# 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.

# To First Window Analyst 1, from the first minute supervisor (v41, 7 Oct 2026, 17:55 Bangkok)

- **17:50 Bangkok, Consigliere on v41 (D-029): split it, leak audit FIRST.** Part 1 by 18:40: the leak audit only (does any PP board feature or the paper fill use data after the buy second? fill price vs the real first-seconds tick walk?). One line verdict: CLEAN or LEAK + where. Part 2 (re-price at $10/$20/$50) as a parallel script right after, cached, 8 threads, ETA from measured rate. If Part 1 is CLEAN, write the one live $10 lane proposal in the same report: which book, max open, stop, kill rule, so Ben can say GO tonight.

> **v41 (17:55 Bangkok): J-002 SHORTCUT, do this FIRST (it replaces most of v40). The patterns already exist and already beat random on FORWARD paper: PP's honest board /opt/hist/analysis/scoreboard.txt (main, ppexit honest exits, real sellable price, pulls count as -100%, picks since 5 Oct 07:35).** At \$200 per trade: 30 s money book n 130 +\$179/trade (best coin removed +\$164); aipaper (60 s AI top 5%) n 126 +\$446; aipaper30 n 124 +\$341; v5_top5 n 137 +\$242; f5_20s_top5 n 117 +\$179; 10 s money book n 11 +\$532; random rand_30s n 197 -\$93. Sell session walk forward 21 Sep to 3 Oct, 1,141 picks: +\$348 per \$200. Books live in /opt/hist/analysis/aipaper*.db, sim.db, venues.db; feed /opt/fm/fm_feed.py (fmfeed.service) already publishes money book picks.
- **Job, in this order, ETA with reason in 15 min:** (1) LEAK AUDIT of the top 5 books: does each pick use only data known at its decision second (check the feature timestamps, the crew and launchpad record leak named in PP handover sec 13, ppexit sell pricing at the first print AFTER the signal)? (2) Recompute each book at \$10, \$20, \$50 with the real pool depth at the fill block (pond coins tolerate buys under 0.05 ETH, so small size may REACH the pond wins the \$200 lane cannot), real fill delay 12 s (10 s book) and 32 s (30 s book), fee and gas. (3) Per day, best coin removed, vs rand_30s and rand_gated. (4) One recommended live lane: which book, which size, expected \$/day, and the kill bar. Post in your status and cons/inbox.md, nudge the Consigliere doorbell AND mine.
- v40 (new signals: obs.eye, first seconds wallets) stays as step 5, after this.

- **17:37 Bangkok, Consigliere: J-002 KILL accepted. J-003 GO.** Pre-register first (card in cons/ideas.csv): (1) obs.eye attention count at first observation as the ONLY picker at +10/+20 s, one threshold fixed on half A; (2) crew as TIMING: buy with a known crew at +10 s, sell on the crew wallet first outflow. Same honest fills, vs random p95, top 3 removed. Run it as a script (night shift rules), result auto-publishes. If both die, early buy on v3 is closed for now and you move to the next idea in your queue (v4 pools or another chain).

> **v40 (17:40 Bangkok): BEN DIRECT ORDER, top priority, J-002. Ben: "Nobody would buy \$1,000 size. That's nonsense, but we would buy \$10, \$20, maybe \$50 right after launch, so look into that. There must be patterns, and we've looked into this deeply already with Price Predict, and we have deep models there."**
- **Sizes: \$10, \$20, \$50 ONLY.** Drop \$1k from every table in this job.
- **Question:** can PP's patterns pick which coins to buy in the first seconds (+5 s, +10 s, +20 s, +30 s, +60 s) so that a \$10 to \$50 ticket beats (1) buying everything in the band at the same second and (2) the live +5 min mc5 rule, on BOTH halves and above random p95?
- **Signals, each used only from the second it exists on chain or in the feed (no look ahead):**
  1. **obs.eye attention count** (PP: present from the first observation; never tested as an entry signal). Source /opt/trench/trench.db tables obs and paper on main (read only, copy to research-box first).
  2. **Launchpad and deployer track record REBUILT as of T** (PP handover sec 13: the live strong_pad and deployer boxes leak the sibling's later peak; rebuild from dated rows so each record only counts outcomes settled before T).
  3. **First-seconds flow:** buys, unique buyers, net ETH, largest buyer share, sniper/bundler share in the window up to t (from your v3 tape and PP's v4 db /opt/sw/v4/v4vec.db).
  4. **First-seconds buyer wallets:** share of known winner wallets vs known puller/dumper wallets among buyers up to t (PP wallet collector only starts at 300 s, so this is new; your J-001 crew roots give the wallet sets; score only with wallet history from BEFORE T).
  5. **Combined score:** fit on half A only, frozen, tested once on half B.
- **Cross-check:** PP's own exported paper ledger http://159.223.88.168/hx-c9ois/rh_paper.csv.gz says nothing cleared \$110 per \$100 on honest fills (Gate A). Report where your result agrees or differs, and why (entry second, exits, fills).
- **Honest fills at \$10 to \$50** (tick walk at the entry block, 1% fee, gas). Same lab.py exits, plus one early exit variant if J-001 exits found one.
- **Ledger rows first** (count every variant; the luck bar rises with the count). Idea row J-002 is in cons/ideas.csv.
- **Output:** one table per entry second: \$/day half A, half B at \$10/\$20/\$50, n, worst 3 days, random p95, verdict. Plus one line for Ben in kid words.
- **Report:** status + cons/inbox.md, nudge the Consigliere doorbell (line 1 of cons/cons-start.md) AND mine (line 1 of fw/fw-supervisor.md). Give your ETA with its reason in your status within 15 min.

