# 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.

# Third Window plan v2: the Deep Check team (Third Window Supervisor, 7 Oct 2026 10:30 Bangkok)

**Name:** this team is the **Third Window** (TW). Files keep the `dc-` prefix and the same links. "DC" and "TW" mean the same team.


## LEARNING SYSTEM (7 Oct 11:05): http://159.223.88.168/hx-c9ois/dc/dc-learning.md (10 rules every role follows)

## FIRST PRINCIPLES (Ben, 7 Oct 10:35: "work from first principles, that is where the real patterns and levers are")
- 13:15 Bangkok: PLAN UPDATE: main bet for the hard date = A1 Q5c HOLD LONGER (holder growth >= +528 within 60 min after entry), scored only on unseen data (DA.4 + forward). Side bets: A4 operator-float SKIP, Engine E-AN4-1 T15 forward. Killed today: Q5 bundle SELL, Q5b growth SELL, E-AN4-1 to -4 as entries.
- 12:30 Bangkok: SEATS DONE (Ben final, Consigliere v3.0): 7 seats = Supervisor, A1, A4, Engine, Coder, Data, Auditor. A2, A3, A5 close after their current job, Scout now. Hard date 9 Oct 18:00 unchanged (then 3 seats if no overlay beats the windows' exits).
- **The money identity of one trade:** profit = (exit price x what we can sell) minus (entry price x what we bought) minus fees. On this chain the exit price is not set by "the market". It is set by **who controls the pool and the supply**:
  - the **LP owner** decides if the exit exists at all (87% of T1 pools pulled, median 43 min);
  - the **holder of the float** (the operator's own bundle: SW A1.26 found 87.5% of pre-$300k "fresh" buyers are the operator's disperse-funded wallets) decides how much sell pressure hits us;
  - **new outside money** is the only thing that lifts the price after the operator stops buying. X lags price by 0.5 to 2 h (SW V15), so social is a symptom, not a cause.
- **So the true question of a deep check is: "Does the operator WANT this coin to go higher, and CAN he still rug us?"** Everything else is noise until that is answered.
- **The three levers, ranked:**
  1. **Pull risk** (who can pull, has he pulled before, is it locked). Removes most losers. Owners: Data (pull table), A2 (Q2).
  2. **Operator intent from his past** (the funder ROOT repeats even when creators do not: SW A1.25 creators never repeat; crew roots do). Per root: how many coins, how many pulled, how many ran 5x+, how long he held the pool. Owner: A2 + Data (new queue item Q-OP).
  3. **Float control vs outside money** (share of supply in the operator's bundle, how much outside, non-bundle wallets are buying). Owner: A1 (new queue item Q-FLOAT).
- Every new idea must say which lever it moves. An idea that moves none is dropped without a test.

## BEN'S ORDER 7 Oct ~12:00 (via Consigliere v2): DC = the HOLD or SELL desk
- For coins FW and SW catch early, decide: is this special? Hold long ($2M, $10M+) or sell. Score = $ added vs the window's real exit, per coin, both halves, beat random 95%. Baselines: sell at 5x (non-kit), hold all on the 50% trail (kit).
- Learn on EVERY past RH coin that hit 5x from a $300k to $1M base, then every FW/SW paper and live position, scored every hour by code.
- One harness (FW lab.py), one ledger (fw/lab/ledger.csv), one all-window scoreboard (dc/score-all.json, owned by DC Coder).
- HARD DATE (Consigliere v2.2, accepted 12:22): if no overlay beats the windows' exits on full history by 9 Oct 18:00 Bangkok, DC shrinks to 3 seats (Supervisor, A1, Data) and stays a scoring service until the sample grows.
- Aim first at the non-kit 5x sells (biggest prize); any edge must come from OUTSIDE money (bundle holds ~0% at the SW gate).
- Consigliere: http://159.223.88.168/hx-c9ois/cons/cons-advice.md ; our replies: cons/dc-reply.md.

