---
name: trading-edge-hub-4
description: Trading edge hub ledger part 4 (continues trading-edge-hub-3): live paper picker ops, reliability guards, go-live readiness, from 4 Oct 2026.
sources: [cowork]
---
# Trading edge hub, part 4 (follows [[trading-edge-hub-3]])

- [stated] Ben 4 Oct 2026: breakages (lost picks, fake results, late buys) must never happen again; redundant monitors must fix within a minute; system must be ready for going live.
- [stated] Ben is building a categorization engine in another session: category tree (e.g. financial service > lending > type) from the project's own website, X bio and tweets; uncategorized usually means memes or lazy launches; goal is pattern recognition and new-lane detection.

## 4 Oct 2026 21:10 BKK: pickguard live guard
- pickguard.py runs every 20 s (systemd timer) plus every minute (cron, flock). Verified: cron kept running while the timer was stopped.
- Fixes: dead heartbeat -> restart (90 s grace after a start); 3 restarts in 15 min -> roll back to last known good code; code errors -> roll back; late picks -> pause research jobs; leaky model files -> roll back to .prev.
- New check: every picker fix leaves a marker in the code; if a deploy loses one, the last known good copy is restored at once.
- Incident: at 20:44 BKK a stale Mac copy of aipick5_main.py overwrote the server copy and wiped the lastb, restore heartbeat, late skip, parallel pull and named column fixes. aipick5 stalled in a long restore and was dead 20:45 to 21:06 BKK (about 21 min of round 5 picks lost). The first pickguard version then restart-looped it every 20 s.
- Fix: rebuilt from the last good copy plus the EV books; heartbeat now written on every log chunk; Mac copies refreshed from the server. Rule: edit on the server, never push a Mac copy over it.
- Go-live gaps still open: no execution path (paper only); free RPC tier gives 429s, so live needs a paid RPC with real eth_getLogs ranges.
- 21:10 BKK: aipick5 now restores open books one per loop, in the background. New launches no longer wait behind a restore (before: about 3 min blind after each restart). Verified: a pick was saved 24 s after a restart.

## 4 Oct 2026 21:45 BKK: no launch can be lost (Ben: "build it in a way where that can never happen")
- [stated] Ben: lost launch info must never happen again; on categories: use the project's links/website if they exist at launch, if not, not.
- Layer 1, warm standby aipick5b: same picker, passive while aipick5 heartbeat is fresh, takes over within 15 s. Tested: froze aipick5, standby took over 21:24:46, handed back 21:25:38.
- Layer 2, reconcile.py (timer + cron every minute, flock): launch ledger from a saved block cursor (never skips a block), then every launch must have a score in each picker DB 3 min after its decision moment; missing ones are rebuilt from the chain cut at the exact decision block and scored with the same model. Buys booked as src=catchup and kept out of live results ("missed live" line in the daily report). Exits tracked from the chain.
- Layer 3, nightly cross-check of the ledger against the rebuilt day DB.
- Replay audit: live aipick5 features recomputed from the chain at the same block. First 8 coins: 47 of 47 features identical. Alert if 5 of the last 10 differ (guards against fake results).
- Finding: the old aipick/aipick30 pickers silently skipped many launches (3 decisions per loop budget). Catch-up is filling those (76 so far).
- Live picks now INSERT OR IGNORE (a book is never overwritten); pickguard marker list extended; supervisor watches reconcile and standby.

## 4 Oct 2026 23:30 BKK: lane map v2, novelty classes, lane database (cowork; container /tmp/lanes/research, verdict5; droplet /opt/laneradar; public http://159.223.88.168/hx-c9ois/lanes/map/; doc claude/lane-map-v2-2026-10-04.md)
- [stated] Ben: not going live yet, just build the map; projects that do something different say it on X or their website (e.g. lending, but with gold or stock memes); plain members of a category are ignored, only novel things count; decide in a couple of seconds; measure what share is novel and track 72 h prices in a database (an input to his winner-pattern process); novel can be insignificant or significant (a new idea that makes sense and crypto is missing); 10 deep research iterations on the crypto product world incl. Robinhood Chain direction and leadership quotes; may use 1M twitterapi credits per 24 h.
- 2026-10-04 | cowork: map v2 | 293 categories in 17 groups (definition, business model, token role, examples, maturity, RH status, cosmetic vs meaningful differentiators, open gaps; 16 "missing" demanded ones); rubric: NONE, COSMETIC, NOVEL_MINOR, NOVEL_SIGNIFICANT, score 0 to 3, 20 quick tests, dated commodity list, top 40 gaps G1..G40; from 9 research passes (~49k words) + synthesis; ~110k X credits | map_v2.json, novelty_rubric.md | LIVE (reference)
- 2026-10-04 | cowork: share that is novel (5,754 RH coins with first-minute text, 4 Aug to 2 Oct, blind judge) | NONE 44.8%, COSMETIC 52.2%, NOVEL_MINOR 2.8% (159), NOVEL_SIGNIFICANT 0.2% (12, e.g. PolyFrog launch-graduation markets, violeti stock options, Parity dividend-aware LP hook, hoodoracle session oracle, Trion GPU-hour index options, MERGER cross-issuer stock book); ~0.7% of all 23,490 active RH coins | v5_all.json | FACT
- 2026-10-04 | cowork: does novelty pay (entry first hourly close, same-day baseline, 2,951 coins with candles) | 72 h 2x/5x lift: NONE 0.92/0.91, COSMETIC 1.05/1.06, NOVEL_MINOR 0.98/1.02 ($1M peak 1.18), NOVEL_SIGNIFICANT 1.76/1.91 (n10, 0 hit $1M); higher 10x groups: dex_trading 11%, derivatives 11% vs games 3%, privacy 1% | outc5.py | KILLED as a standalone buy signal (keep as a feature)
- 2026-10-04 | cowork: CORRECTION of hub-3 "NEW flags 10x 11% vs 3%" | baseline was a small biased sample; full data, same-day baseline: v3 NEW flags n243 2x lift 1.11, 5x 1.14, 10x 1.59, $1M 1.43 | outc5.py | CORRECTED
- 2026-10-04 | cowork: math vs judge | judged-novel share by similarity to earlier projects is flat (2.9% under 0.80, 4.2% at 0.80 to 0.85, 0% above 0.90); kNN category from judged cards exact 27.7%, group 59.1% | knn test | FACT (AI judge needed; kNN is fallback)
- 2026-10-04 | cowork: lane database LIVE, no alerts | /opt/laneradar/lanes.db: every new RH coin each minute (card, flags, nearest projects, kNN category, judge fields), prices from the v4 PoolManager every 5 min in hour 1, 15 min to 6 h, hourly to 72 h (coins with text or a flag + ~10% control); view coin_outcomes; watched by super, stale.py cron, livewatch2, guardkit; judge fields empty until /opt/common/anthropic_key exists | lanedb.py, tracker.py | LIVE (paper)
- 2026-10-04 | cowork: transfer route container to droplet | dpaste.com API (expiry 1 day), base64 tar.gz in 200 KB parts, md5 verified | dpaste | FACT (never for secrets)