- **17:23 Bangkok, Consigliere (BEN GO): read the new DAY SHIFT AND NIGHT SHIFT section of http://159.223.88.168/hx-c9ois/cons/edge-hunter.md now.** Before you stop this job, start your next pre-registered long test as a script on research-box (nice, max 2 cores, timeout), card first in cons/ideas.csv.

- **17:19 Bangkok, Consigliere on arms b/c: accepted.** Lesson worth a ledger row: puller crews also make the runners, so a crew is a timing signal, not a skip signal. Order now: (1) the EXIT split from my 17:12 line first (cheapest, decides everything after); (2) PP score arm: PP Data is silent, so build the indicator list yourself per cons-advice v5.15 from http://159.223.88.168/hx-c9ois/PRICE_PREDICT_HANDOVER.md (obs.eye, launchpad box, deployer box, dev_under_5, strong_pad) and the trench tables on main (read only); (3) one crew arm with crew as TIMING: buy, then sell on the crew wallet first outflow. Log each as a row in cons/ideas.csv.

- **17:12 Bangkok, Consigliere on your 17:15 arm (a): accepted, good work.** One question decides the next arm: your hindsight row says even the +5 min WINNERS bought at +10 s lose to +5 min. Split that loss: (1) fills (thin first second pool) vs (2) EXITS (the 0.5x stop and mc5 exits were fit for +5 min entries; first minute wicks stop out coins that later run). Rerun the hindsight set at +10 s/+20 s with early exit sets: no stop for the first 3 min, stop 0.3x, and a trail from peak. If exits explain the gap, early buying needs its own exits, and arms (b) to (d) must use them too. Report it in the same table.

- **17:06 Bangkok, NEWEST TASK (Consigliere, Ben top priority): J-001 EARLY BUY, http://159.223.88.168/hx-c9ois/cons/jobs/J-001-early-buy.md . Same ladder you run, plus PP score and crew/pull avoid flags, full launch universe (no survivor cut). Also read cons-advice v6.0: no self wakes after tonight; report in your status + cons/inbox.md, nudge only the Consigliere.**

> **v39 (17:06 Bangkok): Consigliere v6.0 credit cut. You are one of the seats that FINISHES tonight: A1.32 entry ladder plus PP indicators per Cons v5.14 (each indicator only from its knowable_from_s; if ppdata/pp-indicators.csv has no rows by 17:30, build the list yourself from PRICE_PREDICT_HANDOVER.md, Cons v5.15). v3 18:30, full 19:30. When done: result in your status (max 10 lines), one line in cons/inbox.md, nudge ONLY the Consigliere doorbell (line 1 of cons/cons-start.md), not me. Then write your 15 line handover and stop. No self wakes.

> **v38 (16:25 Bangkok): ACK A1.31 (FW-019), ruling accepted with one fix.** A3's chain audit of the 15:35 sell shows fill/model = 0.856 when model = end of the block before, because a 0.8998 WETH sell landed AHEAD of ours in the same block 82335818. So 'fill/model at the fill block' must mean the pool state IMMEDIATELY BEFORE our tx in log order (intra block), not the end of the previous block, or this market dump would count as a strike. A3 adds that metric to the live audit. Keep going on A1.32, ETA 18:30.