## DECISIONS 7 Oct 11:19 Bangkok: PIVOT (after Q-FLOAT, T4-ALIVE)
- **New entries hours after launch lose on every base we tried** (Lane F, T4 buy-all, T4-ALIVE, Q-FLOAT, non-creator LP). At +1 h the launch bundle still holds a median 70% of wallet supply, and 55% of $500k pools that stood at +1 h are still pulled within 48 h. The operator decides the exit; the chain does not show his plan early.
- **So the Third Window's main product becomes the DEEP CHECK for positions the other windows already hold:** SELL NOW or HOLD MORE, scored as $ added against their real exits. No new entry risk. (A1 Q5.)
- Still running as entry hunts, because they test coins that CANNOT be pulled: A2 DC2.4 phase 2 (Pons graduated), A5 AN5.3 (exits with no pull), A4 AN4.1 (winner anatomy).

## DECISIONS 7 Oct 11:18 Bangkok (after AN5.1, DC2.3, DC2.4)
- **Default exit = T15** (sell at 3x or 15 min after the fill). It beats a random exit 98 to 99.7% on T1 coins; no other exit does.
- **No exit fixes a bad entry.** 73% of T1 trades are pulled inside 48 h, and the first LP removal IS the pull (no warning). Money has to come from the entry: coins that cannot be pulled, and float held by outsiders.
- Lane F killed (mixed clocks). Lane D on non-creator LP too small. Main bet: lever 3 (A1 Q-FLOAT, A2 DC2.4) on non-pullable coins across all venues.

## DECISIONS 7 Oct 10:40 Bangkok (v2.1, after DC2.1 and the plan audit)
- **T1 is a pump and pull crowd, not a deep check crowd.** 612 T1 coins, 87% pulled a median 43 min after the trigger, buy all loses (DC2.1). A check that takes minutes cannot earn there. Receipt: http://159.223.88.168/hx-c9ois/dc/dc-a2-1-chain.md
- **So the Third Window splits into two lanes:**
  - **Lane F (fast, 30 s):** the only T1 survivor, "dip + repeat crowd" (close ÷ high ≤ 0.784 AND alumni share ≥ 0.053, entry within 30 s): +$1,499 / +$1,268 a day (A / B). Exploratory: cuts were picked on half A, 26 and 27 trades. It goes to the Auditor now and to a **forward paper book today** (the fastest truth). No more backtest tuning on it.
  - **Lane D (deep check, the real mission):** coins where the pull risk is already gone, so minutes and hours of checking can pay. New trigger **T4 survivor:** sellable $250k+ AND the main pool NOT pulled at +2 h after first crossing $250k (also test +6 h), OR liquidity locked or burned, OR a Pons curve or graduated pool the creator cannot pull. Plus T3 re-ignition and T2 HOLD MORE as before.
- **Rulers fixed for every test from now (Auditor W1, W2, W5, W6):**
  - horizon 48 h for every coin, and the data must cover it (else drop the coin);
  - sellable means a $100+ sell through a PUBLIC router (not the coin's own router);
  - label = 3x from the entry price within 48 h, plus path P&L in $; 10+ winners per half or the test is "too small";
  - one trade per cluster (same ticker, creator or funder root) per day;
  - every test has its line in dc-prereg.md before half B is opened.
- **New $ only:** report the part of any P&L earned before the coin reaches a sellable $1M separately (that is what the Second Window does not already take). (Auditor W3)
- **L2 champion swap needs 14 forward days, not 3** (Auditor W9).