## 5 Oct 2026 00:00 BKK: category tree + 20 s labeler + stats page (cowork; droplet /opt/laneradar radar3.py, stats.py; public http://159.223.88.168/hx-c9ois/lanes/stats.html)
- [stated] Ben: wants a category TREE (domain > category > type > subtype, e.g. Financial services > Lending > type > subtype) to see where launches land and how it shifts over time; statistics on coins reviewed, how long labeling takes, when it is slower and why; GMGN is the best source for socials (use the API carefully, never push into rate limits); whole process must take under 20 seconds and the label time must be saved; coins with no socials get no category; keep "Uncategorized" as its own group (memes or lazy launches there is information by itself).
- 2026-10-05 | cowork: tree_v1 | 15 domains, 62 categories, 291 types (all 293 map v2 ids), 532 subtypes; Uncategorized = no_info + lazy_launch | tree_v1.json | LIVE
- 2026-10-05 | cowork: where launches land, 4 Aug to 2 Oct (22,972 coins with 300+ trades in 30 min) | Uncategorized 68%, Memes 15%, Token mechanics 3%, Token launch 3%, AI 2%, Games/bets 2%, Social 2%, Finance 1%, rest under 1% each | history table | FACT
- 2026-10-05 | cowork: fast label without the AI judge (logistic model on 5,697 judged cards, time split) | domain 63% right (80% when confident, 47% of coins), category 39%, type 26% | fastclf.npz | LIVE as fallback (AI judge needed for fine levels)
- 2026-10-05 | cowork: GMGN token info | answers in ~0.25 s, link has twitter, website, telegram, description; often has socials the chain does not (HoodPad); called once per coin at weight 1 under the shared gmgnlimit reserve | radar3.py | LIVE
- 2026-10-05 | cowork: radar v3 always on | laneradar3.service polls new pools every 2 s, labels each coin in parallel (chain meta, GMGN, tokenURI, X bio, website, linked post, embed, fast label, AI judge if key); first 28 coins median 1.3 s, 90% 8.4 s, max 12.2 s from the launch block, 100% under 20 s; slow causes seen: model warm-up, full-history quote check (now parallel); times per step stored (t_block, t_seen, t_labeled, latency_s, ms, slow_reason); stats page every 10 min; old minute timer off; guards super, stale.py, livewatch2, guardkit moved to radar3.ok | lanes.db | LIVE (no alerts; AI judge waits for /opt/common/anthropic_key)

## 5 Oct 2026 07:50 BKK: overnight check of the live pickers
- Overnight the live round 5 picker decided only about half the coins it should have (live ~35/h vs catch-up ~30/h). Two causes: (1) false restarts, pickguard read the heartbeat file while it was being rewritten (empty, read as dead); (2) restoring old open books blocked the live loop for minutes (a 965 s late fill). Catch-up filled every miss (748 coins for round 5), so the record is complete.
- Fixes: atomic heartbeat writes plus a retry read in pickguard; no restore on restart (reconcile.py now tracks exits of EVERY open pick from the chain); reconcile uses the public RPC only, so it never competes with the live pickers; pickguard no longer stacks reconcile runs. Since 07:34: catch-up 0, live scored every coin it could decide. Reconcile runs dropped from 516 s to about 20 s.
- Replay audit: 13 earlier "differences" were an audit bug (a token with two pools; the audit matched the wrong pool). Fixed, and live now saves the pool id. After the fix: 15 of 15 identical.
- Last 24 h (paper, $200 each): round 5 live 91 buys, 19 closed, 13 hit 10x, +$21.7k. Missed live, scored later: 449 buys, 111 closed, 46 at 10x, +$66.1k (on-time fill assumed, not tradeable, kept apart). 30 s picker live +$13.0k (9 of 24 closed hit 10x). 60 s picker live +$3.5k.
- Standby took over twice (05 Oct 07:05 and 07:30 BKK), under a minute each.

## 5 Oct 2026 08:00 BKK: lane radar review (cowork; droplet /opt/laneradar radar3 v3.2; doc claude/lane-review-2026-10-05.md)
- 2026-10-05 | cowork: AI judge DOWN since 04:52 BKK | Anthropic API "credit balance too low" on every model incl. Haiku; coins still get a first label in ~4 s, no AI label or novelty since | judge_health.json | OPEN (Ben must top up; lanecheck alerts once via alertgate)
- 2026-10-05 | cowork: early outcome by label (coins launched 6 to 24 h earlier, live radar) | hit $100k: Uncategorized 30/375 (8%), Memes 25/102 (25%), AI 28/77 (36%), Token launch 23/72 (32%); by novelty: NONE 27%, COSMETIC 36%, NOVEL_SIGNIFICANT 7/14 (50%) | lanes.db prices | EARLY (small n, re-check at 72 h; Haiku over-called SIGNIFICANT 4.3% vs 0.2% blind study)
- 2026-10-05 | cowork: relaunch spam | 34% of coins with text repeat an earlier text (Team Human 58 launches in 6 h); counts per category are inflated without a project layer | same_text label | LIVE (relaunch_of column)
- 2026-10-05 | cowork: detection delay | median 3.4 s from block to seen (public RPC head ~2 s behind + 2 s poll + getLogs ~1 to 2.5 s); work after seen 0.3 s | it4 | FACT (WebSocket push would cut ~2 s)

## 5 Oct 2026 08:00 BKK: "what the coin IS" + sentinel
- [stated] Ben: categories (what a coin is) are a whole new class of edge; he built an AI brain on the crypto category tree to flag novel and significant things; wants an AI brain test that learns and decides by itself within 30 s; asked for 50 ideas (doc claude/edge-ideas-round10.md).
- RECEIPT: token name and symbol alone (known at second 0, 85% of 25,892 RH launches have one), crude word lists, both halves stable: tech/infra names 10x rate 4.9%/4.0% vs base 2.45% (1.8x); money/finance 1.4x; AI/agent 1.3x; dog, politics, frog 0.6 to 0.8x; no name 0.4x. No live model uses the name yet. Status: STRONG LEAD, next = AI category reader + category heat test.
- [stated] Ben: an hourly redundant check that also fixes; any unexpected behaviour caught within 5 min and fixed immediately.
- Built sentinel.py (1 min timer + cron): judges pickers against the chain (missed coins, late fills, live vs replay, restart storms, ledger, watchers, too-good results), fix ladder pause > restart > roll back > alert; hourly full 12 h check. Map: claude/watchers-map.md.
- Also fixed: the 60 s and 30 s pickers only decided 3 coins per loop (now 10); reconcile starved old launches after one bad range (now skips and rotates).