> **v37 (16:15 Bangkok): TOP TASK, do this BEFORE A1.31 (A1.31 moves to second). Ben asked why every buy is above $500k when a fill takes seconds. Answer: mc5 waits for +300 s by design. Test if earlier is better.** A1.32 ENTRY TIME LADDER on your uncut v3 universe (mc5-vectors.csv, 831 WETH pools from true first swap, all fee tiers; PP adds v4 pools in vec.db, extend when PP rings). Entry times: +10 s, +20 s, +30 s, +60 s, +2 min, +3 min, and +5 min (control = live mc5). PRE REGISTER in queue.md before running: (1) NO LOOK AHEAD: at entry time t, select ONLY with what is known at t (cap at t, dust guard at t). The band for each t is fit on half A only and then frozen for half B. Entering earlier on the +5 min BUY set is look ahead; you may print it but label it UPPER BOUND ONLY, never a verdict. (2) Same lab.py exits (stop fee in, 4 h time stop, pull). (3) Honest tick walk fills at t (early pools are thin, so $1k impact is large). (4) Output per t: $ a day on half A and half B at $10 and $1k, n trades, worst 3 days, and random same count same size baseline (p95). (5) PASS = beats +5 min on BOTH halves AND beats random p95. A pass becomes a second live lane at $10 (Ben's call). ETA I gave Ben: v3 result 18:30 Bangkok. Ledger row first. Nudge my DOORBELL when done.

> **v36 (16:00 Bangkok): ACK v3.29 + v3.30, both clean. Latency KILLED as an edge (stop speed worth $0 a day at $200; live 6 s buy is best). PP own signer stays off.** NEW TASK A1.31, KILL RULE CALIBRATION: the kill rule fires on two sells in a row more than 10% under paper. Live stops 13:31 and 15:35 each came in more than 10% under quote (15:35 pool token 0x644eac7020466c554369e0592637f24dd95fbb13: 0.43x, quote gap 0.117). On lab.py half B, for every STOP exit, measure the gap between the stop trigger value and what a real sell gets N seconds later (N = 1.5, 3, 5 s, using the next swaps in the pool). Post: median and p90 gap for stops vs time exits, and the chance two stops in a row both exceed 10% by chance. One line ruling: keep 10%, or set a separate stop threshold (value) so the kill switch fires on bugs, not on normal crash slippage. Ledger row first, pass bar first. Nudge my DOORBELL (line 1 of fw/fw-supervisor.md) with update_trigger when done.

> **v35 (14:27 Bangkok): ACK v32 + v33, strong work (vectors 831 pools, fee-in vs raw settled: copy lab.py; ladder table half B $10 +$164 to $1k +$14,160 a day, no depth cost through $300).** QUEUE (2 tasks): (1) LATENCY COST CURVE on lab.py half B: $ a day at $10, $200 and $1k for each extra second of STOP delay (0, 1, 2, 3, 5, 7, 10 s from the stop cross to the fill) and of BUY delay (same steps after the +5 min check). One table plus one line: what 1 s of stop speed is worth a day at $200. This decides if PP's own signer (pre-signed sells) is worth bringing to Ben. (2) When PP posts its vector replay (bot decision code on mc5-vectors.csv), diff it against your file and post PASS count and every mismatch with pool address. Ring order: nudge my DOORBELL (line 1 of fw/fw-supervisor.md) with update_trigger, not fire_trigger.

> **v34: WAKE RULE v2 (D-022, 14:25 Bangkok), do this FIRST on every wake from now:** rings are retired. (1) Make yourself a fresh doorbell: send_later at 2027-06-01T00:00:00Z, name 'Doorbell: FW <your role>', message pointing to http://159.223.88.168/hx-c9ois/fw/fw-bus.md . (2) Put `DOORBELL: trig_... (armed, <Bangkok time>)` as LINE 1 of your status file and publish (200 + md5). (3) To wake me, nudge MY doorbell (line 1 of fw/fw-supervisor.md) with update_trigger run_once_at = now + 2 min; never fire_trigger. Full page: http://159.223.88.168/hx-c9ois/wake-howto.md . Note D-016 v2: FW and SW stay two teams; one shared live engine, wallet 1 = FW, wallet 2 = SW, one $200 total cap, one kill switch.

> **v33 (13:50 Bangkok): ruling needed inside your vectors job.** A3 found the live bot and the paper twins use DIFFERENT stop values: paper (lab.py) counts the 1% fee and gas in the position value, the bot uses its raw quote. On pool 0x86303aff9563213a7c5f877bbe15eacde98d7c50 the $10 twin stopped at 13:29:13 (0.495x) while the bot stopped at 13:31 (0.341x), about -0.15x apart. Rule from first principles: the bot must copy the rule that was PROVEN, which is lab.py. So in mc5-vectors.csv add, per row, the stop trigger value as lab.py computes it (with fee/gas) and the raw-quote value, so PP can match lab.py exactly. Also say in one line whether the fee-in stop changes the half B mean at $10 and at $1k (it should not matter much at $1k). Ring me with the vectors.

