Recommended state
What Rook thinks the system should look like by early afternoon today, item by item, with what is already live, what ships this morning, what runs in shadow first, and the few calls that are Thomas's.
Recommended state
Rook, updated 2026-09-06 11:36 Beirut. Synthesis of the four lane reports, the Experiment A report (REPORT-experiment-a.md) and the tuning sweep (SWEEP.md, now folded in and applied to the live gate) and everything shipped since 05:29. Status words: live = running now, verified; ship = config or small code this morning, before the US session (about 16:00 Beirut); shadow = new paper book first, live only after n is honest; decide = only Thomas.
0. The read in six lines
- The live book is rugs. All-time live is -$207 over 19 lots; the five liquidity pulls are -$250; everything else is +$43. The gate that refuses rugs is the P&L, not a safety feature. It is live since 10:27.
- Nothing else on the board loses money on median. The heat and chase guards are the two best rules we have. Do not loosen them.
- Exits bank about 30% of what the tape offers, and the single mechanical fix is the first clip at +40 instead of +60, because the trail only arms after a clip. The stepped trail shipped last night comes out.
- Execution spends nothing. Free Alchemy plus the push path has 1.75x headroom. The seconds left are our own code, and the websocket cutover is worth 3.6 s for zero dollars.
- The runners we never saw are the gap, not the ones we refused. Twelve tokens ran +100% overnight with no proven-wallet buy; ten are invisible to the Fomo feed too. Reading block-0 buyers is the only earlier population and it is a paid-logs question.
- Experiment A passes its kill line on paper but on one leg. The +60%-in-60-min leg has never fired because the tape worker is stuck. Fix that before the 23:00 call means anything.
1. Entry gate (what we refuse before spending a ticket)
| item | value | status | evidence |
|---|---|---|---|
| liquidity/mcap ceiling | 0.82, across the whole 2 to 60 min window (was 0.85, before that 0.80 under 10 min only) | live 11:36 | 0.80 / 0.82 / 0.85 are a dead zone: identical refusals, runners and dollars on the tape. 0.82 is the only value that centres the one clean separation the tape gives, nearest runner BUSINESS 0.793 (3.3% below) to nearest rug PEZZED 0.862 (5.1% above). 0.85 left PEZZED only 1.4% of headroom. SWEEP.md section 4. |
| wallet rug rate | 0.18 with >= 10 entries (was off since 11:00; before that 0.15@8 live 10:27 to 11:00) | live 11:36 | The final sweep flips the 11:00 call: minWalletEntries was the wrong number, not the rate. 0.15@8 cost 8 of 14 runners for zero extra rugs; 0.18@10 costs 2 (AJC, PETOI) and is the money-best point on the tape (class 7: -101 paperG2, -226 live saved, all 6 rugs still refused). It hands @fr3ak (and hoodrat, +$43) back and blocks only @seralberttrades (3/3 pool losers). frankdegods stays on the handle deny list; the burned-wallet rule stays. |
| ceilings, tuned | ratio 0.82 / underAgeMin 60 / wallet 0.18@10; underAgeMin is the lever that catches rugs (10 catches 3, 15 catches 5, 20+ catches all 6) and 60 is free margin | live 11:36 (SWEEP.md read and applied to follower.json + rules-v2.json Gold v2) |
Sweep grid data/tuning/sweep.json, 1044 rug-gate points + 6 age points, replayed on 1422 feed rows. Pareto front in SWEEP.md section 3, every lost runner by name. The one uncovered hole: PEZ re-signals at 0.681 / 26 min after an LP top-up; no grid point refuses it, live only the per-token cooldown blocked it. |
| handle deny list | frankdegods | live 06:25 | manual, superseded by the rug-rate gate but harmless. |
| burned wallet | 24 h skip after a liquidity-pull close | live 05:53 | burnedWallet() in follower.mjs. |
| max open lots | 2, and a drained lot must not count against the cap | ship (code, small) | positions report 6: blocks exactly PEZ (-$50), costs no winner on the tape. Without the drained-lot exclusion two unsellable lots jam the book. |
| pre-buy tells, log-only | WETH quote asset; pre-entry sells/buys under 0.30; raw contract creation (no factory); hand-written 8.7 KB owner bytecode | shadow as columns on every signal and a paper twin Gold v2 tells |
each catches 3/3 night rugs with 0/36 false positives, but n=3 from one operator. Log for a week, then decide. |
| heat ceiling +500%/1h, chase 25% over the wallet fill, cohort <= 1, mcap under $1M, liquidity $25k to $400k | unchanged | live | capture report 2: heat refusals median -70%, chase -73%. Mcap cap refused the most rows at median -7.5%, raising it buys volume not edge. |