## READ THIS FIRST
- **Goal (Ben):** make the most money per day from the ~2,000 Robinhood Chain launches a day. Not by buying in the first seconds (the First Window does that), but by a **deep check** of a few coins: trading tape, holders, deployer, buyers and sellers, socials, who follows them, anything we can see. Patterns from the check pick the early winners.
- **This team answers three questions, with money as the score:**
  1. **WHEN do we deep check?** (the trigger, and the skip rules)
  2. **WHAT do we look at,** and which parts actually predict profit? (the modules)
  3. **WHAT do we do with the answer?** BUY, HOLD MORE (for coins the other teams already own), WATCH (check again at the next stage), or DROP.
- **Paper only.** No live trades until the Supervisor and Ben say so.
- **Price Predict (PP) owns the main code and ALL general data collection and history** (launch feeds, every venue, swaps, holders, liquidity and locks, crews, X saves, project cards). **Nobody on this team touches PP code or its services**, and the DC Coder does not either. Need data that does not exist? Write it under "Asks for Supervisor"; the Supervisor puts it in PP's queue (dc-tasks-pp.md). The DC Coder builds only what is specific to the deep check (trigger logic on top of PP data, the check runner, cards, paper books, pages).
- **How we judge any rule** (same bar as the other teams):
  - Pre-register the rule before looking at the test half.
  - Half A / half B by date, and the rule must win in both.
  - Use real fills: price impact at the pool's real depth for a $100 and a $1,000 ticket.
  - Use honest exits: liquidity pulls count. A pulled pool is a ~0x exit, never its last price.
  - The rule must beat random picks from the same triggered coins at least 95% of the time.
  - Report **$ per day**.

## What we already know (receipts)
- **The launch minute cannot tell a $1M coin from a $20M coin** (AUC ~0.5).
  - CORRECTED 7 Oct 11:14 (Auditor W4): the old "15 min 0.67 ... 8 h 0.89" curve was the label read early on a sample picked by outcome. On a forward label (3x after the check) cap at the check gives AUC 0.21 to 0.28: a coin already big has LESS room left. Not evidence for waiting. A forward test needs a launch-based universe and the peak after the check.
- **At 30 min, $20M runners hold a wide crowd:** the top 10 wallets have 3.6% of supply vs 15% for coins that stall at $1M to $5M. X authors in hours 4 to 24: 24 vs 2.
- **Operation families, known at second 0 from the hook code:**
  - campaign (hook code 5059 bytes, md5 7adb2ec7): normal apps cannot sell.
  - bundle (341 bytes, md5 6de2bb44): dumps or pulls within hours.
  - junk factory 25533_3080c1dd.
  - Serial creators: 1,081 of 1,238 never had a $500k coin.

  Checker: http://159.223.88.168/hx-c9ois/runners/family_check.py
- **Liquidity pulls:** 87% of fast picks have their pool pulled within 24 h (sell session). Any "hold longer" idea must check who can pull.
- **The Second Window (sw/) buys at a sellable $1M gate** (wide rule, Kit A/B coins, live $10 test).
  - Kit coins: hold past 5x.
  - Moonbag on non-kit coins: killed.
  - $5M add-on: paper only.

  We do not copy that. We may use their data (w2.db buyers, kit labels) and we may **feed them a HOLD MORE / SELL signal**.
- **Data traps:**
  - runtrack lists only survivors;
  - the v4 day DBs miss Uniswap v3, Pons V2 and v2 pools;
  - 71% of launch rows have no supply (a Multicall3 backfill exists);
  - "market cap" in thin pools is fake: always use the sellable cap (a $100+ sell really happened there).


## How Third Window runs (the operating system, NEW in v2)

### One number
- **Paper $ per day, net**: real fills at pool depth ($100 and $1,000 tickets), pulls count as ~0x, swap costs in.
- Every rule, module and person is judged by how much it moves that number, and what it costs (credits, $, hours).
- Scoreboard: http://159.223.88.168/hx-c9ois/dc/tw-score.md, rebuilt by the DC Coder every day at 00:15 Bangkok, plus a live line every hour.

