 Aug..3 Oct, 505 hit $1M, 97 $5M, 20 $20M; hourly closes for 2,071) | $5M rate by graduation week 8% (5-9 Aug) -> ~1.5% late Aug -> 1.1% early Sep -> 0.25% late Sep; 16 of 20 $20M before 7 Sep; new quote sublane first 3 days $5M 3.7% vs 0.35-0.7% older (after 7 Sep), 2.3% vs ~1.1% (before); coin traits flat (curve speed, hour, weekday, ticker length, quote type) | g1.py, g2.py (/root/w2) | STRONG (regime + newness)
- 2026-10-06 | LANE LEADER: first coin per Pons quote lane to close an hour >= $1M, buy $1k at that close, half at 2x / trail 50% / stop -50% / 14 d, 1% fee each way + depth slippage | n31 (~0.7/day) +71%/trade, LOO +25..+77%, boot 90% +1..+159%; before 7 Sep +91% (n20), after +35% (n11, median +47%); rank<=1 +71%, <=2 +43%, <=3 +25%, <=5 +17%; all $1M crossers after 7 Sep -24% | g4.py | LEAD (small n; NOVAAI +1,440% inflates mean; Pons only)
- 2026-10-06 | plan v2 for PP: lane table (venue x quote; lane age, launches/day, money-tested hourly close at $250k/$500k/$1M/$2M with rank), paper books w2_leader (+ rank 2/3 controls), w2_newlane, w2_pons100; backtest leader rule on long.xyz/bankr/v3 burned LP once b*.db lands; go/kill ~20 Oct | http://159.223.88.168/hx-c9ois/brx/second-window-plan.md (v2, md5 51e80ae93934db3555ad858d0097e588) | PLAN
- 2026-10-06 14:10 BKK | [stated] Ben: winners are a social chain reaction; map the virality steps. VIRALITY LADDER (163 coins >= $500k, aligned at first $500k; 10-min holders/buyers, hourly wallets for 152, X CA posts -24 h..+48 h; survivor-skewed tracker) | $20M rate 14% base -> quiet (<=1 X post 24 h before) 19% (no 6%) -> + fresh buyers (<=6% repeat trench wallets in h0-2) 31% -> + buyers h2-4 >= half of h1 57% -> + cap >= $2M at h4 75% (n8); winners ~4% repeat wallets vs stalled 14-28%; buyers 11x/19x/39x stalled at h4-12/12-24/24-48; first 10k+ X post after $1M in 83% of winners (median cap $3.8M), stalled get it at ~$0.5M (shill); money at h4: steps 1-3 +31% (n13), steps 1-2 +13%, rest -16% | /root/w3 r1-r10.py | LEAD; doc http://159.223.88.168/hx-c9ois/brx/virality-map.md (md5 ffb0ef00143261b18fc9b59d36e7055a)
- 2026-10-06 ~11:30 BKK | PRELAUNCH ROUND 2, winners first (prewatch session; 17,796 clean launches 19 Sep..3 Oct, all 932 winners + 1,472 sampled others; WIN = 10x@40s or 10x within days or $500k; blind Sonnet label of what each project is; GitHub commits before launch; X posts before launch) | announced coins win 2.6% vs 6.4% unannounced, both halves; announced coins launch in the pond 0.0% of the time vs 2.5% (pond wins 26.6%) = announcement marks launchpad-template launches; best inside announced: project type launchpad/NFT/stocks/game 4.7% (x1.8, early 6.9 late 3.4) still below all coins 5.2%; GitHub 2-6 active days 4.2% (n9 winners), 7+ days 0 of ~115; start cap <$10k lifts every coin (11.6%), announced adds nothing stable (late half 11.6% vs 18.2% unannounced); walk-forward rules all lean on cap <$10k with 4 to 6 late winners | r2_an.py, r2_combo.py (/opt/prewatch/bt) | KILLED: no pattern beats the base; matches virality-map "quiet before is better" | doc claude/prelaunch-watch-v1.md v2
- [stated] 6 Oct 11:30 BKK supervisor edge round 2: c500 HOT out of sample 9/21-9/29 positive every day (+91..+235%/coin). c500 launches falling (~250/day late Sep, ~90/day pace 6 Oct). NEW c10k class (fee 10000/spacing 200, token 5274 bytes, 3 or 5 ETH launch via PositionManager): ~700/day, sell all at 3x, +96..+105%/coin, every day positive 9/30-10/4; in lane1 v5 watch only. Dead ends: pond fail-coin small-target scalp, 81-88% fee pools, per-deployer signal (fresh wallets), time of day.
- 2026-10-06 ~12:30 BKK | PRELAUNCH x CREWS (r3_crew.py, PP crew.db roots, crew = 10+ earlier launches) | announced coins come from crews as often as quiet ones (29% vs 30%), good crews rarely announce (1% vs 4%); crew + announced 4.1% vs crew quiet 8.9%; good crew + announced 17.3% vs quiet 26.4%, week 2 walk-forward 8.0% vs 29.1%; announced winners cluster in roots 0x5879317331054e0e5554f950eed57df573b39ab1 (9 wins, all week 1) and 0x59442072d89adc833993407ae20bb5ede83fc918 (famous borrowed X accounts, 20 of 64 launches 10x, both weeks) | FACT: crew grade is the signal, announcing lowers it; crew.db launch.t is insert time