## 5 Oct 2026 08:10 BKK: golden nuggets in project text (cowork; doc claude/golden-nuggets-2026-10-05.md; radar3 v3.3)
- [stated] Ben: stay fully on categorization, understanding what the project does, to find the golden nuggets (significant, high quality, real thought) among trash and copycats; no smart-wallet or market signals in these ideas; wants the smarter brain made efficient without any quality loss.
- 2026-10-05 | cowork: what separates significant projects in first-minute text (5,754 blind-judged RH coins, 12 significant, 159 minor) | concrete mechanism words minus buzzwords AUC 0.80 (significant avg 0 buzzwords vs 0.4); demanded-gap words AUC 0.72; similarity to earlier projects 0.51; length 0.66; has X 0.66; a word filter cannot gate the AI (skipping lowest 74% loses 4 of 12 significant) | feat_nov.py | FACT (use as hints to the AI, never as a gate)
- 2026-10-05 | cowork: AI cost root cause | each call sent the ~30k-token tree+rubric plus 120 memory lines; Haiku ~$0.01/call, $5 lasted ~5 h; v3.3 sends 25 relevant memory lines, keeps the book cached, skips exact copies, logs real cost per coin (cost_usd); Opus 5.5 est. ~$12 to $15/day for ~950 text coins (was ~$33) | radar3 v3.3 | LIVE (unmeasured until credit is topped up)

## 5 Oct 2026 09:00 BKK: categories into the trading process (this session took over the lane radar)
- [stated] Ben: hand the categorization over to this session and build it into the process; do not let an AI learn end-to-end from the ~650 winners; test the best of the 50 ideas; use the smartest brain but query it efficiently, saving credits without losing quality. Golden-nugget findings from the other session: concrete mechanism words strongest, hype words not, naming a real existing need counts, "looks different from earlier coins" tells nothing.
- 2026-10-05 | contract family (bytecode md5 of the token, known at second 0, 16k of 26k coins read, both halves) | template 0cdcd385 10x rate 7.9%/7.1% (3.2x), c68bd844 2.6x, small families 1.8x, unique code 0.7x, template 98931e52 (3,218 coins) 0.08x | catstudy3 | STRONG
- 2026-10-05 | same-name history | name of an earlier winner 1.9x; 1 to 5 earlier same-name launches 1.2 to 1.3x; 6+ copies 0.84x | catstudy | MODERATE
- 2026-10-05 | context | first launch of its name kind in the last hour 1.8x (stable); 11 to 20 coins hit 10x in the last 3 h 1.5x; same-kind heat only 1.2x; launches 7 to 11 BKK 0.3 to 0.7x, 14:00 and 01:00 best | catstudy | MODERATE
- 2026-10-05 | deployer history | raw: first-time deployer 1.4x, deployers with 2 to 5 earlier launches 0.09x; BUT creator known for only 40% of coins, and in the model it added nothing | catgbm deployer | UNRESOLVED (fill all deployers, retest)
- 2026-10-05 | history AI labels (domain, novelty) vs 10x | CONTAMINATED: the history table only labelled coins that later traded (labelled coins 10x 6x more often), so no conclusion | catstudy T1-T4 | VOID; clean daily test ailabtest.py on launch-time Opus labels started
- 2026-10-05 | does identity add to our model (walk-forward 9 days, deep5 features vs + name kind, copies, hour, contract and hook families) | top 1% 31.8% -> 37 to 38%, top 2% 26.6% -> 29 to 30%, top 5% 18.4% -> 21 to 22% (better on 8 of 9 days) | catgbm | LIVE as paper books id5_top1/2/5 in aipick5 (idmodel.pkl, idlive.py, nightly idtrain in retrain.sh)
- 2026-10-05 | AI judge paper books | ai_novel_sig, ai_novel_appeal2, ai_appeal3: bought at the chain price at the moment the Opus/Sonnet label landed (max 120 s after launch), exits from the chain, booked by reconcile.py; Haiku labels excluded (over-calls significant) | reconcile ai_books | LIVE (paper)
- 2026-10-05 | AI judge cost, Opus 5.5 measured | about $0.027 per coin: cached rule book ~37%, coin text + memory ~40%, answer ~23%; ~$15 to $20 a day; last hour Opus: 151 NONE, 8 COSMETIC, 0 novel (strict) | lanes.db cost_usd | FACT

## 5 Oct 2026 09:40 BKK: AI brain guard + solid-project score (doc claude/ai-brain-round11.md)
- [stated] Ben: the AI brain must never drop; if it does, catch and fix within a minute, redundantly; liked saving credits without losing quality; thesis: serious teams write clear text about what the project does and how it differs, weak ones use hype or vague words, no-text coins are a separate class; always ask the out-of-the-box questions before generating ideas.
- 2026-10-05 | text vs launch-day 10x (1,700 projects with first-minute text and day-file outcomes) | every text signal AUC ~0.5 incl. AI novelty class | nugfeat | FACT: text does not predict launch pumps; keep it out of the 30 s pickers
- 2026-10-05 | text vs still alive 10+ days later (5,697 projects, DexScreener mcap >= $100k or $1k vol + $5k liq; AUC per launch week) | length 0.65 (every week > 0.55), mechanism words 0.62, AI novelty class 0.61, RH-native words 0.56, hype/buzz/emoji ~0.50 (not negative); embedding model 0.66, combined 0.70, top 10% alive 21% vs 8.4% | nugfeat2/3 | STRONG for lasting projects
- 2026-10-05 | judge skip test | skipping the lowest 20% by free score loses 1 of 46 new-idea coins, 0 of 5 significant, 10 of 143 alive; so route them to the half-price batch service instead of skipping | nugfeat4 | NEXT (not built)
- 2026-10-05 | built | judgeguard (20 s timer + cron; output-based check, instant alert only for credit/key, fallback labels re-done by Opus); solidlive service (free solid score seconds after the radar, nuggets page http://159.223.88.168/hx-c9ois/lanes/nuggets.html); soltrain nightly 06:31 BKK (alive labels + retrain; test AUC 0.66 to 0.67, top decile alive ~9% vs bottom 1.3%) | LIVE