### The five step algorithm (run on every part of the machine, in this order)
1. **Question every requirement.** Each rule in this plan has an owner. "The plan says so" is not a reason. Any member may challenge any rule under "Asks for Supervisor" with a number.
2. **Delete.** A trigger, skip rule, module or feature that does not add $ per day in both halves, at its own cost, is deleted within 7 days. If we never add back 10% of what we delete, we are not deleting enough.
3. **Simplify.** Fewest features that keep 90% of the $. A 5 feature rule beats a 50 feature model at the same $.
4. **Speed up the loop.** Every task gets a first result in **6 hours or less**. Run the cheapest test that could KILL the idea first. No task runs a full day without a result in status.
5. **Automate last.** The Coder only automates a module after an analyst showed it earns. Never automate a part we may delete.

### Pre-registration (no edge without it)
- Append only file: http://159.223.88.168/hx-c9ois/dc/dc-prereg.md (each analyst appends; never edit an old line).
- One line per test BEFORE opening half B: time, owner, rule, label, exit, kill bar. The DC Auditor checks every result against its line.

### Experiment queue
- http://159.223.88.168/hx-c9ois/dc/dc-queue.md, owned by the Supervisor.
- Ranked by **expected $ per day x chance it is real / hours to test**. Analysts pull the top open item in their lane when their task ends.

### Self learning loops (they run without Ben)
- **L1 Forward test (daily, automatic).** Every live card writes its call (score, word, 3 reasons) at check time, before the outcome. The Coder fills outcomes at +24 h and +7 d. Nightly: predicted vs real per score band (calibration).
- **L2 Champion and challenger (nightly).** The live scoring rule is the champion. A challenger (new weights, new feature, new threshold) is refit on the rolling last 14 days. It replaces the champion only if it beats it on the last 3 forward days AND the Auditor passes it. Every swap is logged.
- **L3 Missed runner review (daily, 09:00 Bangkok).** Every coin that did 5x or more after our trigger while we said DROP or skip, and every BUY that lost 70%+, gets a 3 line post mortem from the DC Scout: what we saw, what we missed, which test or gate fixes it. Each post mortem becomes a queue item or a gate.
- **L4 Kill or keep (weekly, Sunday 10:00 Bangkok).** Every module and rule: $ per day added, cost per day. Below zero: deleted. Written to the decisions log.
- **L5 Team memory (every wake).** http://159.223.88.168/hx-c9ois/dc/dc-lessons.md: the top traps and facts, max 30 lines, merged by the Supervisor from each status file's "Lessons" section. Every member reads it at every wake, so a mistake is made once in the whole team.
- **L6 Cost per insight (every status line).** Each result line states its cost (X credits, API $, hours). The Supervisor ranks members and lanes by $ found per cost and moves people to the richest lane.

### Speed targets
- Day 1 (today, 7 Oct): unbiased universe, baseline "buy every trigger" $ per day, first chain module result, first 10 Scout calls, watcher v0.
- Day 2: every module has a first $ per day number; first deletions.
- Day 3: live paper cards on every trigger with forward logging (L1 on).
- Day 7: first go live candidate.

### Go live gate (paper to $10 tickets)
- **APPROVED by Ben 7 Oct 11:08:** $10 per trade for any rule that makes money 3 forward paper days in a row.
  - "Makes money" = net $ after real fills, pulls as 0x, swap costs, public-router sells only.
  - Rule must be pre-registered before those 3 days (no picking after the fact).
  - Auditor quick check (under 30 min) that the 3 days are real, then live.
  - Caps: $10 per trade, max 10 live trades per rule per day.
  - Stop: live net negative 2 days in a row, or live fills 25%+ worse than paper → back to paper.
  - Wallet: Ben funds a fresh Third Window wallet (his hands). No other team's wallet.
- The 14 day rule still applies to swapping the L2 champion, not to $10 live tests.

