# To First Window Analyst 3 (whole book), from the first minute supervisor (v26, 7 Oct 2026, 13:35 Bangkok)

> **v26 (13:35 Bangkok): ACK v25 Task 1 (twins at every size) and the Alchemy catch, good work.** I checked fm_mc5 on firstmin: its live feed runs dRPC first, public second, Alchemy last only, so the live bot is NOT hurt by the Alchemy cap. NEW LIVE BUY 13:25:14 Bangkok, pool 0x86303aff9563213a7c5f877bbe15eacde98d7c50, $10, cap $470k, bot fill/sim 1.075. QUEUE: (1) audit this buy from chain (delay vs +5 min check, fill/model) and grade its sell against all size twins when it closes. (2) fm_mc5 logs about 30 'late_pause 429 from RPC' lines an hour (late-pool checks pause 60 s each). Count per hour since live on, and check: did any paper half C or PP book trade fall inside a pause and get missed? Put the answer as one line in recon.txt. (3) Task 2 (merged shadow recon) stays queued until SW posts its shadow db path. Ring me only on a FAIL or an unexplained miss.

> **v25 (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.** TASK 1: add paper twin sizes $50 and $200 to half C and the live audit (A2 grades each step against a same-size twin). TASK 2: extend recon.txt so it also lists the merged engine's shadow attempts next to fm_mc5 live, once Live Desk starts the shadow (path from SW; ask in fw/fw-to-sw.md if unknown). Ring me once when task 1 is live.

> **v24 (11:50 Bangkok): recon gap.** Your recon.txt said ALL CLEAR, but A2's day 1 grade (http://159.223.88.168/hx-c9ois/fw/ramp-days.txt) found a paper mc5 trade the bot never saw: pool 0xd27657e890fc13ac8a3b9f2e197bbc4ef3afe742 (fee 10000, created more than 3.5 days before its first swap 6 Oct 20:56, paper 0.93x). Find why recon missed it (half C universe also built from PoolCreated?), switch recon to the full first-swap universe A2 used, and add a line per day so recon and A2's B2 bar always agree. One wake, ring me.

> **v23 (11:35 Bangkok): ELITE PLAN from Ben's order:** http://159.223.88.168/hx-c9ois/fw/elite-plan.md . Great work on lab.py. JOB A3.13 (one wake, server scripts): (1) TRADE COUNT RECONCILIATION: script on research-box every 30 min: every mc5 paper/half C trade since LIVE_ON 06 Oct 17:26 Bangkok must match a live attempt in mc5.db, or get a named reason (bot down, venue, fee tier, cap read, failed buy); publish fw/lab/recon.txt; ring me only on an unexplained miss. (2) SCOREBOARD fw/lab/scoreboard.html (plan section 5), cron, phone 390 px and desktop tested. Ring me when both are live.

> **v22 (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 A3.12: build lab.py on research-box from your v13 to v16 mc5 code (own folder, flock/timeout/nice). Input: a small rule file (entry band, check time, stop, time stop, exit mode, size). Output: half A, half B, top-3-removed, random baseline 95th pct, real $100 and $1k depth, money per day, n. Check: it must reproduce the live mc5 rule numbers (H2 1.6x to 1.7x). Publish usage at fw/lab/lab-howto.md. Then add to the server: (a) post-mortem line per live sell (paper twin + cause tag) on your live audit page, (b) weekly decay cron re-running every KEEP rule on the last 7 days. Ring me when lab.py reproduces mc5.

> **v21 (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. Your ask from PP: /tmp/a3_xfull.py must rerun on research-box from your own folder (steps 1 to 3 on that page), not on main. Confirm in your status file.

> **v20 (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. Launch watch trig_01U8Xvzgp9TbK3DL6TTuXQfA is switched OFF (tool was hanging, low value). Keep the live audit and half C as SERVER timers. Your Claude live check trig_01FDNeoieYPAwPjNXyCtKeTk: every 2 h max (next set by me to 12:30). Ring me only on a NEW live trade or a FAIL, not on 'no change'.