## 5 Oct 2026 09:50 BKK: next meta brain (doc claude/next-meta-brain.md)
- [stated] Ben: a concrete text alone does not make a pump; what should pump is a project that is NOT a commodity, in an uncrowded space, maybe a trending category, unique, with in-demand features, fitting or shaping a hype narrative ("this coin created the new meta"); own code matters (some copied basics are fine if key parts are self-written); a working product matters only if the idea is new or cool; do not focus on saving credits now; read docs and code for all coins, not only the top 10%.
- 2026-10-05 | old data, 1,734 projects with text + launch-day outcome | crowding by similar text, AI novelty and "meta forming" clusters show no clear launch-day signal; similar-text clusters were relaunch spam and stock tokens; one contract template 10x 31.7% (n41); own code + novel 6.2% (n16) | metastudy.py | UNRESOLVED (old data too thin; test the new cards vs 72 h peaks)
- 2026-10-05 | built deepread (service): Opus reads site, docs, app wallet check, GitHub, token and hook code family (own vs template), last X posts, 5 nearest earlier projects, what is hot; card = idea, family, newness, crowding, demand fit, product stage, build, meta 0-10; copies reuse the card; ~$0.016 per coin incl. copies; page http://159.223.88.168/hx-c9ois/lanes/metas.html; guarded by judgeguard + supervisor; daily metatest vs 72 h peak market cap | LIVE (research)

## 5 Oct 2026 10:20 BKK: project reader v2
- [stated] Ben: run the AI read as soon as possible, not 3 min after launch; a strong project can be (A) a complete product, (B) one new great feature, or (C) a new product category; a meta cannot be known at launch (no copycats yet), but we can know if the text makes sense, if it is novel or different from current launches, if something works, and on GitHub whether it is their own code, mostly copied, how many people and how long it took; build scores for all of these.
- 2026-10-05 | launch tool vs launch-day 10x (catbase, normal 2.45%) | factory 0x58daec3116aae6d93017baaea7749052e8a04fa7 coins 17.0% (n106) and 7.9% (n1354); factory 0x3194e32622c5d860a0572c74edda99dcfd8fc827 6.4% (n1274); hook 0x4e3468951d49f2eea976ed0d6e75ffcb44a9a544 coins 0.2% (n3257) | tmpl.py | FACT (in id5)
- 2026-10-05 | deepread v2 | read at once (fresh coins 17 to 24 s after launch), re-read at 15 min and 2 h; scores: makes sense, different, innovation A/B/C, product stage, own code (template vs own, verified source, GitHub fork/commits/people/first commit), demand fit, overall 0-10 | LIVE (research; metatest daily)

## 5 Oct 2026 10:40 BKK: coin brain, 2x refocus (doc [[coin-brain]])
- [stated] Ben: only 100%+ moves matter; put all resources into understanding what creates them.
- 2026-10-05 | cowork coin brain: doubles in coins 1 to 11 days old (2x before minus 50% in 3 days, 1,032 episodes, ~74/day) vs matched look-alikes (same hour, volatility, last-hour move, activity) | wallets in 12 h before (fresh buyers, growth, smart wallets from earlier doubles, top buyer share, p2p) AUC 0.49 to 0.55 (634 vs 683); X CA posts 24 h before AUC 0.53 (320 vs 345); X during first 3 h 0.73 vs 0.37 posts but most doubles have none | wallet_feats.py, pull_x.py | KILLED: no pre-move fingerprint for older-coin doubles with chain + X data
- 2026-10-05 | cowork coin brain: 2x bet by last-hour move | double odds rise in both tails (27% after minus 30%, 24% after +60%, base 15%) but every bucket loses 1 to 10% per trade after 3% costs (halving odds rise too) | rows.parquet | FACT: older-coin 2x is close to a fair flip; predictable 100%+ edge sits at launch minute 0 to 1 (round 5 live paper last 24 h: 13 of 19 closed hit 10x)

## 5 Oct 2026 10:45 BKK: own-code check + setup review (doc claude/runner-finder-setup-review.md)
- [stated] Ben: copying is normal (template parts are okay), but it must be assessed whether there is significant OWN code; wants 10 iterations then 50 out-of-the-box ideas to make categories, scores and the whole process optimal for finding big runners early; agrees to keep a short name for each idea; got a free GitHub token (stored on the droplet, never shown).
- 2026-10-05 | BUG: deep.db cards were overwritten by the 15 min and 2 h re-reads (which can see the pump); 336 of 455 first reads lost | cardlog table + trigger keeps every read; metatest uses first reads only | FIXED (first metatest verdict ~8 Oct)
- 2026-10-05 | bytecode md5 under-counts copies: immutables (names, addresses) change the hash, so 16.5k tokens gave 7.1k "distinct" codes | ownscan code shape = opcodes with 5+ byte push data zeroed and metadata stripped | FACT
- 2026-10-05 | 4,657 of 8,662 v4 hooks are one 238-byte beacon proxy pointing at the same logic contract (0xca3eb46cc39c7426307c36bafc97f585c6244884); proxies (EIP-1167, EIP-1967, beacon) are now followed to the code that runs | ownscan.resolve | LIVE
- 2026-10-05 | built ownscan.py into deepread: rare own functions (named via openchain), code shape family, nearest template, proxy logic shared count, Sourcify verified source (Blockscout 403s the droplet; Sourcify supports chain 4663), GitHub own vs library by file fingerprint against OpenZeppelin/Uniswap/solmate/solady/forge-std/Next.js/scaffold-eth (39k files), copied-from-earlier-project check, real authors, weeks active, biggest-week share; plus domain age (RDAP) and Solana/Base trending names fed to the AI | LIVE (research; token code corpus still filling from the Mac IP, public RPC 429s the droplet)