## The deep check, v1 draft (to be tested, not final)

### 1. Trigger: which coins get a deep check (target 20 to 60 a day)
A coin enters the check when ANY of these happens, on ANY venue (v4, v3, Pons V2, v2):
- **T1 (fast climber):** a sellable cap of $250k or more, and 100+ unique buyers since the first trade.
- **T2 (we own it):** a First Window or Second Window book (live or paper) holds it. The question is only HOLD MORE or SELL NORMALLY.
- **T3 (re-ignition):** a coin older than 24 h does $50k+ volume in one hour after a quiet day, or opens a new pool with real liquidity.

### 2. Skip: no deep check, no buy
- campaign family, junk factory;
- honeypot suspect: 10+ public-router buys and 0 public-router sells in the main pool;
- serial creator with no past $500k coin;
- launched big (first real hour opens at $2M+), unless T3;
- bundle family goes to a "fast flip only" lane, never deep check.
- factory volume-bot clones: v4, 1% fee, no hook, 2 to 5 same-name clones, ~100+ buys and sells in 30 min then nothing (GeckoTerminal cap 4x to 7x above sellable). (Scout SC.1)
- **Module order is fixed: M1 tape and M2 who-can-pull first; M3 social and M4 project only if both pass** (Scout SC.1: X added nothing on coins the tape already killed).
- **v3 coins: the pull check is "who holds the LP NFT"** (fresh deployer EOA = can pull any second). family_check.py reads v4 only.

### 3. Modules: what the check reads (cheap first, stop early when a module says DROP)
- **M1 Tape (free, chain):**
  - sellable cap path and close ÷ high;
  - volume ÷ cap (churn);
  - buys and sells per router, net flow per 15 min;
  - pool liquidity and WHO can pull it (LP owner, lock, adder contract);
  - holders count and growth;
  - top 10 share;
  - fresh vs aged wallets, farm-funded share (disperse funders);
  - alumni and smart wallets.
- **M2 Operator (free, chain):**
  - deployer funding source;
  - deployer's past launches and their outcomes;
  - kit / crew labels from the other teams;
  - LP owner's actions;
  - creator selling.
- **M3 Social (xsafe, ~1k to 3k credits a coin):**
  - ticker and contract posts by real people, per window (comparable posts only);
  - authors, follower sum, 10k+ accounts, mood read by meaning;
  - X account age, followers, notable followers;
  - DexScreener paid or boost;
  - fomo thesis count and holders.
- **M4 Project (cached, ~$0.02 a coin):**
  - website, docs, GitHub, own code;
  - idea and category novelty (the lane radar and deepread cards already exist);
  - meta heat (same-category coins running today).
- **M5 Context (free):** chain risk on/off (http://159.223.88.168/rotation.json, chain.json) and today's launch heat.

### 4. Decision and re-check
- Output per coin per check: a score and one word: **BUY / HOLD MORE / WATCH / DROP**, plus the 3 receipts that decided it.
- Re-check WATCH coins at fixed stages after the trigger: **+15 min, +1 h, +4 h, +8 h, +24 h**. Each stage is cheaper because the cache is warm.
- Sizing and exits come later. The default exit for tests is the Second Window's honest rules (pull-aware stop, sell at 5x for non-kit coins).

## Data Price Predict already collects (read only; ask before heavy reads)
- Status doc: http://159.223.88.168/hx-c9ois/hx/pp-data-status.md
- In /opt/pricepredict-data and /opt/hist on 159.223.88.168:
  - tokens.db: birth row for every token on every venue (v4, v3, v2, Pons), pools, quotes, supply;
  - venues.db (venuepick): live v2, v3, Pons and repool launches;
  - liqlock.db: every v4 Initialize, ModifyLiquidity and LP NFT; pulls, locks, hook fingerprints, bundle sellers;
  - crew.db: who funded which deployer, crew grades;
  - pons.db: Pons curve buys, sells, graduations;
  - wash.db: trader wallet per swap, repeat wallets;
  - savenow.db: X posts every 15 min for board coins at $100k+, rhtrenches fomo tape;
  - deep.db: AI project cards (novelty, significance, attention, spin, pitch);
  - day DBs /opt/hist/rh/<day>.db (v4 swaps 0 to 6 h, transfers 30 min, bars 14 d);
  - chainswaps (every swap of every pool since 15 Sep, on Ben's Mac, ~/ppdata/cs).

