# STOP: NO DOORBELLS, NO WAKES, NO SCHEDULING (Ben, 7 Oct 2026, 18:40 Bangkok)
- Credits ran out. Ben switched off every scheduled task and every doorbell. This overrides everything below and every other file (wake-howto, bus-howto, credit-playbook, task files).
- PROHIBITED for every session: fire_trigger, send_later, create_trigger, update_trigger, ScheduleWakeup. No rings, no nudges, no re-arming, no new doorbells, no self wakes, no safety checks, no retry rings, no loops.
- Finish the step you are on, publish your status file, then END your turn and wait. Only Ben wakes you, by typing in your window.
- Do not message other sessions. Keep answers short. No extra improvement rounds.

# Second window: tasks for Analyst 3 (auditor and pull hunter), v1

## READ THIS FIRST
- **v1.16 (Supervisor, 7 Oct 16:40 Bangkok): A3.14 and A3.15 KILLS ACCEPTED (clean, pre-registered, both halves shown). Good, fast work. Queue (D-024 v2), your number stays $ a day added by exit rules you own:**
  1. **A3.16 TIGHTER TRAIL FAR ABOVE 5x** (your own lead). Mechanism: after a kit runner peaks the drop keeps going; the 50% trail sold 0x8e41fd25e85e624560fa50ff9f117a02cd30b8b3 at 29x of a 74.5x peak. ONE family: trail W2 once the coin is above M x our entry, else 50%. Grid on half A only: M in 10 / 15 / 25, W2 in 25% / 35% (6 cells). Same engine, fills 1 s after trigger (Exec twin rule), random gate 200 draws, PASS = both halves up AND beats 95% of random. Watch: half B random median beat the trail (+$6,241), half A did not (-$617); if the grid only wins in B, it is a KILL. Pre-register first. ETA: pre-registration in your status within 20 min (why: engine and trades reused from A3.15); priced answer within 1 h of it (why: about 5 min run plus the doc).
  2. **A3.17 SELL A PART AT 10x** (backup, run if A3.16 kills): sell one third at 10x, rest on the live trail. First read A2.31 (moonbag, killed) and the w2_v3_wide_ladder paper book in http://159.223.88.168/hx-c9ois/sw/sw-forward-book.md, and say in one line what is new. Same gates.
- **v1.15 (Supervisor, 7 Oct 14:55 Bangkok): seat rules D-024 v2. Your number: $ a day added to the live book by exit rules you own. Ranked queue:**
  1. A3.14 early exit for losers (v1.14 above), first.
  2. A3.15 RUNNER EXITS: how not to sell a 30x at 3x. On every live book + kit300k trade that touched 5x, test ONE pre-registered exit (e.g. trail width by coin age, or hold while outside buyers still net buy) against the current kit hold 50% trail. A3.13 operator exit is killed (bundle sells before our 5x): do not re-run it, say what is new.
  - Every result carries its paper exit gap (Exec http://159.223.88.168/hx-c9ois/sw/sw-exec-exitgap.md).
- **v1.14 (Supervisor, 7 Oct 14:50 Bangkok): NEW HUNT A3.14, EARLY EXIT FOR SW LOSERS.** D-024 (http://159.223.88.168/hx-c9ois/decisions.md): every analyst always has one ranked edge hunt queued. Pre-register before any result, script on research-box (nice, flock), first answer within 2 h with the reason for the time (D-023), result as a ledger row in sw-registry.csv, every finished line ends with `cost: credits, API $, hours`. On every wake: fresh doorbell (send_later 2027-06-01) on LINE 1 of your status (D-022).
  - Why: most SW live book trades end at the 0.3x stop. One early signal that fires before the stop and saves +0.05x per loser is worth about $50 per loser at $1k, with zero effect on winners if it never fires on them.
  - Candidate signals, pick ONE on half A only (live book + kit300k trades, your tick walk engine): (1) in-range LP removal started (Burn of 20%+ of in-range units), (2) the coin's deployer or top pre-$300k bundle wallet sends tokens to the pool or a router, (3) net sell pressure in the prior 3 blocks over X% of pool WETH. Freeze it, then price half B.
  - Bar: half B $ a day up vs the current exits, winners lost (any 5x trade the signal exits early) counted and priced, beat 95% of 200 random early exits of the same count. Note: FW Analyst 2 runs the same idea on FW's mc5 book; share the signal code, do not copy their pick.
  - First answer: pre-registration within 45 min (why: your bundle and pull tables from A3.12/A3.13 exist), priced result within 2 h of it (why: tick walk reruns about 30 min per signal).
- **v1.13 (Supervisor, 7 Oct 10:56 Bangkok): A3.11 KILL ACCEPTED (smart-wallet lists are just busy early buyers; matches A1.26: pre-$300k buyers are the operator's own bundle). A3.12 SMALLER ACCEPTED: kit300k stays a paper book, judged at 40+ half C trades.**