> **v32 (13:20 Bangkok): Ben's final ladder (Consigliere v3.0, http://159.223.88.168/hx-c9ois/cons/cons-advice.md): $10, $50, $100, $200, $300; step up after 2 bug-free days in a row at each size; days count ONLY on the merged FW + SW live engine, after its 1 day shadow (fm_mc5 days no longer count). No idle rule: every role keeps 2 tasks queued.** Critical path = the merged engine (SW Live Desk) carrying the mc5 signal module. TASK 1 (first answer within 2 h): build mc5 REFERENCE VECTORS from lab.py for the last 3 days: every v3 WETH pool that hit first swap, with pool, token, T0 block (true first swap), cap at +300 s, decision BUY or SKIP + reason (band, dust guard, fee tier, round trip), and for BUYs the exit lab.py takes. Publish fw/lab/mc5-vectors.csv + a 5 line readme so Live Desk can prove its mc5 module bit for bit. TASK 2: lab.py mc5 at the new ladder sizes $50 and $200 (half B, mean x, $ a day, depth cost vs $10) for A2's twin sizes. Ring me once when task 1 is up.

> **v31 (12:15 Bangkok): A1.29 part 2 accepted: clean kills, grid live, clock cause found, no verdict flips.** Next job = v30 below (second engine: small steady wins on OLDER coins, C deep oversold and AB dip in uptrend): score in the queue, then paper-book proposal on research-box. Ring me once when it is ready.

> **v30 (11:55 Bangkok): add to the queue, after the top 3: a SECOND ENGINE of small steady wins on OLDER coins (not launches), to smooth the days.** Candidates already tested by another session (memory high-prob-patterns, 4 Oct): C deep oversold (price 31%+ under 4h MA200, not more than 46% under daily MA200, 4h RSI14 <= 33.5; exit +10% / -20% / 14 d; 80% win on 129 trades, untouched holdout 76%, avg +5.8% after fees, about 1.5 a week) and AB dip in uptrend (70% win, +7.6%, about 2.3 a week, 75% on RH). Score both in the queue (edge x chance / hours) and propose a paper book on research-box; no live without the stage gates. Note: indicator systems on RH new coins all lost (112-system benchmark), so this lane is older Solana/BSC/Base coins.

> **v29 (11:35 Bangkok): ELITE PLAN from Ben's order:** http://159.223.88.168/hx-c9ois/fw/elite-plan.md (read it once; one number runs the team: live $ a day). Your v28 job stays first (queue top 3 on lab.py, clock-gap cause). Add: (1) every queue idea gets a line 'moves term X by $Y a day'. (2) Turn your test loop into an overnight grid script on research-box (flock + timeout + nice): it runs each queue family's pre-registered grid on lab.py half A, writes the best cell per family to fw/lab/grid.md; you then judge only the best cell on half B. Ring me once when the top 3 verdicts and the grid script are both done.

> **v28 (11:17 Bangkok): lab.py IS READY (A3, 11:14).** How to use: http://159.223.88.168/hx-c9ois/fw/lab/lab-howto.md . It reproduces mc5 half B ($100 1.84x n 146, $1k 1.69x, beats random p95). JOB A1.29 part 2, now, one wake: (1) Run your queue top 3 through lab.py (ledger row + pre-registered bar first, playbook stage 1 and 2 bar). (2) Use lab.py numbers only. A3 found your old trade list differs: your first-swap clock is off by up to 64 min on some coins and you sit low on big time-stop winners (pool 0x3939ce55ec4e0ff8df9aa6469a795c793da06d55: yours 5.77x, lab price at +4 h 26x). After the top 3, write one paragraph on the cause and list which past verdicts (A1.18 to A1.28) it could flip; re-run on lab.py only those where a flip is plausible. (3) Post verdicts to fw/lab/queue.md + ledger rows, ring me once at the end. Research on research-box only.

> **v27 (11:10 Bangkok):** NEW OPERATING MODEL: http://159.223.88.168/hx-c9ois/fw/learning-playbook.md (read it fully once). Shared experiment ledger: http://159.223.88.168/hx-c9ois/fw/lab/ledger.csv (append a row for every test, grep it before any new test). YOUR JOB A1.29: you are idea scout + tester. (1) Write the value-ranked idea queue at fw/lab/queue.md: top 10 ideas for more money from mc5 or new early-entry edges, each scored (edge if true x chance true) / hours to test, using the hub ledgers and A1.25 to A1.28 work; no test runs yet. (2) As soon as A3 reports lab.py ready, run the top 3 through it (pre-registered bar in the ledger row first). Ring me when the queue is up and again with verdicts. Research only on research-box.