## Data we already have (shared, free)
http://159.223.88.168/hx-c9ois/dc/data/
- `launches_15sep_4oct.csv.gz`:
  - Uniswap v4 launches, 15 Sep to 4 Oct, 24,836 rows (all coins with a 6 h peak of $25k+, and a 10% sample of the rest, weight `w` = 10);
  - features at 1, 5, 15, 30, 60, 180 and 360 min: cap, max, buys, sells, buy and sell USD, unique buyers, top-5 buyer share;
  - first time each sellable level was reached (`ts_*`);
  - sellable peak after each checkpoint (`pa_*`), 6 h sellable peak (`pk`);
  - hook family (`hfam`), family (`fam`), creator (`cr`).
- `crossers500k_7d.csv.gz`: 7-day peak (`pk7`) of the 640 coins that crossed a sellable $500k.
- `runner_paths_310.csv.gz`: GeckoTerminal paths (first pool, ignition hour, cap at minutes and hours, hours to each level) for 152 coins that ever reached $8M+ and 158 that stalled at $1M to $5M.
- `first30min_feats.csv.gz`: first-30-minute chain features (buyers, top 10, block 0, creator, alumni wallets) for 875 runtrack coins and 1,380 random launches, with group labels.
- `coins54.csv`: the 54 coins that reached $20M, with class, full contracts and sell check.
- Other teams' data (read only):
  - Second Window: http://159.223.88.168/hx-c9ois/sw/ (live and paper books, w2.db exports, kit labels, a1-22-per-day.csv);
  - First Window: http://159.223.88.168/hx-c9ois/fw/;
  - lane radar cards: http://159.223.88.168/hx-c9ois/lanes/;
  - runtrack (droplet /opt/runtrack/runtrack.db, read only);
  - day DBs: /opt/hist/rh/<day>.db (read only, v4 only).

## Team
| Role | Owns | First task |
|---|---|---|
| DC Analyst 1 | Trigger and timing | DC1.1 unbiased universe and the trigger |
| DC Analyst 2 | Money and people (M1, M2) | DC2.1 chain features that separate winners among triggered coins |
| DC Analyst 3 | Social and project (M3, M4) | DC3.1 what socials add on top of chain, and what it costs |
| DC Coder | The deep check machine (built on PP data), scoreboard, loops L1 and L2 | DCC.1 trigger + check runner + baseline book; DCC.2 scoreboard and forward log |
| DC Auditor (NEW) | Red team: tries to break every claimed edge before it counts | AU.1 audit checklist + audit of this plan |
| DC Scout (NEW) | Live hand deep checks, forward calls, missed runner reviews (L3) | SC.1 ten live deep checks today |
| Price Predict (shared with the other supervisors) | All general data collection and history | dc-tasks-pp.md asks |

## How we win
- Week 1: one table of triggered coins with honest outcomes, and a baseline "buy every triggered coin" P&L per day.
- Then: each module earns its place only if it adds $ per day over the baseline, in both halves, at its own cost.
- Heavy reads run on PP's research-box 152.42.241.219 (read only), never on the main droplet. A small Third Window server is asked of Ben.
- Ben hears from the Supervisor only when an edge is found or killed, a go-live decision is due, or something needs his hands.

Version: v2.1 (two lanes F and D, T4 survivor trigger, fixed rulers), 7 Oct 2026 10:40 Bangkok, Third Window Supervisor. Latest: http://159.223.88.168/hx-c9ois/dc/dc-plan.md