| age ceiling | 60 min live; 180 min as a paper twin Gold v2 age180 |
shadow | capture report: the only loosening the night supports, and only with heat and chase intact. Not live until the twin has 30 entries. |
2. Exit ladder (what we do with a lot once we hold it)
| item | value | status | evidence |
|---|---|---|---|
| first clip | 50% at +40 (was +60) | ship this morning | positions report 2: the trail arms only after a clip (follower.mjs around line 281). 22 of 39 calls never saw +60 and are worth -876 points together. Leave-one-out keeps the gain between +120 and +219 points. |
| second clip | 25% at +200 | live, keep | |
| trail | flat 40% relative, trailSteps: [] (drop the 25-after-200 step shipped 23:48 yesterday) |
ship | positions report 3: on the swap-tick path the stepped trail is the worst variant tested (-54 vs +346 for a flat 40) once LIQLORACLE is in the sample. My 23:48 change was fitted on 32 calls that did not include it. |
| stop | -45 | live, keep | no variant in the sweep beats it once rugs are removed from the sample. |
| tighten by market cap | no | decided by data | only 4 of 39 calls crossed $1M peak mcap; mcap band and pnl band are the same axis on this tape. |
| tighten at low pnl | trail 25% until the peak passes +67, then 40 | shadow twin Gold v2 lowtrail |
positions report 5: with a relative trail the stop sits below breakeven for any peak under +67. That is where the give-back is. Untested as a full ladder. |
| no-progress 30 min under +10, time stop 90 min, breakeven after clip, wallet exit off | unchanged | live | |
| exit retry | ladder 1000/1000/1500/2200 bps, 10 s cooldown, plus a per-token backoff after 3 failed quotes and a Relay key | ship (code, small) | execution report 6: 508 Relay 429s in two hours were our own retry storm against a 50/min limit. |
| in-position liquidity watch | PoolManager liquidity read on every mark, write off at zero | live 05:53 | it buys truth and a fast write-off, not money. Honeypots refuse the sell before the pull. |
3. Detection and execution (how fast we see and fill)
| item | status | evidence |
|---|---|---|
| Alchemy websocket logs subscription becomes the buy source, poll loop stays as fallback | ship this morning (shadow.mjs, tests, watchtower restart) | saw 119 of 119 feed rows overnight, median 4.0 s before the poller. Detection 4.2 s to under 1 s. |
sequencer-feed detector (feed-watch.mjs tx.from path) |
retire | 2 hits in 3.05M transactions: every proven buy is relayer-submitted, tx.from is never the wallet. |
| one-plan follower build | live since 05:53 (verified: process newer than the file, reusePlanMs present) |
signal to ready-to-send 9.0 s to 1.8 s in the A/B. Not yet measured on a live fill. |
| Alchemy tier | free, plus enable Pay As You Go as an insurance switch ($0 base, $0.45 per 1M CU, uncapped eth_getLogs) |
decide, one toggle |
watch.lastAt liveness sentinel on the crontab |
ship, no LLM, $0 | the 87 min outage was invisible because the process never died. Same shape as the tracker freeze. |
| Relay API key (free) and Kyber as-is | decide (free signup) | raises the 50/min quote limit; the backoff above matters more. |
| target after this morning | wallet buy to our fill about 6 s (from 21 s) |
4. Capture (what we look at)
| item | status | evidence |
|---|---|---|
| proven-wallet feed, 56 wallets, re-ranked with rugs/entries next to every row | live, re-rank in the Experiment A page | 12 of 24 mature wallets fall to the rug gate; the list needs its next 20 wallets from the scout scorer, not from the leaderboard. |
Gold v2 age180 paper twin |
shadow | see section 1. |
| block-0 buyer population (who buys in the first 3 blocks of a pool) | decide: needs uncapped logs (Alchemy PAYG) or an indexer key | the only population that is structurally earlier than a 95-min-late proven-wallet feed. 10 of 12 invisible runners are invisible to Fomo too. |
| Fomo-triggered shadow book | shadow, spec from the Experiment A report | Fomo leads our wallets by 5.4 min median on the same tokens: a latency edge, not a new population. Worth a paper book, not a live one. |
| copycat guard | ship (small): refuse a symbol that already exists in the registry with over 10x the mcap | we bought a copycat PEZ while the real one, +5312%, sat in our own registry refused for mcap. |
5. Experiment A (Bolo's lane)
Report landed: REPORT-experiment-a.md, 11.7 h of tape.