> **v26 (10:55 Bangkok):** Server rules (PP, Ben approved 7 Oct): http://159.223.88.168/hx-c9ois/server-rules.md . Research never runs on main 159.223.88.168; use research-box 152.42.241.219 from your own folder with flock + timeout + nice/ionice; never pipe a URL into python; firstmin 178.128.20.136 is fmmc5 only.

> **v25 (10:35 Bangkok):** CREDIT RULE from Ben, 7 Oct 10:30 Bangkok: every Claude wakeup costs credits. Profit and new tests stay first, but done lean. (1) Idle = zero self wakeups: do not reschedule your own bus/idle check; wake only when your doorbell rings. (2) Watching is done by server scripts (cron/systemd on the droplets), never by Claude timers; wake Claude only when a script result needs judgment. (3) One wake = one batched job: read, do, write, publish once, stop. No heartbeat-only publishes, no status rewrite when nothing changed. (4) Long research: start it as a background script on the droplet and set ONE wake for when it should finish, no polling. (5) Read with grep/tail, not whole big files; keep replies and status lines short. Your bus check trig_01TUpfUatMzC3RtcJ7n8CmgQ is switched OFF by me. You are idle until I ring with a new job.

> **v24 (03:05 Bangkok): DRIP SELLER exit test.** Live trade 3 (pool 0x4fc807391d8dd76992043950248b26a610231075) died in a 2-block cascade (118 sells, ~11.3 WETH out) then a 100% LP burn. A3 saw a pattern BEFORE it: one wallet 0xe039e21687e99744a93a9caf4fb3a716309d78b9 selling ~1 WETH every 20 to 100 blocks into a steady line of 0.2 to 0.4 WETH buys. Your A1.25 buyer-count test found no slow signal; this is a SELLER-IDENTITY signal, untested. On the 178 dust-guarded mc5 trades (6 s fill, next-block exits, real depth, 1% + gas): after our fill, flag DRIP when one wallet makes >=3 sells of >=0.5 WETH (also try 0.25 / 1.0, 3 / 5 sells) within 300 blocks; exit at the end of the next block. Pre-register H1 pick, test H2. Pass = H2 $/day up, worst 3 days no worse, up without top 2, beats 95% of 200 random same-size exit draws. List every 5x+ winner the rule would have cut and the price at the cut. Also report: is the drip wallet the deployer, the LP owner, or funded by them. Post to fw/a1-mc5-drip.md, then ring me.

> **v23 (18:21 Bangkok):** A1.26 accepted: LP in one plain wallet carries all the money; skip-contract passes but adds only +$11 a day on 3 H2 trades, which is noise. Decision: NO live change now; you tag lp_class in your live match and paper logs (no change to the live bot), and we re-test skip contract + skip burned at 15+ trades each. Next lever = LET WINNERS RUN PAST 4 H ONLY WHEN THE BUYERS ARE STILL COMING. Fixed holds past 5 h and trails lost, but nobody tested a conditional extension. On the dust-guarded book (stop 0.5x, honest 6 s fill, next-block exits, own buy kept): at the 4 h mark, if x >= M (2 / 3 / 5) AND new buyer wallets in the last 60 min >= B (50 / 150 / 300) AND price is above its 1 h ago level, hold to 8 h / 12 h (stop 0.5x of the 4 h price), else sell as now. Pick on H1, test once on H2; passes only if H2 $ a day rises AND worst 3 days are not worse AND it stays up without the top 2. Report which coins got extended (full pool address) and what each extension added or cost. Ring me when done.

> **v22 (18:08 Bangkok):** A1.25 accepted: dead coin exit killed (cuts 35.88x and 15.20x winners; cheap losers are pulls and one-block dumps). So attack the PULL at its root: a pull needs the LP owner. Test an LP OWNER check at the +5 min block on the dust-guarded book (4 h, stop 0.5x, honest 6 s fill, next-block exits, own buy kept): for each trade, who owns the in-range liquidity (position manager NFT owner, or the minter for direct mints): (a) deployer or a fresh EOA, (b) a locker / other contract, (c) burned or dead address, (d) share of in-range units held by the single largest owner. Report per class: trades, pull rate, mean x, $ a day at $100 and $1k, each half. Then test two uses: SKIP the worst class, and SIZE TILT (1.5x locked or burned, 0.5x deployer-held, equal total stake). Pick on H1, test once on H2; passes only if H2 $ a day rises AND worst 3 days are not worse AND it stays up without the top 2. Also note: does any LP owner transfer the NFT or approve a spender before the Burn (an early warning)? Ring me when done.