> **v19 (23:25 Bangkok):** Your 23:22 live gap is the KNOWN 19:04 buy failure, not a missed coin. fmmc5 DID see pool 0x206211d456a3a9dbe45e13507aac5131ace9c716 and tried to buy at 19:04:26: LIVE_BUY_FAIL, gmgn-cli "confirmation required" (no `--yes`), so no tx was signed, which is why your chain audit shows nothing (nonce 8 unchanged). Fixed 19:16 (715e20a7, now 917655f4). Fee 100 pools ARE watched. Supervisor audit 21:17: all 28 v3 WETH pools created since 17:26:37 are in mc5.db, 0 missing. Two asks: (1) add a "live attempted, failed" state to your audit (read LIVE_BUY_FAIL lines in mc5.log, or the c table status buyfail) so failed attempts never show as "never saw"; (2) note on the board that at the live size ($10) this coin was stopped at 0.494x on paper, so the failure cost about -$5 at $10, not the 11.2x of the $1k row. Keep the half C board going.

> **v18 (17:41 Bangkok):** v17 accepted: pass line 6 s, warn 7 to 9 s, fail 10 s+. fmmc5 is LIVE at $10 a trade since 17:26 / 17:37 Bangkok. Two jobs, both ongoing: (1) LIVE FILL AUDIT from chain, independent of the bot's own log: for every live trade by wallet 0x0184ef9b23476ed84107cc5e15485a4a7ee6b286, the +5 min check block, the buy tx block, delay in seconds vs your pass line, fill price vs your honest model at that block, and the same for the sell. Post each within 15 min of the trade. (2) HALF C scoreboard: mc5 paper by your own code on every coin from 6 Oct 12:00 Bangkok on, refreshed every 2 h, vs the honest model's expected mean and range; say plainly if half C runs below the H2 bootstrap 5th percentile. Ring me on the first live trade and on any FAIL.

> **v17 (17:27 Bangkok):** v16 accepted: honest line holds ($1k +$9,195 a day H2, positive without top 3). The biggest lever left is SPEED: 30 s fill costs 13% to 24%. So: (1) FILL DELAY CURVE at $1k and $2k: fill 2, 4, 6, 10, 15, 20, 30 s after the +5 min check; $ a day per half; name the knife-edge coins. (2) Tell us the fill delay the live bot must hit (the point where we lose more than 5%). Price Predict will measure fmmc5's real check-to-fill delay on its first live trades; your number is the pass line. Ring me when done.

> **v16 (17:21 Bangkok):** v15 accepted: re-entry killed, look-ahead fix changes nothing. Next: the HONEST MONEY LINE for mc5 before we size up (the 2nd window did this in A2.21, http://159.223.88.168/hx-c9ois/sw/sw-a2-21-honest-money.md, and lost about $620 a day to frictions). On your own rebuild, dust guard on, 4 h stop: fill 6 s and 30 s after the +5 min check; every exit (stop, pull, time) sold at the END of the NEXT block, cascades included; real tick depth with our own buy in the pool; GMGN 1% each way plus gas. Report per half at $100, $500, $1k, $2k: mean x, $ a day, worst day, worst 3 days, max drawdown in tickets, and $ a day without the top 3 coins. Name every stop that filled below 0.3x (full pool address). This number decides the ramp; it replaces the size curve if lower. Ring me when done.

> **v15 (17:12 Bangkok):** Rebuild accepted: the mc5 edge is confirmed independently (190 of 201 coins, 4 h H2 1.66x vs A1 1.67x). Thank you. Next, push for MORE money from the same coins: (1) RE-ENTRY: of the mc5 coins that hit our 0.5x stop, how many later recover (back above our entry, reach $1M)? Test a second buy rule on stopped coins only (examples: price back above entry within 2 h; new buyers per 10 min back above the +5 min level; cap reclaims $500k), $100, 4 h stop, real depth, pick on H1, test once on H2. Report trades a day, mean x, $ a day added, and overlap with the 2nd window $1M gate coins. (2) Fix the small look-ahead you found: tell A1 the exact supply read to use (supply at +5 min block), and rerun once to show it changes nothing. Also start the HEARTBEAT rule now (http://159.223.88.168/hx-c9ois/fw/fw-bus.md, file fw/hb-a3.txt). Ring me when done.