  - **A3.13 THE OPERATOR'S EXIT (one pass, then stop).** A1.26 (http://159.223.88.168/hx-c9ois/sw/a1-26-funders.md) shows we can name a coin's own bundle wallets: fresh wallets funded through disperse contracts 0x84157d3d24630c297849c84d49ae7f1805c1b60f and 0x03802649e469ed5680c58dff9e65425e22d21cc5 shortly before they buy. Idea: the operator sells his bundle near the top, so his first big sell is our exit.
    - On the kit coins held past 5x (live kit hold) and the kit300k trades from A3.12: tag each coin's bundle wallets; find the first moment the bundle has sold X% of its tokens (X 10/25/50, pick on half A only). Exit rule: sell the rest at that moment, else the live trail.
    - Compare with the live 50% trail. Already killed, do not redo: drip-seller exit (FW A1.28), net-buy exit (V14), pull tell exit (A3.3). Say in one line why this is different (we key on the operator's OWN wallets, not on flow).
    - PASS = $ a day above the live exit in both halves and beats 95% of random exit times with the same count. Else KILL.
  - Run on research-box or the Mac. Output http://159.223.88.168/hx-c9ois/sw/sw-a3-13-operator-exit.md , ring me only when done.
- **v1.12 (Supervisor, 7 Oct 10:28 Bangkok): A2.31 audit accepted (moonbag stays killed). New audit, priority over A3.11 when A3.11 finishes or stalls: A3.12 KIT AT $300k.**
  - Analyst 1 lead (http://159.223.88.168/hx-c9ois/sw/a1-27-early-map.md, section "Unregistered lead"): buy EVERY Kit A / Kit B / 3,587-family coin 30 s after its fair $300k milestone ($15k bought at a cap of $300k or more), live exits (0.3x stop, 5x, kit hold 50% trail, pull exit, 12 h). A1 got A $3,416 / B $14,471 a day vs live $3,117 / $12,038, catching 9 of 9 and 10 of 10 $10M coins. A1 used virtual-reserve depth.
  - Rebuild on YOUR data with a full tick walk of in-range depth. Report $ a day, worst day, median x, without top 1 and top 2, beat 95% of 200 random draws of ALL v3 coins at their fair $300k milestone, and live + kit300k combined (two tickets when overlap). Also the stop-fill reality: how many stops gap through 0.3x and the real fill.
  - Verdict: CONFIRMED / SMALLER / NOT REAL. It cannot go live before half C passes (40+ paper trades), but your audit decides if the Coder's paper book is worth the slot.
  - Run on research-box (bus v1.3). Output http://159.223.88.168/hx-c9ois/sw/sw-a3-12-kit300k.md , ring me.
- **v1.11 (Supervisor, 7 Oct 09:45 Bangkok): EDGE SPRINT E1, Ben's order: buy early and hold to multimillions. You are auditor AND hunter this round.**
  - **A3.11 SMART WALLET COHORT (earlier entry).** From half A only: list wallets that bought before $300k in 2 or more coins that later reached $3M+ (drop known bots, routers, snipers that sell inside 2 min, and wallets of the coin's own creator/funder). In half B, test: buy when K of these wallets have bought within N minutes of launch (grid K 1/2/3, N 5/15/30), entry at that block's real price, exit with the live book's rules.
  - Leakage check is the whole game here: the list must be built ONLY from half A outcomes. Report how many half B 10x coins the signal would have caught, and at what cap.
  - Also audit A2.31 when Analyst 2 posts it.
  - **Sprint rules (all of E1):** pre-register the rule in your status BEFORE you look at results; split half A (to 30 Sep) / half B (1 Oct on); price at real depth with $1k; report $ a day, worst day, and beat-random %; PASS needs both halves positive vs the live book AND beat 95% of random. A clean KILL is a win too, write it up the same day.
  - **Already killed, do not redo:** buy at hour 8 after $500k (V10), regime hold on non-kit (V13), net-buy exit (V14), drip-seller exit (FW A1.28), $750k gate (A2.16/A2.25, half C 0.16x), operator flag hold (A2.29). Cite them, build past them.
- **v1.10 (Supervisor, 6 Oct 17:10 Bangkok): A3.9 ACCEPTED. Kit hold goes live under the pre-registered rule, with your kill switch idea built in** (auto revert after 3 kit holds in a row below selling at 5x). Excellent audit; the two extra tagged coins and the "dumped to 0.3x" row made it solid.
  - **A3.10:** your rebuild has no 3-open cap. Kit holds last up to 48 h and fill slots. Rerun all out at 5x vs KIT HOLD with max 3 open (skip new gates when full), and again with the new slot rule (a hold past 5x does not count, at most 2 extra). Report $ a day, both halves. If kit hold under the cap loses its lead in either half, ring me at once.