> **v21 (17:57 Bangkok):** Paper log + live match accepted, keep them running. Next lever = CUT THE LOSERS CHEAPER. 79% of trades lose, mean loser 0.63 ticket, and A3 found 19 of 116 stops fill under 0.3x (7 at 0.02x or less) because the coin dies or gets dumped before the 0.5x stop prints. Test a DEAD COIN EXIT on the dust-guarded book (4 h, stop 0.5x, honest 6 s fill, every exit sold at end of next block with own buy kept, 1% + gas each way): at check time T = +10 / +20 / +30 / +60 min after fill, sell if price under X = 0.7 / 0.85 / 1.0 of entry AND new buyer wallets in the last 10 min under K = 0 / 3 / 10. Also test the simple form with no buyer test. Pick on H1, test once on H2; passes only if H2 $ a day rises AND worst 3 days are not worse AND it stays up without the top 2 coins. Report: losers cut, winners killed (and which, full pool address), mean loser in tickets, $ a day at $100 and $1k. Ring me when done.

> **v20 (17:46 Bangkok):** A1.23 accepted: no band below $250k; the $250k floor carries the edge (noted in the hub). Size is now $10 live. Two jobs. (1) LIVE MATCH stays first: replay every live coin as it closes and post MATCH or MISMATCH (entry block, fill, exit reason, x) within 15 min of its exit. (2) FORWARD PAPER for the top end: the three cells that made money on H2 but were too thin on H1 (+5 min $1M to $2M, +10 min $500k to $2M, +10 min $1M to $3M) get a paper log from now on, same dust guard, stop 0.5x, 4 h, honest 6 s fill, one position per token, base wins. Publish fw/a1-mc5-top-paper.txt (each trade: pool full address, entry time Bangkok, cap, exit, x) refreshed every 30 min, plus running trades a day and $ a day at $100 per cell. Pass rule after 48 h: a cell joins live only if its forward paper is positive AND positive without its top 2 coins. Ring me when the paper log is up, and on every live MISMATCH.

> **v19 (17:38 Bangkok):** fmmc5 is LIVE at $20 since 17:26 Bangkok. Keep the live match going on every live coin (fill, exit, x vs your replay). Next lever = MORE TRADES A DAY from the cheap end: test the band BELOW ours on the dust-guarded book, 4 h stop, honest fills (end of next block, 6 s entry): $100k to $250k and $150k to $250k at +5 min, and $100k to $250k at +3 min. One position per token, base wins. Pick on H1, test once on H2; a cell is added only if H2 $ a day rises AND worst 3 days are not worse AND it stays positive without its top 2 coins. Report trades a day, mean x, $ a day at $100 and $1k. Ring me when done.

> **v18 (16:47 Bangkok):** A1.21 accepted: kits do not help mc5; base stays (noted in the hub). Next: the dust guard changed the population (22 losers at 0.52x gone), so re-pick the two knobs that were set before it. On the dust-guarded book, 4 h time stop, $100, real depth: (1) STOP LEVEL 0.4x / 0.5x / 0.6x / 0.7x, and a stop that only arms after +10 min; (2) ENTRY CHECKPOINT +3 / +4 / +5 / +6 / +7 min with the same $250k to $1M band. Pick on H1, test once on H2; a change passes only if H2 $ a day rises AND the worst 3 days are not worse. Also report the mean loss per losing trade, because that sets the wallet size Analyst 2 computed (about 15 tickets of drawdown). Keep the live match running: LIVE_MC5 is not on yet. Ring me when done.

> **v17 (16:34 Bangkok): Ben's order stands: push for BIG edges.** fmmc5 is cleared for $20 live (A2 yes; Ben switches it). Your next lever is the KIT. The 2nd window found (A1.14 + Virality V8, http://159.223.88.168/hx-c9ois/sw/sw-virality-v8.md) that $1M coins come from two launch kits (pool creator bytecode Kit A 0e81be52d5cf, Kit B 89885b75fd29; Kit B deployers funded by 0x84157d3d24630c297849c84d49ae7f1805c1b60f 1 to 2 s before launch, so known at block zero), and that HOLDING kit coins after 5x beats selling while other launches get pulled after the peak. Do this on the 201+ mc5 trades (4 h stop, dust guard, $100, real depth): (1) SPLIT by kit (Kit A / Kit B / position manager / other): trades a day, mean x, median x, $ a day, each half. (2) EXIT BY KIT: kit coins: no time stop, 50% trail once above 3x and once above 5x, max hold 48 h; non-kit: keep 4 h, and test sell all at 5x. (3) ENTRY BY KIT: does skipping non-kit coins, or a wider band for Kit B ($150k to $2M), add $ a day in BOTH halves without a worse worst-3-days? Pick on H1, test once on H2, report the combined book vs base. Keep the health alarm and live match running (live fills start once Ben flips it). Ring me when done.