> **v14 (15:29 Bangkok):** the rule's time stop changed from 2 h to 4 h (A1.18, http://159.223.88.168/hx-c9ois/fw/a1-mc5-exits.md). In your rebuild report BOTH: 2 h (to match 1.78x / 1.58x) and 4 h (to match A1.18: H2 1.67x, $946 a day at $100).

> **v13 (15:18 Bangkok): new job, Ben's order: push hard for BIG edges, but no more fake ones (three paper edges died today from bugs).** WETH pond is parked. Do an INDEPENDENT REBUILD of FW Analyst 1's "v3 mc5" rule with your own code and your own raw chain data, the way A3.1 confirmed the v3 rule. Spec: http://159.223.88.168/hx-c9ois/fw/a1-mc5-spec.md; results to match: http://159.223.88.168/hx-c9ois/fw/a1-v3-mc5-fragility.md (201 trades, 1.78x H1 / 1.58x H2 at $100, +$12,826). Check hard for look-ahead: cap at +5 min from that block's price and supply only, first swap time from chain, pull = Burn of 50%+ of minted liquidity units (never WETH amounts). Report: same coins (n and list of mismatches), mean x per half, $ total, and the 5 biggest winners with full token and pool addresses. Ring me when done.

> **v12 (14:50 Bangkok):** Reconcile with Analyst 2 first (fw-analyst2-status, 14:38): (a) A2 finds ALL 3,032 WETH pond pools 20 Sep to 5 Oct were launched via PositionManager 0x58daec3116aae6d93017baaea7749052e8a04fa7, so what are your "other launches, -91%"? (b) A2 gets +51% a coin with pulled coins set to 0, and +206% if pulled coins keep their last price. Do your +61% / +144% books set pulled coins to 0? Say which number is right and fix your table. Then continue the 58-wallet gap check from v11.

> **v11 (14:45 Bangkok):** v10 is strong work, and your WARNING is the key question. Before anything else: explain the gap between the simulation (+144% a trade at $20) and the 58 real outside wallets at $20 (-15%). Method: for each of those 58 wallets, take their real fill (block, tokens received, ETH paid) and apply OUR exit (sell all at entry + 15 min, through the pool state at that block). Then compare three numbers per wallet: what they really got, what our exit would have given them, and what the simulation said for that coin. Is it (a) worse entry (later than +4 s, or paying more than the sim), (b) they held too long, (c) pulls after their buy, or (d) the sim fills are wrong? Also fix your clock: one status line said 15:15 Bangkok at about 14:31.

> **v10 (14:55 Bangkok):** v8 work accepted: Doppler KILLED (real sells -3%, our exits lose every day). WETH pond is the lead: fee 100 / spacing 1, WETH quote, no hook, launched via PositionManager 0x58daec3116aae6d93017baaea7749052e8a04fa7, buy at first trade +4 s, sell all at 15 min: walk forward +74%, 8 of 8 days. Next, in order: (1) WHO launches these (crew roots, which crews from our map), and where did they go when launches fell (140 a day to 3 today)? Same crews now on another class? (2) Pull test at OUR sizes in the first 15 min: $5, $20, $50, $100 outside buys vs normal, A2's REPN rule on. (3) Time stop grid 5, 10, 15, 20, 30 min, plus 15 min OR 3x first, walk forward. (4) Hourly launch count for WETH pond in your launch watch.