- **v1.9 (Supervisor, 6 Oct 16:50 Bangkok): A3.8 ACCEPTED (keep all out at 5x). Next, A3.9: audit Virality V8, the KIT HOLD exit. This one could go live.**
  - Claim: at 5x, if the coin's pool was created by Kit A (bytecode keccak 0e81be52d5cf) or Kit B (89885b75fd29), hold everything on a 50% trail from the peak (or pull, or 48 h); else sell all at 5x. $1k: $3,638 / $7,546 a day vs $3,068 / $3,054 for all out at 5x, 13 holds. http://159.223.88.168/hx-c9ois/sw/sw-virality-v8.md
  - On YOUR rebuild (wide gate, 0.3x stop, block-end fills, range-aware depth, $1k): price both. Report per half: mean x, $ a day, worst day, worst 3 days, max drawdown.
  - Attack it: (1) kit tagged from the pool create tx only (no hindsight); (2) holds still open when the data ends: price them at the last real sale quote, then again as if dumped to 0.3x; (3) drop the single best kit hold, still ahead in both halves?; (4) Kit B only (Kit A is dead since 1 Oct): what does "Kit B hold, else sell" do in half B?
  - **Pre-registered verdict:** SHIP LIVE if, on your rebuild, $ a day is higher in BOTH halves, still higher in both after dropping the best hold, and max drawdown is no more than $1,500 worse in either half. Else PAPER (re-test on half C from 7 Oct). Ring me.

- **v1.8 (Supervisor, 6 Oct 16:10 Bangkok): A3.7 ACCEPTED (V7 paper only). Next, A3.8: should LIVE switch to ladder (b)?**
  - Virality V6 on the wide list (http://159.223.88.168/hx-c9ois/sw/sw-virality-v6w.md), $1k: all out at 5x 2.19x / 1.83x ($3,068 / $3,054 a day, worst day +$714 / -$2,000) vs (b) 50% at 5x + 50% trail 2.13x / 2.64x ($2,915 / $6,023, worst day +$1,493 / -$2,000). (b) is a fixed rule, no regime guess.
  - On YOUR rebuild, wide gate, 0.3x stop, $1k, range-aware depth: price (a) and (b). Report both halves: mean x, $ a day, worst day, worst 3 days.
  - **Verdict rule:** if (b) loses no more than 5% of $/day in half A, has a worst day no worse than (a) in both halves, and gains in half B: say SWITCH. Else KEEP. One line.

- **v1.7 (Supervisor, 6 Oct 15:55 Bangkok): audit Virality V7 (hold-or-sell regime signal): http://159.223.88.168/hx-c9ois/sw/sw-virality-v7.md**
  - On your own rebuild: at each 5x hit, prior_hold_share over earlier 5x trades from the last 4 days whose HELD version has fully exited (no lookahead). Hold on a 50% trail if 0.3 or more, else sell at 5x. Both halves, $500 and $1k, real depth.
  - Also: what happens if the regime flips back (hold-share rule says hold, then the next coins get pulled)? Measure the worst case: the cost of the first 2 to 3 wrong holds after a switch.
  - Verdict in plain words: ship to paper only / ship to live / kill.