| item | status |
|---|---|
| collector, joiner, crowd | healthy: 7,040 rows, 0 holes, 0 collector errors, 8.6 s median ingest lag, wallet resolution 31% |
| A3 tape worker | stuck: peak60mPct set on 5 of 185 marks; two dead alerts have burned 736 and 731 retries with no ceiling, 92% of marker cycles error, every error string truncated before the child's stderr. That single bug kills the +60% leg of D4, pUp60, and any ladder replay. One fix unblocks the most. |
| the 5.4-min lead is conditional | measured unconditionally against feed.jsonl: +38 min across all 55 overlapping tokens, +13 s on tokens under 6 h, -40 s (proven wallets first) under 60 min. The Fomo edge lives in stale mid-caps, not fresh launches. |
| the bottleneck is the universe, not detection | of 15 young tokens both feeds hit, Gold v2 skipped 12 on its own gates (heat 7, mcap 4, age 2, liq 1, band 1). UNIPCS, best token in the experiment (+112% median over 24 marks), refused for mcap $3.5M. |
| follow-list gap | smartmoney shortlist bluntz_capital, NotARandomUser, macdegods, threem, econoar, FullPinkYak; macdegods and threem are not in FOLLOW-LIST.json yet sit top-4 by m60 hit with zero rugs. ship: add them. |
| scout bug | --before now is a silent zero (Date.parse('now') = NaN); would have frozen an empty scout set at 23:59 tonight. ship (small). Also: 0 traders eligible at --min 2 in 11.7 h, so the scout threshold needs re-basing. |
| D4 kill call at 23:00 | decide: above the line on the proven-within-15-min leg alone, but 144 marks sit on 14 tokens (four carry 115), and the brief's own example behaviour (age <15 min, mcap <$300k, proven buy <5 min) has n = 0 in 11.7 h, so it cannot be tested as written. My recommendation: judge on both legs or extend day one by the hours the tape was dark. |
6. Risk frame
| item | value | status |
|---|---|---|
| ticket | $50 | keep; wallet $509, two open lots = 20% at risk |
| max open | 2 | ship (section 1) |
| daily loss stop | off | keep off: the rugs realize hours after the buys, a realized stop never fires in time. The entry gate is the loss stop. |
| kill switch | touch data/executor/KILL |
live |
| funding | no top-up needed for two lots; a third lot needs about $60 more ETH | decide |
7. Sequencing for this morning (what I do, in order, each with tests and a dry buy and sell before the follower restarts)
- Ceilings set from the sweep grid: ratio 0.85 across 2 to 60 min stays, wallet rug-rate gate off (done 11:00, config only).
- First clip +40, trailSteps empty (config, tests, dry buy and sell, restart).
- Websocket cutover in shadow.mjs, poll as fallback, feed-watch tx.from path retired (code, tests, watchtower restart).
- maxOpen 2 with the drained-lot exclusion (code, tests, restart).
- Exit retry backoff per token (code, tests).
watch.lastAtsentinel on the crontab, tracker build fixed (em-dash guard has failed 771 builds since 21:36 yesterday).- Paper twins added to
rules-v2.json:Gold v2 tells,Gold v2 age180,Gold v2 lowtrail, plus the Fomo-triggered book once the Experiment A page defines it. - Copycat guard.
Nothing above touches the wallet without a dry buy and a dry sell first. Steps 1 to 6 are about three hours. Steps 7 and 8 fit before the US session.
8. The milestone this sets up
Gold v3, 48 hours. Same trigger, rug-gated entry, +40 first clip, websocket detection, two lots. Success at Tuesday 09:00:
- zero
liquidity_pulledwrite-offs on live money; - wallet buy to our fill under 8 s p50, detection under 1.5 s;
- live book positive ex-nothing (no "without the rugs" qualifier);
- every paper twin at n >= 20 so the next config change is chosen from data with LIQLORACLE not driving every sum;
- Experiment A with both D4 legs measured and a scout list of at least ten wallets with n >= 5 each.
9. Needs Thomas, all of it
- Alchemy Pay As You Go toggle (no base fee). Buys uncapped logs for the block-0 population and insurance for the push path.
- Relay API key, free signup, the same dashboard shape as Alchemy.
- A history indexer for Robinhood Chain (Blockscout is Cloudflare-blocked from this box): without it, "who funded these burner deployers" stays unanswerable. Options and prices go on the execution page once I have checked them.
- The D4 call at 23:00: both legs, or extend day one.
- Whether the wallet rug-rate gate should come back in any form. On this tape it adds no rug coverage and costs eight runners, so it is off; say so if you want it as a shadow column instead.