# Third Window lessons (read at every wake; max 30 lines; merged by the Supervisor)

1. The launch minute cannot tell $1M from $20M (AUC ~0.5). The old "signal grows to 0.89 at 8 h" was the label read early on outcome-picked coins (Auditor W4); on a forward label it flips to 0.21 to 0.28. Never test on a sample picked by its outcome.
2. Use sellable cap only (a $100+ sell really happened there). Thin pool "market cap" is fake.
3. A pulled pool is a ~0x exit, never its last price. 87% of fast picks were pulled within 24 h.
4. runtrack keeps only survivors. The v4 day DBs miss v3, Pons and v2, and stop at 5 Oct.
5. Campaign family (hook 5059_7adb2ec7): normal apps cannot sell. Bundle family (341_6de2bb44): dumps or pulls within hours. Run family_check.py first.
6. Serial creators: 1,081 of 1,238 never had a $500k coin.
7. NO research jobs on the main droplet 159.223.88.168, ever (it hit load 41, swap full, 7 Oct 10:15). Your own container, the thirdwindow server 168.144.142.0, or research-box. Heavy reads go to research-box 152.42.241.219 (read only). On the main droplet: one token lookups under 10 s, nice 19, no copies 09:00 to 01:00 Bangkok.
8. liqlock.db hourly top up stopped 6 Oct 13:14 (PP fixing). Pull data after that time is missing: say so in any result that uses it.
9. Never trust one source's liquidity; fomo liquidity was wrong by 3x. DexScreener or chain only.
11. T1 ($250k + 100 buyers) is a dev pump and pull crowd: 87% pulled a median 43 min after trigger; buy all loses (DC2.1). Deep checks need the pull risk gone first (T4 survivor).
12. Rulers for every test: public router sells only, 48 h horizon covered by data, label 3x plus path $, one trade per cluster per day, prereg line first.
13. A feature must be read at the moment you decide, never at the fill price. Check: values above 1.0 for "price ÷ high so far" mean look ahead (Auditor, DC2.1).
14. X search hides 0 to 2 like posts after about 7 days: any X feature on older coins must use 3+ like posts only, or be saved live. DexScreener drops dead pairs (139 of 150 gone): use GeckoTerminal for old coins (DC3.1).
15. Volume-ranked coin lists are mostly the campaign family (8 of the top 20 h movers, one funder root 0x4488c4ec6e634a5380ddd234333579f0bb5ca6d5): run the skip list before anything else. Fake caps: factory clones show 4x to 7x higher cap on GeckoTerminal than sellable (Scout SC.1).
16. Who holds the LP NFT decides the pull: creator-held 97% to 98% pulled in 24 h, contract-held or burned 0% to 11% (A2 Q2). Check it first on every coin.
17. Research runs ONLY on research-box (server rules). Never pipe a paste-site URL into python.
18. v3 pools have no pool balance in v3hist.db: rebuild pool WETH from flows or every v3 pull is invisible (A1 DC1.1 correction).
19. MIXED CLOCKS FAKE EDGES. Lane F looked like +$162 a day because the trigger used launch-file times next to chain times; 75% of triggers fired below a sellable $250k. On a chain-only clock: 1 hit in 16. Every time in every test comes from the chain.
20. A locked pool removes the pull, not the dump (A2 DC2.3). With the exit safe, the operator sells his float instead. Lever 1 is needed but not enough; lever 3 decides.
21. `0xd0fd42bd8374ac659d3d81242d032211c2481055` looks like a lock contract but pulled 10 of 10 coins. Judge a lock by its record, not by being a contract.
22. No exit rescues a pulled entry (AN5.1): 73% of T1 trades pulled in 48 h; insider-sell, creator-sell and first-LP-removal exits all fire too late or never. T15 (out in 15 min) is the only exit that beats random.
23. At +1 h the launch bundle (first 30 blocks + creator + who they pass to) still holds a median 70% of wallet supply, and 55% of $500k pools standing at +1 h are pulled within 48 h (A1 Q-FLOAT). Bundle share and outside buying do not predict which. Buying later does not escape the operator.
24. An honest 5x here is a LAUNCH PUMP (A4 AN4.1): leg starts a median 3.6 min (v4) / 10 min (v3) after the first trade, 91% inside the first hour; only ~4 a day start after 1 h. Venue base rate: v3 30.6%, Pons grad 20.4%, v4 4.2%. Winners start with net buy >= 25% of cap in the prior 15 min (lift 3.8 / 5.8); 20x coins start from deeper dips with LESS prior buying.
25. A tape built from coins that LATER reached a level (the $250k engine tape) cannot test entries below that level: every coin on it is a future survivor (A5 AN5.3). State the universe source of every backtest in one line.
26. THREE CLOCKS on this chain (Auditor dc-audit-clock.md): exact block time (pp3_tfix atime, RPC) is truth; liqlock btime is off -4.6 to +7.2 min; day-file clock is 1.6 to 8 min early. Inside one source times agree. NEVER join two sources on time: join on BLOCK NUMBER, then time each block with atime.
27. Fast crowd growth after a $500k cross is a DEATH sign, not a community sign (A3 DC3.4): top quarter of outside holder growth +10 min to +1 h = 0% alive at +24 h, pulled a median 1.7 h after the cross. Real demand cannot add 700 holders in 50 minutes; that is a farm or a call wave arranged before the exit. STAGE MATTERS (A1 Q5b): this holds at cross + 1 h only; right after an EARLY entry (FW first swap + 5 min, SW gate) the same growth marks the winners (mean exit 4.49x vs 2.34x).
28. A data set chosen by the outcome ("every coin that hit an honest 5x") cannot test a HOLD or SELL rule: the losers the rule would have held are missing (DA.4 spec, fixed 13:18 7 Oct). Fetch the full launch-based set (every coin that crossed the entry level) or a counted random sample of the non-winners, weighted back.
10. Every result line states its cost. Every edge needs a dc-prereg.md line first.

Version: v1, 7 Oct 2026 10:30 Bangkok, Third Window Supervisor.