- **v1.6 (Supervisor, 6 Oct 15:12 Bangkok): after the V5 audit, audit A2.16 too** (http://159.223.88.168/hx-c9ois/sw/sw-a2-16-gate-grid.md). On your own rebuild, run the one cell: gate $1M sellable, age <= 3 h, top <= 0.08, army <= 0.015, no-stop exits. Both halves: trades/day, mean x, median x, $ a day at $500, worst day. One verdict line for V5 and one for A2.16, then ring me. These two decide the live config.

- **v1.5 (Supervisor, 6 Oct 14:55 Bangkok): NEW FIRST TASK, before A3.2: audit Virality V5.** It says the 0.6x stop kills winners: no stop 2.20x / 2.19x, stop 0.3x 2.20x / 2.07x, vs 0.6x 2.07x / 1.89x (http://159.223.88.168/hx-c9ois/sw/sw-virality-v5.md). If your own rebuild confirms it in both halves, the live rule drops the stop. So:
  - Rerun your A3.1 rebuild with stop 0.6x, 0.4x, 0.3x, none (target 5x, half-pool pull in-range only, 12 h). Both halves, without the best 3 trades, median trade, worst day.
  - Check the 5 "stopped then 5x" coins it names. One of them, 0xabe99cf268cdc8bd70a71a5283fdacbcb36714fd, appears in Analyst 1's A1.11 as a FARM WALLET, not a coin. Which is right? If it is a wallet, V5 has a data mix-up; say how big.
  - Pull exits with no stop average 0.13x: check they are priced at what the pool really paid after the pull, not at 0.
  - Verdict line: "V5 CONFIRMED" or "V5 NOT CONFIRMED", ring me.

- **v1.4 (Supervisor, 6 Oct 14:48 Bangkok): A3.1 ACCEPTED. Strong audit. Two fixes to your status, one question, then A3.2 to A3.4 as planned.**
  - Your "Needs Ben: dRPC out of money" is stale (checked 13:18; Ben topped it up ~13:22, the Coder measures 0.26 s per paid call since 14:05). Remove it. The Supervisor doorbell works if you fire it WITHOUT the text field (sw-bus.md v1.2); remove that ask too.
  - **Question (decides 1.915x vs 1.99x):** in your rebuild, is 0x978177408c2f741c24ddf5ea9eb14cecbf9b3926 a pull or a 5x target? Analyst 2 (A2.14) found its "pull" was a far out-of-range 1-tick burn (position ticks 137355 to 137356, price tick 140410) and it really hit 4.83x. Your exit counts (3 pulls) match A2's OLD count. If your pull test counts out-of-range burns, rerun with in-range only and give the new total, both halves.
  - A3.3 next priority after A3.2: your stop near-miss (0xa2508b15ab16826c720631b9afa0a90c453ae8a4: stop missed by 1%, -43% became -97%) is exactly what costs money live. Coordinate with the Virality Expert's V5 (looser stop test) by reading its status; do not duplicate.

- **v1.3 (Supervisor, 6 Oct 14:00 Bangkok):** Analyst 2's latest number is 1.99x on the 58 at $500 (A2.14). Add to your A3.1 bug checklist: does a Mint or Burn outside the current price range get counted as a pull or as liquidity? Compare your result with 1.99x and with the old 1.90x, and say which one your rebuild supports.

- **v1.2 (Supervisor, 6 Oct 13:50 Bangkok):** the paid dRPC key is topped up and working again since about 13:30 Bangkok (Coder moved back to it at 13:35), so you can use it to speed up A3.1. The Coder is about to rotate the dRPC key; read it from the shared config at use time, never print it. Order stays A3.1 first.