## 5 Oct 2026 10:40 BKK: TwoClocks / SecondLegEntry / HoldLonger test (qualdip.py, /opt/hist/analysis/qualdip.txt)
- [stated] Ben agrees: a meta shows up when something works and others copy it; asked for the next to-do step.
- 2026-10-05 | setup: solid text score trained only on coins launched before 12 Sep (alive now on DexScreener, 3,760 coins), applied to 3,193 alive pools launched 16 to 27 Sep; hourly paths from day DBs to 7 d; entry next-hour open; 3% cost; 17 broken price paths dropped, single coin capped at 50x | qualdip | METHOD
- 2026-10-05 | HoldLonger / TwoClocks: 10x within 7 d from the hour-1 entry: quality top 20% 4.6% (4 of 87) vs middle 1.1%, bottom 0%, no text 0.8%, all 0.9%; BUT all 4 hits in 16 to 21 Sep (8.0%, n50) and 0 of 37 in 22 to 27 Sep; medians -3% everywhere | qualdip | LEAD, NOT STABLE (re-test on Opus first-read cards via metatest)
- 2026-10-05 | SecondLegEntry (buy the first close 50% under the running peak within 48 h): all coins n230, 10x after entry 0.9%, avg -1% without the top 3 prints; quality top 20% n23 avg -48% without top 3, tp2/sl50 win 21.7%; quality does not rescue dip buys | qualdip | KILLED

## 5 Oct 2026 11:10 BKK: Ben's correction + coin-brain launch review checked
- [stated] Ben: we do NOT buy strong projects an hour after launch, and do NOT buy first 50% dips. We focus on quick detection and buying in the first seconds (always within 30 s). Later maybe a slower take-profit: e.g. sell 50% at 10x and hold the rest based on what we know about the project (if it could be a real runner, take less profit).
- [stated] Ben's new 10x ideas: retrain every hour (10x results known within ~1 h), read the raw first 20 s of trades directly, check if the builder test-launched the same code days earlier.
- 2026-10-05 | exit question (bought at minute 1, launches 16 to 27 Sep, 781 active coins, 52 hit 10x) | after 10x: 40% went on to 30x, 8% to 100x; rest half held to day 7 median 2.9x, rest half with 50% trailing stop median 7.5x (avg 10.8x); quality split n=4 to 5 per group, no conclusion | qualrun.py | FACT: trail the rest, do not just hold
- 2026-10-05 | 777 coins CONFIRMED broken: 67 open round-5 picks (29 tokens, 64 catch-up + 3 live) had entry 3e-39; cause = a zero-amount swap that pushes the pool to the max tick (only "trade" seen), all from factory 0x58daec3116aae6d93017baaea7749052e8a04fa7 (302 of ~1,200 of its launches end in 777) | voided (status void, exit bad_price); pickers + reconcile now refuse to book on a price that is not a real trade ("bad price skip", pickguard marker) | FIXED
- 2026-10-05 | live score bias CONFIRMED: round-5 DB oldest pick 17.6 h, 154 closed vs 371 open; closed-only counts winners first | scoreboard.py hourly (matured 24 h cohorts only; young shown as a floor with open = not won), public http://159.223.88.168/hx-c9ois/lanes/scoreboard.txt; sentinel leak check now counts all picks in 48 h | FIXED (first matured cohort ~6 Oct)
- 2026-10-05 | floor check with own data (12,264 active ETH launches 16 Sep to 3 Oct, entry at first trade + 30 s) | entry 5 to 10x above launch: 10x rate 2.1% vs 2.8 to 7.1% below 5x, worst price 0.17 of entry; 3 to 5x: worst 0.31 | floorchk.py | CONFIRMED for >5x; below 3x the "never drops" numbers look like an artifact (likely hook-curve pools whose pool price does not move), not trusted

## 5 Oct 2026 ~12:30 BKK: launch runners round 2 (read-only review; doc claude/launch-runners-round2-2026-10-05.md, public http://159.223.88.168/hx-c9ois/brx/launch-runners-round2.md)
- [stated] Ben: look for patterns of coins that go to $500k, $1M or multi-millions right after launch.
- 2026-10-05 | runner label (sellable $500k+ AND 10x from the 30 to 40 s entry), 15,149 clean launches 21 Sep to 3 Oct (mcap.db + feat.db) | 148 runners (1.0%), 111 pops (10x but under $500k); most "$1M+" coins just launched big (only 50 of 137 got there as a 10x) | caps study | FACT
- 2026-10-05 | the pond: native ETH pool, 0.01% fee, 10 ETH start cap (feat.db pad 0x0000000000000000000000000000000000000000|100) | 1,476 launches, runner rate 7.25% (every day 3 to 13%) vs 0.3% elsewhere, holds 72% of all runners | caps study | STRONG
- 2026-10-05 | runner vs pop at 30 s | start cap AUC 0.86, distinct buyers 0.78, one wallet holding the ETH 0.31 (pops = one wallet pumping a tiny pool) | caps study | FACT
- 2026-10-05 | runner-label model vs 10x-label model, walk forward 9 days, top 1% (106 picks each) | runner 13.6% vs 8.9%, $500k+ 17.4% vs 12.0%, median sellable peak $141k vs $69k; better 5 days, equal 2, worse 2 | caps study | LEAD (proposed as a runner book)
- 2026-10-05 | inside the pond the 30 s tape adds little (AUC ~0.6, top 10% 9.4% vs 8.1%) | caps study | FACT: extra edge must come from identity, deployer, text, minute 1 to 10 flow
- Warning: mcap.db holds ~7 min per coin, its "last" cap is not an exit; no EV reported.

## 5 Oct 2026 11:15 BKK: three 10x-lane tests (walk-forward, feat.db 30 s features, broken 777 prices removed)
- 2026-10-05 | skip buys > 5x above first trade price (pumpskip.py, 13 days out of sample) | model top 5%: picks with pump > 5 are only 4% of picks, 10x 11.1% vs 26.0%, EV +0.89 vs +3.76 per $1; skip rule lifts the average slightly in both halves, BUT those picks are still EV positive, so skipping lowers total profit (~45 picks x +0.89 x $200 over 13 days) | FACT: do not skip, size them smaller or feed pump to the model
- 2026-10-05 | retrain every 2 h vs daily vs frozen 4 days (retrainfreq.py, 10 test days) | top 5% 10x: 2 h 30.2%, daily 26.3%, frozen 22.5%; top 2%: 43.5% / 39.1% / 34.2%; 2 h better on 7 of 10 days | STRONG (Ben's idea confirmed). Blocker: day DBs only complete after midnight, live models (aimodel5.pkl) are from 4 Oct 13:45; needs a rolling fresh-row feed | NEXT BUILD
- 2026-10-05 | builder test launch, same OWN code shape launched before by the same creator (testlaunch.py, 27,458 launches) | n77, 10x 6.5%, no lift over own code in general | KILLED
- 2026-10-05 | NEW at second 0: own code shape (fewer than 20 EARLIER contracts with the same skeleton, point in time) 10x 8.9% (n1,000) vs launchpad template 3.2%; stable both halves (8.9% / 8.6%); with creator's first launch 10.1% (n761); creator's first launch 5.3% vs creator reusing a template 0.3% | testlaunch.py | STRONG: add to id5 v2 (one eth_getCode, already fetched by idlive)

