2026-10-20
Trench lessons — 2026-10-20
Key takeaways
- Alpenglow switch: Anza now says ~9:00 AM CT Wed Oct 21 (slot 459,653,000); what changes for traders around it
- White Whale Labs, "The Audit Of The Memecoin Casino" (Oct 19): copying swarms 2 min late loses 41%/coin, "good" solo wallets lose 20%/trade out of sample, −40% stops filled at median −97%
- Platform watch: nothing breaks
Alpenglow switch: Anza now says ~9:00 AM CT Wed Oct 21 (slot 459,653,000); what changes for traders around itevidence: B official + live chain; C trackers
- What new guidance says (Triton, updated Oct 20; BAM blog):
- Preconfs and BAM: Harmonic preconfs are off from ~1 h before the boundary until the first incremental snapshot. JitoBAM block packing and preconfs are off from ~12 h before activation to ~12 h after (Triton's estimate, keyed to 13:20 UTC: ~8:20 PM CT tonight to ~8:37 PM CT Wed). BAM's own post says Block Engine bundles keep working as usual; only BAM-native MPP and DBI go away.
White Whale Labs, "The Audit Of The Memecoin Casino" (Oct 19): copying swarms 2 min late loses 41%/coin, "good" solo wallets lose 20%/trade out of sample, −40% stops filled at median −97%evidence: A-size dataset, self-published; C for verification
- Data: every pump.fun and Stonk launch trade since late August: 384M swaps, 7.7M wallets, >1M pools. The author's database isn't public, and the method is described only in the post and replies.
- Base rate: of 22.8M first buys (2 weeks), 17% doubled before a −40% drop. Mechanical 2x-or-stop lost ~19% per attempt after costs.
- Clusters: swarms of 450–2,000 wallets (plus tight groups of 7–50) enter in the first 1–2 min; 9–33 groups a day, with members rotating. Their coins double 28–46% of the time, and their entries hit 2x-before-−40% 50–60% of the time (vs 17%).
- Copy tests (2 min behind, 2x-or-−40%, out after a day): big swarms −41%/coin over 93 coins (median −95%); positive-record groups only, −11%/coin; one "good" wallet, −20%/trade. Only small tight groups (7–50 wallets) were positive: ~+10%/coin over 16 coins (author: "not enough to call it").
- Neural-net picks: 2x rate 26% vs 10% base, but every one of 8 days lost money (−11%/pick, 771 picks).
- Two replies that matter more than the headline:
- "The −40% stop is a rule, not a fill… half the trades stopped out, 80% of those stops filled below −90%, median stop fill −97%" (measured on 1-minute candles).
- Wallets whose records held up were not block-0 snipers: median entry was buyer #390, ~10 min in, but they enter together, hundreds at a time. "Solo wallets with great records mostly decayed out of sample (3.4M wallets, luck finds you a 60% hit rate over 15 trades)."
- Also: Stonk coins (transfer tax) show far fewer cluster fingerprints than pump.fun.
Platform watch: nothing breaks
- Jito: Block Engine bundles unaffected through the switch (see §1). No tip-floor or API change found.
- Raydium LaunchLab, GMGN, Axiom, Photon, Birdeye: nothing that touches execution (GMGN trading competition; Birdeye teaser for Oct 21). No new Token-2022 transfer-fee incidents found for Oct 19–20.
Sources
- https://x.com/anza_xyz/status/2112520675365900505
- https://x.com/anza_xyz/status/2112246102506565825
- https://x.com/anza_xyz/status/2112246167602081918
- https://solana.com/upgrades/alpenglow
- https://www.anza.xyz/alpenglow
- https://blog.triton.one/alpenglow-migration-avoid-the-misinformation/
- https://blog.triton.one/alpenglow-and-your-triton-stack-what-to-update-before-mainnet/
- https://bam.dev/blog/update-on-bam-readiness-for-alpenglow-on-mainnet/
- https://x.com/davidrhodus/status/2112241173003948474
- https://pumpkinspool.com/watchtower/alpenglow
- https://x.com/MaxUrPointsHQ/status/2112545000244617590
- https://x.com/WhiteWhaleLabs/status/2112213836098908460
- https://x.com/WhiteWhaleLabs/status/2112265145963647155
- https://x.com/WhiteWhaleLabs/status/2112240920389403058
- https://www.tokenpost.com/news/investing/35409
- https://github.com/pump-fun/pump-public-docs/commits
- https://x.com/a_milz/status/2112286532795269621
- https://x.com/MeteoraAG/status/2112505606401909057
Evidence grades: A = large dataset, B = official docs, C = vendor/thread/small sample. Research notes for education only. Not financial advice.