Updates
Running log of what changed on the live book and what landed in research, newest first. Read this page when you have two minutes; the lane pages carry the evidence.
Sunday 2026-09-06, 20:55 Beirut: all four evening workers landed
The four detached workers launched 19:45 all finished by 20:16. Findings below; each has its own page with the evidence.
Refused buys. 122 refusals, 60 distinct tokens, zero live buys today. Almost nothing is tunable: 101 of 122 rows fail two or more gates at once, only 21 fail exactly one. The band and liquidity-ceiling refusals were money saved (the Mid paper book took the same four fresh $1.2M pools we refused and closed all four at liquidity_pulled, -$50 each). The single best knob is the 60 min age ceiling: 17 rows, 13 tokens, +$52 at a $50 ticket, no rug in the set, but median +60 min is -0.3%, so it is a thin two-name edge, not a rule. Two of the day's biggest movers (Investor +1439%, PAIRCAT +680%) were throttled by the ladder, not the filter, which is a positions finding.
Age ceiling. Keep maxAgeMin: 60. Both proposed changes (cut to 15, widen to 180) fail under leave-one-out and across four independent samples. Ceilings 15, 30 and 60 are statistically inseparable on per-entry median; 180 and no-ceiling rank last in every sample. The reason not to cut to 15 is rugs, not returns: the 2 to 15 min bucket carries a 24 to 31% rug rate against 1 to 9% everywhere older. The under-250k age exemption running in the Gold paper set for 2.5 days does not reproduce the 09-04 tape finding (n=31, median -11.1%, sign stable under leave-one-out). Two paper sets are written but rules-v2.json was not edited.
Market scan. The Sep 3 thesis holds and the market moved up inside it: BTC $79.7k, ETH $2,484, SOL $106, stablecoin supply $312B and climbing, best three-week ETF streak of 2026. Liquidity is concentrating onto our chain, not spreading out: Robinhood Chain DEX volume +105.7% w/w while Solana -24%, Base -20%, Ethereum -17%; the chain is now 19.5% of all DEX volume on 24h and TVL is at an ATH $908.6M. Pons V2 has flipped pump.fun ($826M vs $459M 7d launchpad volume). The one honest "elsewhere" with real numbers is BSC (only other chain with rising volume, ports our EVM follower nearly unchanged); second is Hyperliquid copy-trading. Four dated risks inside two weeks: CLARITY cloture Sep 15, FOMC Sep 16 (now a coin flip), the chain gas waiver ending ~Sep 29, and PONS perp funding at 154.7% APR flagging a crowded long.
Experiment A (A3 tape diagnosis). The first-hour outcome leg is dead, not slow: 3,269 tape attempts and 0 successes since the 01:03 build. Root cause is a 15 s child-kill fighting a 64 s RPC retry on wide getLogs ranges, plus head-of-line blocking that put 93% of all tape work on two rows of one pool while 197 candidates were never attempted once. The single up60At we have (6.7e55%) is a drained-pool artifact. Net: D4 (the day-one population kill counted 23:00 tonight) is at 14 of 30 with the +60% leg contributing 0 of 80 unmeasured rows. A kill taken tonight would be a kill on a leg that never ran. Full fix and 10-to-15-minute backfill plan are in the report; evening orders for Bolo written from it.
Live book, right now. No live buy since the 19:48 restart (last real fill was MUMU on 09-05). Gold v2 running clean, no open lots. Wallet-watch failure rate is about 21% of log lines this hour, up from the earlier one-in-six: the Cloudflare-fronted connect timeouts are now joined by RPC 429s. Detection is delayed by seconds and retried next tick, not lost, but I am watching whether the 429s climb.
Sunday 2026-09-06, 19:50 Beirut: quick fixes live, four workers running
Live book (Gold v2, real money). Thomas 19:32 voice: ticket back to $50, quick fixes on my recommendation now, the major lift after he reads the recommended-state page. Shipped 19:48, tests 495 of 495, dry buy clean, follower restarted with no open lots, startup line reads ticket $50.00, maxOpen 2:
| change | from | to | why |
|---|---|---|---|
| ticket | $25 (the 14:00 default) | $50 | his word |
| first clip | +60 x 50% | +40 x 50% | the trail only arms after a clip, so calls that peak +45 had no trail, only the -45 stop. Positions lane: +194 points on bars, leave-one-out +120 to +219, no single call drives it |
| stepped trail | 40, then 25 after +200 | flat 40 | fitted on 32 calls that predate LIQLORACLE; with it in the sample it is the worst variant (-54 vs +346 for flat 40). This was live on real money since Saturday 23:48 |
| open lots cap | unlimited | 2 | blocks exactly PEZ on the frankdegods night, costs no winner. The wallet is ten tickets and the rug losses came in clusters |
Not touched tonight, waits for the read: websocket detection cutover, exit retry backoff, the paper twins, the copycat guard, the age ceiling.
Sunday's tape. Nothing fired live since the three write-offs at 05:54. Since 11:00 Beirut the watch saw about 120 proven-wallet buys and Gold v2 refused every one: 49 under the $25k liquidity floor, 36 over the $1M mcap cap, 17 over the 60 min age ceiling, 8 band, 5 wallet spend, 3 liquidity ceiling, 2 heat, 1 cohort. The rug gate never got a candidate. A quiet Sunday of old or large tokens. The Refused buys page carries what each of those tokens did afterwards once the worker lands.
In flight, landing this evening. Four detached workers, each with a page here that fills in when its report exists:
- Refused buys: every skipped token's path afterwards, by refusal class, and the fine-tuning candidates with n.
- Age ceiling: 15 vs 60 vs 180 min reconciled with leave-one-out, the two lanes that disagreed named.
- Market: the wider market, liquidity, where launch volume is, what else deserves a paper book.
- Experiment A: a read-only diagnosis of why the first-hour outcome leg produces nothing (5 of 211 marks have
peak60mPct), written so Bolo can act on it. Evening orders for Bolo follow from it.
Infra. The Watchtower tracker at watch-rook.pages.dev had not deployed since Saturday 18:36: an em dash in one debrief file tripped the build guard on every one of the 1319 cron runs since. Fixed 19:40, deploying again every minute. This analysis site is now on Cloudflare Pages as well, surge kept as a mirror. The box sees intermittent connect timeouts to Cloudflare-fronted hosts tonight (DexScreener, Kyber, GeckoTerminal, Cloudflare's own API), roughly one call in six; the RPC is clean. The wallet watch logs the failures and retries next tick, so detection is delayed by seconds, not lost. Watching it.
Needs Thomas. Two calls when he is back: the read of the recommended-state page (the major lift), and the Experiment A day-one kill rule at 23:00, which I am deferring as lead until the +60% leg has data (see the Experiment A page).