- **v1.1 update (Supervisor, 6 Oct):** Analyst 2's loser count changes the priorities. Of 38 losers only 7 went near zero; 28 stopped at about minus 43%. So a pull warning saves little (about +0.05x). The bigger lever is spotting the losers early. A3.3 is changed below. A3.1 stays first, and must use the live definitions listed in Analyst 2's task A2.8 (no look-ahead in pool choice, supply, army memory or hubs): http://159.223.88.168/hx-c9ois/sw/sw-tasks-analyst2.md

- From: the 2nd Window Supervisor, 6 Oct 2026. Your job: make sure our best rule is real, and find a way to dodge its biggest loss.
- Our top rule is the v3 runner rule (`w2_v3_1m`) from Analyst 2: http://159.223.88.168/hx-c9ois/hx/second-window-v3.md. Paper result: about 1.9x per trade on 59 trades, about 1 in 3 hits 5x, the rest mostly lose about 43%. Pools get pulled by the creator 5 to 20 h after start (78% of v3 coins).
- Other task lists, same folder: PP http://159.223.88.168/hx-c9ois/sw/sw-tasks-pp.md, Coder http://159.223.88.168/hx-c9ois/sw/sw-tasks-coder.md, Analyst 1 http://159.223.88.168/hx-c9ois/sw/sw-tasks-analyst1.md, Analyst 2 http://159.223.88.168/hx-c9ois/sw/sw-tasks-analyst2.md
- **No Dune.** Ben wants to avoid its cost. Use Price Predict's own chain data (day files in /opt/hist, /opt/pricepredict-data) and the paid dRPC node. If something truly needs Dune, ask the Supervisor first.
- Bangkok time, full addresses, paper only.

## Tasks, in order

### A3.1 Rebuild the v3 rule from scratch (independent check)
- Do NOT reuse Analyst 2's scripts. From raw swaps, rebuild the gate (sellable $1M, under 1.5 h, top receiver 8% or less, army share about zero) and the exits (5x, minus 40%, half the pool pulled, 12 h), with 1% fee each way and price impact from pool reserves.
- Check for the usual bugs: using data from after the buy moment, a "sellable" level that was only one wash trade, fills at prices nobody could get, pools counted twice, USD prices that are wrong (ETH vs WETH).
- Output: your result next to Analyst 2's, trade by trade. Every trade where you disagree, say why.

### A3.2 Can these coins be sold? (sell block census)
- For the 59 rule coins and all ~926 v3 coins since 13 Sep: did outside wallets sell successfully after our buy moment? Any token with a sell tax, blacklist, max wallet or pause function? Read the contract, don't trust a site.
- Output: share of coins that block or tax sells, and whether any of the 59 trades would have been stuck. This sets how strict the Coder's sell check (C2) must be.

### A3.3 Early fail exit (changed in v1.1), plus pull warning
- Main question: in the first 15 to 60 minutes after our buy, what separates the 19 trades that hit 5x from the 35 that stopped out (counts from the A2.1 spec replay) (buyers per minute, new wallet share, sell pressure, price vs entry, army wallets arriving)? Find an early exit (for example at minus 15 to 20%) that cuts losers without cutting winners. Report the money result on the 59 trades, both halves.
- Second, smaller: the pull warning below, for the 7 coins that went near zero.
- 7 losers went near zero, mostly pulls. If we can see a pull coming, we sell first.
- For every v3 coin that got pulled: what did the LP owner, deployer and their linked wallets do in the 60 minutes before the pull? Look for: moving ETH out, funding fresh wallets, approvals, small test removals, selling their own tokens, buy pressure stalling, a new coin launched by the same funder.
- Find signals that come before the pull at least 1 to 5 minutes ahead and rarely fire on coins that were not pulled. Report hit rate, false alarms, and the money result of selling the moment the signal fires (on the 59 trades).

### A3.4 Bad day risk
- On the 59 trades: how many trades overlap in time, the most money tied up at once, the longest losing streak, and the worst 3 day drop at $100 and $500 a trade.

## Hand back
- One short doc per finished task in this folder, linked from http://159.223.88.168/hx-c9ois/sw/sw-analyst3-status.md. A3.1 first. First principles score at the end of each.

Version: v1.6 (then audit A2.16 wide cell), 14:48 Bangkok, 6 Oct 2026, 2nd Window Supervisor. Latest version: http://159.223.88.168/hx-c9ois/sw/sw-tasks-analyst3.md