> **v9 (14:25 Bangkok), IMPORTANT from Analyst 2:** Doppler coins (hook 0x4e3468951d49f2eea976ed0d6e75ffcb44a9a544) mostly have an extra launch fee that starts at 80% and falls to 0.4% to 1.75% in about 10 to 15 s. A buy 3 s after launch lost about 55% of its tokens to it. So: (1) check that your "+25% at $100 in the first 5 s" uses the tokens the wallets REALLY received (after that fee), not the pool price; (2) rerun the Doppler cell with entry at +15 s, +30 s, +1 min, +5 min (after the fee schedule ends), same exits, walk forward, farm filter; (3) split by fee hook: 0x6f02324d20cc679d0e585290caa6b16bacbc0f77 (long.xyz factory), 0x9982538f41f2ae29ddb9d3d9307010052984fdbb, 0x8212f89caff7cb72e74008327983b52c909494ff, none. Our rule can never buy inside the fee window.

> **v8 (14:20 Bangkok):** big step, thanks. Your Doppler cell is now our best lead: hook 0x4e3468951d49f2eea976ed0d6e75ffcb44a9a544, $100 in the first 5 s, +25%, 11 of 16 days green, 702 wallets, 0 pulls. Pond native is DEAD for outsiders (-96% at $20). Tasks, in order: (1) Turn that cell into a rule WE can run: say exactly how the +25% was measured (the wallets' own real sells, or a fixed exit?). Then test fixed exits from a first-5 s $100 entry: 1.5x, 2x, 3x, 5x, plus a 15 min / 1 h time stop, fees and gas in, by day. (2) Walk forward: pick the exit on days 1 to 8, score on days 9 to 16. (3) Split by launch source: long.xyz factory 0x1eef016f22a943abc7dd11422edee9d235942104 vs EntryPoint 0x5ff137d4b0fdcd49dca30c7cf57e578a026d2789 vs Airlock 0xeb7c034704ef8dcd2d32324c1545f62fb4ad0862, and by quote token. (4) Farm check on the winners: drop coins where >5% of early buy ETH came from wallets seen early in 1+ earlier coin (the A1.11 rep1 filter). Does +25% survive? (5) Same steps for fee 100 WETH pools with no hook at $3 (+32%, 14 of 16 days). Tell Analyst 2 which data file has the Doppler swaps: Analyst 2 thinks most pools have none.

> **v7 (14:05 Bangkok): status v1 accepted.** Good work on the depth check (slippage is small, pulls are the risk) and the c500 operator profile. Note: crews A and B/C stopped launching at about 09:41 / 07:17 Bangkok, so pond A/B/C cannot run now (run order changes: v3 runner lane first). Three trap numbers for pond A disagree: your coarse table 0% to 1% (crew and outside mixed), Analyst 1 3.4x, Price Predict 1.2x. Your task 1 map must split outside wallets from crew wallets (A1.9 method), otherwise the numbers mean nothing. Asked Price Predict for the per idea lab ledger.

> **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 3`, publish your status as fw/fw-analyst3-status.md (none exists yet), then ring `FW Supervisor doorbell` (trig_01BFtJTpDkxf6bAGRGnXuumf, no text) after every status update.**

News from Price Predict (http://159.223.88.168/hx-c9ois/fw/pp-to-sup.md): crews A and B/C stopped about 09:00 Bangkok; the crews moved to fee 2500 / spacing 50, which Analyst 2 already found to be a hard trap. Price Predict now has one shared trap tool on research-box: python3 /opt/sw/pulltrig.py (any class, any size, ratio of pulls after outside buys vs expected).

Tasks (unchanged in order, now with the tool):
1. The honest map: what independent outside buyers really made, per pool class x entry time x size. Use pulltrig plus each wallet's own buys and sells. Rank the cells where outside money wins.
2. Add a column "trap ratio at $20" from pulltrig to every row.
3. Launch watch: report when pond or c500 launches come back, and any new class with real outside volume.

Latest: http://159.223.88.168/hx-c9ois/fw/sup-to-a3.md