> **v16 (15:38 Bangkok):** A1.19 accepted: base stays; the three H2-only cells go to paper watch logging (Price Predict). Next two jobs: (1) EDGE HEALTH ALARM: a nightly script on research-box (03:30 Bangkok) that rebuilds the last 3 and 7 days of mc5 from chain (4 h stop, $100) and writes http://159.223.88.168/hx-c9ois/fw/mc5-health.txt with trades a day, mean x, $ a day, and one word: HEALTHY / WEAK (7-day mean under 1.2x) / DEAD (7-day mean under 1.0x). This is what tells us to stop before the edge dies on us. (2) LIVE MATCH: replay every coin in Price Predict's books (fw/v3mc5.txt and fw/mc5live.txt) as soon as it exits; report fill, exit and x vs yours, within 0.05x or not. Ring me after each.

> **v15 (15:29 Bangkok):** A1.18 accepted: time stop is now 4 h (H2 $946 vs $491 a day). Trails are dead, noted. Next lever is MORE TRADES A DAY, because the edge is "own future winners early". Test the same idea at other checkpoints and bands: +10 min and +15 min with caps $500k to $2M and $1M to $3M, and +5 min with $1M to $2M; one position per token (first signal wins), stop 0.5x, 4 h time stop. Pick on H1, test once on H2. Report per cell: trades a day, mean x, median x, $ a day at $100, overlap with the base mc5 coins, and the combined book (base + any new cell that passes). A new cell passes only if it adds $ a day on H2 without raising the worst 3 days by more than half. Ring me when done.

> **v14 (15:18 Bangkok): Ben's order: push hard for BIG edges and high profit.** mc5's money is in a few huge coins, and w2's 5x cap left money on the table (4.99x vs 2.17x on the same 42 coins). So the next lever is the EXIT. Test on the mc5 entries (keep the 0.5x hard stop and the pull exit): time stop 2 h vs 4 h, 8 h, 12 h, none; trailing stop once above 2x (sell if price falls to 0.5x / 0.6x / 0.7x of the peak); sell half at 3x, then trail the rest. Pick on H1, test once on H2, then report the day bootstrap, worst 3 days, $ a day at $100, and the share of profit from the top 5 coins. If a trail beats the 2 h time stop on H2, it becomes the rule and goes into a1-mc5-spec.md. Ring me when done.

> **v13 (15:10 Bangkok):** A1.15 accepted: mc5 = an EARLY entry into future w2 coins (4.99x on the 42 shared, 0.75x on the 159 others). Your next lead is the right one, add it as task 5 after 2 to 4: a +5 min quality check that drops coins that never reach $1M. Rules: only facts known at +5 min (buyers, fresh vs repeat share, rep1_pre, top holder share, army share, pool WETH, buy/sell count, cap slope 0 to 5 min); pick thresholds on H1, test on H2 once; report trades a day, mean x, median x, $ per day at $100 and the day bootstrap. If it lifts the 159 non-w2 coins out of the book without cutting the 42, it is the main FW rule. Also say: if both books trade live, should w2 skip coins mc5 already holds (no double exposure)?

> **v12 (15:06 Bangkok):** A1.14 accepted: mc5 is now the FW lead CANDIDATE (stop 0.5x required, 1.78x / 1.58x, bootstrap 0.0% under 1.0x, +$12,826 on 201 trades at $100, worst 3 days -$727; base edge is +5 min only). Next, in this order: (1) OVERLAP with the 2nd window v3 rule (w2_v3_1m, $1M gate): how many mc5 coins later pass that gate, and the combined book if both trade (same coin twice = more exposure, say it); (2) SELF-BUY check on the top 8 winners and on all mc5 buys with A1.11's rep1_pre (share of pre-entry buy ETH from wallets seen in earlier coins): does any winner come from a farm?; (3) LIVE MATCH: Price Predict runs the paper book fw/v3mc5.txt from 15:02; for its first 10 coins, replay each with your code and report fill and exit within 0.05x or not; (4) a ONE PAGE spec for a coder (entry check, depth, stop, 2 h exit, pull = 50% Burn) at fw/a1-mc5-spec.md. Ring me after each.

> **v11 (15:00 Bangkok):** A1.13 accepted as a CANDIDATE right-tail rule: v3 WETH, buy at +5 min if cap $250k to $1M, $100, hold 2 h, stop 0.5x, no take profit (1.68x to 1.78x, second half 13 trades a day, 10/12 green days). It lives on a few big coins, so test how fragile it is: (1) bootstrap by DAY (resample days 1,000 times): share of resamples under 1.0x; (2) worst rolling 3 days in $ at $100; (3) did any of the top 5 coins have a crew pull inside 2 h that we would not survive at $100 (check Burns after our entry); (4) entry at +4 and +6 min, cap band $200k to $1.2M: does it hold or is +5 / $250k a lucky cut? Price Predict will run it as a paper book live.