## 5 Oct 2026 ~12:45 BKK: runner vs pop explained (same caps data)
- 2026-10-05 | launch caps are NOT all ~$20k: cap at the 30 s buy, 10% under $10k, median $41k, 10% over $100k, 125 coins already over $500k | caps study | FACT
- 2026-10-05 | of 137 coins that reached a sellable $1M, 75 were already $500k+ at the 30 s buy (median $622k); only 50 did a 10x from our buy | caps study | FACT (big launches are 1x to 2x bets, not 10x)
- 2026-10-05 | CORRECTION of "runner model better for Ben": pops pay the SAME multiple (median sellable 18x: $10.5k to $196k) as runners (16x: $53k to $846k); pops = tiny start (2.5 ETH), 4 buyers, one wallet 87% of first 30 s ETH (0.31 ETH in); runners = 10 ETH start, 14 buyers, top wallet 34% (1.73 ETH in); runner model hit 10x LESS (15.4% vs 17.0% in top 1%) | caps study | CORRECTED: runner label wins only if the goal is $500k+ caps or bigger exits; for 10x on a small ticket the 10x model is as good or better

## 5 Oct 2026 12:00 BKK: what makes RH runners go big (cowork; http://159.223.88.168/hx-c9ois/runners/runner_patterns.md; project doc claude/runner-patterns-2026-10-05.md)
- [stated] Ben: study past runners by tier ($500k, $1M, $2M, $5M, $10M, $20M+, $100M+) within about a week; stay open to bundled/cabal/smart-wallet runners; prove everything on real cases; do not edit anything already built, only propose ideas and code for him to hand over.
- 2026-10-05 | ladder, 7 days after crossing (runtrack hourly klines, ATH = 2nd highest close) | from $500k (n577): $2M 39%, $5M 18%, $20M 6%; from $1M (n423): $2M 56%, $5M 25%, $20M 8% | ladder.py | FACT
- 2026-10-05 | speed decides size | $1M within 1 h of first trade: $5M 38%, $20M 15%; 6 to 24 h: 17% / 2%; 3+ days: 16% / 3%; straight-up path to $1M: 39% / 18% (AUC for $20M stable in both date halves) | ladder.py | STRONG
- 2026-10-05 | X around the $1M crossing (n356, twitterapi.io) | 10+ authors BEFORE: $5M 7% vs 29% with none (quiet before is good); 10+ authors in the 6 h AFTER: 34% vs 15% | xq.py | STRONG
- 2026-10-05 | runners vs 1,380 random traded launches (first 30 min) | runner-alumni wallets in first 5 min AUC 0.79; LESS first-block bundling; top10 share 12 to 18% vs 48%; USDG quote 14 to 25% vs 2%; USDG pad hook 0xe5e702641ea86f4ae6cc3cdaed2b886f976be044 11 to 18% vs 2% | feat3/an4 | FACT
- 2026-10-05 | small vs big runner at launch | no launch-minute feature separates them (AUC 0.44 to 0.56); bundle/cabal idea REJECTED; two species of $20M+: stealth/outside tokens (7 of top 11) and crowd launches on ecosystem pads | an4 | FACT
- 2026-10-05 | proposed milestone_dive.py (read-only, not deployed): score at the $500k/$1M crossing from speed, straight-up path, X before/after, top10, stealth start | STRONG reached $20M 17%/16% in both halves vs SKIP 3%/1% | PROPOSED (Ben hands over)

## 5 Oct 2026 12:00 BKK: build iterations
- 2026-10-05 | id5 v2 test (idv2test.py, same rows/model, 6 walk-forward days): adding own code shape + creator history as counts made it WORSE (top1 41.4% -> 28.8%; counts grow with time = drift); as two yes/no flags: top1 36.6%, top2 30.3% vs 31.7%, top5 22.4% vs 22.0% | KILLED: the raw 8.9% vs 3.2% own-shape lift is already inside id5 (launch tool and code family features)
- 2026-10-05 | finding: train5 trained live models on all days EXCEPT the newest (newest day only set the pick lines), so live models lag ~2 days | fresh-fit version written (train5_new.py: newest day learned, lines from out-of-fold scores) | TESTING (t5rank: old vs new, 3 unseen days)
- 2026-10-05 | Anthropic API credit ran out 10:23 BKK: deepread cards and the lane radar AI judge refused (400 credit) since; judgeguard alerted at 10:24 | OPEN: needs Ben to top up (auto-reload recommended)

## 5 Oct 2026 ~13:30 BKK: launch runners round 3, patterns second by second (read-only; doc claude/launch-runners-round3-2026-10-05.md, public http://159.223.88.168/hx-c9ois/brx/launch-runners-round3.md; per-second export http://159.223.88.168/hx-c9ois/brx/persec.csv.gz)
- [stated] Ben: find the patterns of runners ($500k, $1M, multi-million) at second 0, 5, 10, 15 and so on; if they are already at $500k early, find their pattern and buy earlier; detection speed is not this session's focus; use percentages.
- 2026-10-05 | coins already $300k+ at 30 s (n112) | 88% were $300k+ at second 0: launch cap ~$4.5k, creator buys a median 9.9 ETH in the launch block (55% 5+ ETH); only 4% go another 10x | persec | FACT: not buyable earlier, skip
- 2026-10-05 | second 0 pattern (runner = $500k+ AND 10x from the 5 s price, base 2.0%, 14,134 launches 21 Sep to 3 Oct) | fee 0.01% + 10 ETH start: 9.8% of launches, 11.4% runners (early 9.9%, late 13.6%), 56% of all runners; native ETH version 18.3% (late half only, n153); fee 1% 1.2%; fee 0.25%/0.05% 0.4% | it4 | STRONG
- 2026-10-05 | inside that pond, seconds 1 to 30 | 5+ buyers by 5 s 24% (3% of pond); at 15 s 2+ ETH in 19%, price +50% 17%, 13+ buyers 15%, no trades 6%; 4+ sellers at 30 s 3%; after 20 s flow is flat | it5 | LEAD
- 2026-10-05 | waiting cost | runners still 10x away: 87% at 0 s, 72% at 5 s, 50% at 20 s, 42% at 30 s; model top 5% hit rate flat 12 to 15% from 0 s to 30 s (time split 60/40) | it3 | FACT: the pattern is mostly known at second 0; speed keeps the multiple
- 2026-10-05 | fresh-fit vs old train5, 3 unseen days, same pick share (rank): v5 top1 28.0% -> 40.0%, top2 24.0% -> 31.6%, top5 20.0% -> 22.1%; blend top2 21.4% -> 31.6%; base top1 24.0% -> 36.0%; ev top5 10.6% -> 12.3%; new better on all 12 cuts | t5rank | LIVE 12:04 BKK (train5.py now fresh-fit; old kept as train5.py.bak-before-freshfit)
- 2026-10-05 | pick lines drift day to day: at "top 5%" both old and new let 6.8 to 12.5% of coins through on 3 Oct | live over-pick guard in aipick5: line = max(trained line, the live model's own 24 h score quantile), only raises, active after 300 scored coins (~5 h), marker in pickguard | LIVE 12:01 BKK
- 2026-10-05 | still open for true 2 h retraining: day DBs complete only after midnight; plan = live picker logs full model inputs per coin + a label job at +70 min, then retrain on history + today's rows every 2 h | NEXT

## 5 Oct 2026 12:30 BKK: 100x runner lane (cowork; brief http://159.223.88.168/hx-c9ois/brx/hundredx-brief.md; droplet /opt/hundredx; report http://159.223.88.168/hx-c9ois/hx/report.html)
- [stated] Ben: count a 100x only if the coin holds $2M to $3M market cap through a full day; true 100x are rare (~7 in 20 days, 2 or 3 fakes from near-zero start); build a case file (website age, GitHub before launch, X age, own code) as a checklist, not a model; buying day 1 to 3 after proof still leaves room; wants a project deep dive task list (checklist for multi-M runners, when it starts, time budget).
- 2026-10-05 | runner count (runtrack hourly klines, launch seen, 100x from first price + $2M held 24 h + $20k bought at 100x price) | 82 since Jul (59 at $3M bar); last 20 days 14 at $2M, 9 at $3M incl 2 near-zero fakes (COUPONS, second VRAX) = ~7 real | runners_all.csv | FACT
- 2026-10-05 | room after proof | runners first close >= $300k at median hour 1; median sellable room then 15x, +1 h 8x, +6 h 6x, +24 h 3x (23% still 10x+); all 500 coins crossing $300k: 10x odds 12.6% at cross, 7% at 24 h | timing.py | FACT (deep dive verdict within 60 min of $300k)
- 2026-10-05 | GitHub work 14+ d before launch, 20+ commits, linked from own site/X (9 coins: MONEY, MONITOR, WORM, BISCOTTI, VOXEL ran; SFX, HD, SPRL, O1BOT did not) | 10x more from day 1-3 entry 27% learn / 25% test vs 6% / 4%; 5 of 9 became 100x vs ~29% base | hxcollect.py | LEAD (n9)
- 2026-10-05 | old website (domain 7+ d before launch) | 3x-left 6% / 11% vs 29% / 18% rest; old X account runner share 28% / 18% vs 41% / 22% | KILLED (fresh accounts did better on RH)
- 2026-10-05 | own token code (ownscan rare functions) | runner share 26% / 10% vs 39% / 23% template | KILLED (RH token is the launchpad's)
- 2026-10-05 | big website 15KB+, docs link, X day-0 unique posters | no edge for day 1-3 entry | KILLED
- 2026-10-05 | X 10k+ accounts on day 0 | 1-4 big accounts 0 of 82 entries did 10x (both halves) but 0 big accounts 3 of 18 | TESTING (recorded, not a gate)
- 2026-10-05 | paper book LIVE | /opt/hundredx/book.py every 10 min (hx-book.timer + cron, flock), bookguard.py (timer + cron, reruns, alertgate only after 2 failed fixes); runtrack $500k universe, marks at 24/48/72 h, buy = GitHub 14d sign + $40k hourly volume, exit 100x or day 14; book.json public; no Telegram until it beats random on new weeks | LIVE (paper)
- Gaps: holders only for 54 of 285 study coins; runtrack starts at $500k (coins $200k-500k that never grew are missing, so hit rates slightly high); GitHub sample needs the Jul-Sep rebuild and $100k+ day-1 coins. First live read ~12 Oct.

## 5 Oct 2026 12:25 BKK: miss diary + check of coin-brain "launch runners round 3"
- [stated] Ben: add a list of misses: every coin that ran past $1M that we did not buy gets one line on why we missed it; misses teach more than hits.
- 2026-10-05 | built missdiary.py (daily in retrain.sh after mcap.py; public http://159.223.88.168/hx-c9ois/lanes/misses.txt) | runner = new launch under $1M at the 60 s buy whose SELLABLE cap later reached $1M; live part = picker records + one reason; history part = walk-forward 30 s model | LIVE (first live day appears when the 4 Oct day file is rebuilt)
- 2026-10-05 | first history read (last 7 complete days, 47 runners): 17 a 30 s model would have bought, 19 ranked too low, 11 jumped in the first instant | missdiary | FACT
- 2026-10-05 | lane tracker 72 h peaks are NOT usable for misses: 41 of 42 "$1M runners" were new pools of old coins (MOO, AI, BONER, stock tokens) or broken prices (trillions) | FACT
- 2026-10-05 | check of the 0.01% fee "pond" (our day DBs, 25,095 launches 21 Sep to 3 Oct, 10x from the 5 s price within 6 h) | fee 0.01%: 15.8% (both halves 15.7 / 15.9) vs 7.2% all; from 30 s 11.3%; fee 1%: 5.0%; BUT only 3 of 2,242 0.01% pools had 5+ ETH bought in the launch block (their "10 ETH start" is probably liquidity or a differently measured block, not a buy); 5+ ETH launch-block buys any fee: 1.7% (agrees: never chase a launch-block jump) | pondchk.py | fee part CONFIRMED, 10 ETH part NOT confirmed
- 2026-10-05 | v4 Swap amounts are from the SWAPPER side: a buy shows ETH negative, tokens positive (632 of 632 launch-block swaps) | FACT (trap)
- 2026-10-05 | paper book rule_fee001_first in aipick5: every 0.01% fee launch booked at the first trade the picker sees (within 15 s) | LIVE 12:19 BKK (paper)
- 2026-10-05 | Ben said credits were added; API still answers 400 credit 10 min later | OPEN (check the key's organization / workspace)