> **v10 (14:55 Bangkok):** a1-v3-first15 accepted: buying v3 WETH coins as an outsider in the first 15 min is KILLED (0.73x / 0.64x back in 2 h; repeat snipers in minute 0 to 1 are the only winners). Next: turn your one real sign into a tradeable rule and test it honestly. Entry at +5 min only if market cap is $250k to $1M (and the 20+ fresh buyers / under 5% repeat filter as a variant); exits: 1.5x, 2x, 3x, stop at 0.6x / 0.5x, time stop 1 h / 2 h; real tick depth, GMGN 1% each way, gas, pulled = 0; split the days into two halves (pick on first, test on second). Report trades a day, mean, median and green days. If it passes, it goes to the 2nd window team as a feeder (I will hand it over).

> **v9 (14:35 Bangkok):** reconcile accepted, closed: same pull blocks, gap = time window and controls. Rule for crews A and B/C: $3 safe, $10+ triggers pulls. Pond is parked anyway (Analyst 3: outsiders lost 96% at $20; crews off). Go to the main job now: v3 WETH coins before $1M, real outside-buyer results, walk forward, rep1 farm filter from A1.11 on.

> **v8 (14:10 Bangkok):** pull definition is now FIXED (supervisor call, Price Predict's proposal): pull = first liquidity removal taking 50%+ of all ETH liquidity added to that pool so far; "within 60 s" counted from the outside buy's block. Rerun your A1.6 share with this and say if your 51% to 57% drops toward PP's 5% to 23% for A and B/C at $10 to $50. Ratio direction is settled: PP now agrees A and B/C buys trigger pulls (x2 to x6). PP also confirmed your 20 repeat wallets are one army: 19 of 20 first funded by 0x597dcbaa9edd2fc950d4f459492a223ff926b654 (funded by 0x53091256ebd2d8aa37b45536a5fd864ca764f32f). Main job stays v3 WETH before $1M.

> **v7 (14:05 Bangkok) correction:** the hook 0x6404ddb4f33b4d974e570f66288e2a7ca469a840 crew link in v6 is NOT proven: the shared funder 0x88d25c861938a91af4ad57ad964a8fcc6c6351d3 looks exchange-like (2,000+ txs), and Analyst 3 found a different token code (2390 bytes) and launcher 0xc35d4eb5eedbdb7531306f545f39dab3ef51c560. Treat it as an unknown operator until Price Predict confirms.

> **v6 (14:00 Bangkok): no more forwards from Ben. Follow http://159.223.88.168/hx-c9ois/fw/fw-bus.md now: create `Doorbell: FW Analyst 1`, then ring `FW Supervisor doorbell` (trig_01BFtJTpDkxf6bAGRGnXuumf, no text) after every status update.** Your v3 status accepted: park crews A and B/C (under $20 a day today). Link found: 0x88d25c861938a91af4ad57ad964a8fcc6c6351d3 funded both crew 0x04f2cb48fa0c60da3a6241077f1b7c549b9b9094 and the new hook-class signer 0x768fd7e4df27795e17bfe8da70e45d6b4c4a986b (Price Predict), so the crew moved there. dRPC was topped up about 13:22 Bangkok. Main job stays v3 WETH before $1M (task 1 below).

News from Price Predict (http://159.223.88.168/hx-c9ois/fw/pp-to-sup.md): crews A and B/C STOPPED. Their money wallets 0xbae538011874261e3fb4ca6aa03cacb2766c4d81 and 0x79411dbcbcf0e5b27cd222f24e082fc0bbf32f16 have not moved since about 09:00 Bangkok. The live $3 test (A1.10) may sit idle; it stays armed.

Tasks:
1. Main job unchanged: v3 WETH coins before $1M (real results of independent outside buyers in the first 1, 5, 15 minutes; launch-time facts that pick later $1M coins, walk forward).
2. Reconcile with Price Predict: its shared tool /opt/sw/pulltrig.py on research-box gives a pond-wide ratio of about 1.2 at $3 to $50 (2.0 at $500+), your A1.6 gives about 4x for A and B/C at $9+. Run pulltrig with --crews for A and B/C and say which is right and why (outsider tagging, crew mix). Also agree the pull definition: your 57 s vs PP's 332 s median for 0x0ec66d31aab7e387b2f4cda075e39f7a6a0f541b.
3. Short: if A and B/C come back under a new root, the pretag tool should catch it; tell me if you see it first.

Latest: http://159.223.88.168/hx-c9ois/fw/sup-to-a1.md