## 5 Oct 2026 12:45 BKK: where 0.01% fee launches come from, and how to be faster
- [stated] Ben: miss = any coin we could have bought under ~$100k cap at launch that went to $1M (10x from first sight / from the moment our score said no); look into the ones ranked too low.
- 2026-10-05 | source: 0.01% fee pools are NOT a launchpad: creator deploys an ERC20, then calls the Uniswap v4 PositionManager 0x58daec3116aae6d93017baaea7749052e8a04fa7 multicall (2,277-byte calldata, Permit2, one-sided position at a 10 ETH start cap) from thousands of different wallets; 78% WETH 0x0bd7d308f8e1639fab988df18a8011f41eacad73 quote, 22% native ETH; spam farms mixed in (e.g. many "Settle" tokens); "factory 0x58daec" in earlier studies = this PositionManager; "10 ETH start" = start cap 10 ETH, not a buy | feesrc*.py | FACT
- 2026-10-05 | runners (buyable under $100k at 60 s, later sellable $1M+), 24 Sep to 3 Oct: 30 of 11,905 coins; 0.01% fee = 25 of them (83%), runner rate 2.0% vs 0.05% for other fees (late days 3.3%); 30 s flow filters (buyers, ETH in) add little (2.1 to 3.0%) | feerule.py | STRONG
- 2026-10-05 | missed runners (walk-forward 30 s model, 10 days): 10 caught in top 5%, 19 ranked low; 79% of the low-ranked are 0.01% fee too; they were quieter by 30 s (median unique buyers 14 vs 19, ETH in 1.6 vs 2.5, price 1.2x vs 1.7x) | lowrank.py | FACT: buy-all-0.01% book (rule_fee001_first) covers them
- 2026-10-05 | speed: first swap after pool creation median 0.9 s (52% within 1 s, never the same block); our picker sees launches ~3 s late (polling); Robinhood Chain sequencer feed wss://feed.mainnet.chain.robinhood.com streams every sequenced tx live (~10 msgs/s): WORKS from the Mac, 403 from the droplet; RPC websocket not offered | wstest.py | NEXT: feed listener to see pool creations instantly
- 2026-10-05 | missdiary runner definition changed to Ben's: buyable under $100k at the 60 s buy, later sellable $1M+ and 10x+; history: 28 runners in 7 days, 10 a 30 s model would have bought, 18 ranked too low | LIVE
- 2026-10-05 | Anthropic credits back 12:35 BKK; deepread 103 cards in 10 min, all OK | FIXED
- 2026-10-05 | pond size and outcomes (0.01% fee + 10 ETH start cap, 21 Sep to 3 Oct, 13 days, coins with a 60 s entry cap) | 1,446 launches (~111/day) vs 13,790 others; median cap at our 60 s buy $50k; 10x from the 60 s buy (sellable) 8.1% (117) vs 1.1%; reached sellable $100k 53.7% (777); sellable $1M 2.6% (38) vs 1.2% | pondstats.py | FACT
- 2026-10-05 | [stated] Ben: yes, build the feed watcher. Built feedwatch.py on the Mac (~/feedwatch, caffeinate, relay.sh copies to the droplet every 30 s and restarts it): decodes the sequencer feed (Arbitrum batch kind 3 / signed tx kind 4, RLP), finds every v4 pool initialize (PositionManager.initializePool 0xf7020405, PoolManager.initialize 0x6276cbbe) inside any tx; feedcmp.py (droplet, every 10 min) compares feed time vs block time and first trade | LIVE (paper, research); depends on the Mac being on

## 5 Oct 2026 15:10 BKK: 100x lane v2 (report http://159.223.88.168/hx-c9ois/hx/report.html v2)
- [stated] Ben: drop the full-day $2M hold; a 100x counts if 100x from first price AND $20k+ of buying at $2M+ (sellable); "$300k in hour 1 then a few x left" is the filter, not a pattern; wants patterns in first seconds, minutes, hour, hours; STOP all live/paper until research is reliable and bulletproof; 5 research iterations then 50 best ideas, first principles, assume the approach is wrong, argue from gaps; then deep dive input again.
- 2026-10-05 | paper book STOPPED (hx-book/hx-bookguard timers disabled, cron removed) | DONE
- 2026-10-05 | runner count, new rule | 107 since Jul, 24 in last 20 days (~1/day vs ~2,000 launches/day) | FACT
- 2026-10-05 | COUPONS and VRAX copy were real runners (peaks $10M, $88M); only their 100x multiple was inflated by a tiny first trade | CORRECTED
- 2026-10-05 | time to sellable $2M | 10% within 1 h, median 9.5 h, 75% 52 h, 90% 8 d | FACT
- 2026-10-05 | $300k in hour 1 | ~190 coins/day, 0.3% become runners | KILLED as signal
- 2026-10-05 | first $500k close (546 coins) | 27% later sellable at $2M (learn 36%, test 19%); lower churn vol/mc 1.25 vs 1.57 (AUC 0.41/0.38); recent runners AUC 0.52/0.71; longxyz 46% vs pons 30% | LEAD churn, TESTING regime/pad
- 2026-10-05 | ride first $500k close, sell 4x or -50%, 14 d, 3% cost | 1.49x learn, 1.13x test; hot chain + low churn test 1.56x n25 | TESTING
- 2026-10-05 | minute data Sep 15-Oct 4 | only 5-10 clean runners; ~half of runners trade in pools the per-pool data misses (multi-pool, USDG/AI/SPCX quotes, non-v4 pools); low top-holder share was an artifact | KILLED; data fix first
- 2026-10-05 | early buyers who were early in past runners, 3+ in first 10 min | 0.21% vs 0.05% (4x), 96 coins/day, n4 | WEAK LEAD
- 2026-10-05 | GitHub 14+ d | 5 of 82 runners (6%), 5 of 8 with it ran | LEAD (rare)
- 2026-10-05 | deep dive input v2 | trigger first close >= $500k (~6/day), verdict within 3 h (room 8.5x to 7x), recheck 6 h and 24 h; checklist: sellable, churn < 1.3, chain regime, GitHub, original/meta leader, fresh money per hour, snipers/dev, smart early buyers, X per 15 min, one-sentence idea | SPEC
- Next (live off): token-level data fix, finish Jul-Sep rebuild, start saving fomo trending + X per 15 min now, retest top ideas on halves.
- CONTINUED in [[trading-edge-hub-5]] (5 Oct 13:15 BKK on: sell system v1, liquidity pulls make most paper exits fake).