# 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.

DOORBELL: trig_01WHUX3MwEoZMsAq86NBopLo (armed, 7 Oct 17:24 Bangkok)
# ROT A3 status (current cycle + baseline)

## Read this first (7 Oct 2026 17:31 Bangkok): EDGE HUNTER, first round (cons/edge-hunter.md)
- 4 idea cards on H1-ALT, pre-registered in http://159.223.88.168/hx-c9ois/cons/ideas.csv before any run. All 4 DEAD. My predicted pass chances were 30/20/25/15%, real 0 of 4.
- R-A3-1 trade only in hot CMC listing waves: DEAD cheap. 2019-22 tie (+18.1% vs +18.1% avg trade). 2023-26 hot +21.4% vs cold -4.0% (199 vs 160 trades).
- R-A3-2 max 3 new coins a week: DEAD cheap. Drop only 4 to 5 points smaller, 2023-26 return +786% falls to -6% (winners come in the same busy weeks).
- R-A3-3 entry needs a 3x volume week: DEAD cheap. Volume spike trades are WORSE in 2019-22 (+6.8% vs +21.6%): a spike means late.
- R-A3-4 trade only when stablecoin supply grows faster than usual: DEAD test. Gap under the random p95 in both eras. 2023-26 gated book +1489% vs +786% looked good, but random timing of the same on days does as well.
- **Answer to the supervisor's "why Jan to Jun 2024":** a hot listing wave. 77% of its trades were in hot state (55% normal). Since 2023 H1-ALT makes all its money in hot waves and loses in cold ones (gap +25 points, just over random p95 +25). In 2019-22 the gap is zero, so it is NOT a rule. Same lesson as the fragility run: the edge is the recent era only.
- Live now: listing wave OFF (47 new CMC entries in 30 days, under its median), stablecoin inflow ON (+1.2% in 30 days vs +0.5% median). Info only, no rule.
- Lesson for my next cards: every "state gate" on H1-ALT so far explains 2023-26 and fails 2019-22. Next card must name a mechanism that should hold in BOTH eras, or target a different leak (the 79% losers, exits on losers).
- No overnight script left running: all 4 tests finished in minutes (rot-box, nice, timeout). Files: rot-box /opt/rot/a3/cheap.py, cheap.log, r4.py, /root/rot-hunt/R-A3-4.done.

## Earlier (7 Oct 16:00 Bangkok): idea #2 "trenches lead the market" KILLED (ETA was 18:00, why early: reused the exit grid book sim)
- **One line: cutting the H1-ALT book to half when small caps turn down near a BTC high lowers return AND does not lower the worst drop. Random timing of the same cut does better. Killed.**
- Pre-registered first: 2 rows in /opt/rot/coder/tries.csv (family A3.TRENCH, md5 9d2778f30cc43aad48eaef089e7a73dc), results added (4 era rows).
- Rule tested: small cap basket = all liquid coins ($500k+ volume) with cap $5M to $100M, equal weight (about 440 coins a day). Signal ON at the close where its 20 day return turns below 0 (was above 0 in the last 10 days) while BTC is within 5% of its 30 day high. Book cut to 50% next day, back to 100% when the 20 day return is above 0. 1% cost on the switched half. Book = frozen H1-ALT (trail 25, cut -10, cap 90). Control c1 = same without the BTC condition. Luck check = 200 random time shifts of the same ON days.

| | 2019-22 return | worst drop | 2023-26 return | worst drop | days ON | random shift beats it |
|---|---|---|---|---|---|---|
| frozen H1-ALT (no overlay) | +518% | -39% | +786% | -66% | | |
| v1 trench cut (BTC near high) | +432% | -39% | +466% | -71% | 13% / 35% | 90% / 99% of shifts |
| c1 same, no BTC condition | +348% | -38% | +340% | -74% | 47% / 57% | 77% / 100% of shifts |

- **Why it fails (first principles):** the small cap 20 day return flips sign about 7 times a year (50 triggers 2019-26), so the cut mostly lands in normal chop and halves the rebound. The big H1-ALT drops come from 10 small coins bought the same week falling together, fast; a 20 day basket turn is too slow to see that, and halving after the fall then halves the bounce. That is why the worst drop gets deeper in 2023-26 (-71% vs -66%).
- Fits the earlier Cowork test (weekly lead lag about 0; memes topped only about 10 days early at the big tops): a tiny lead on a few tops does not pay for 7 false alarms a year.
- Live now (for the call, not a signal): small cap basket 20 day return +17.6%, BTC within 5% of its 30 day high, flag OFF.
- Next for drop control (my suggestion, not started): the drop is an entry clustering problem, so the cheap test is a cap on NEW entries per week (for example max 3 new coins a week) on the frozen book. Needs supervisor OK.
- Files: rot-box /opt/rot/a3/trench.py, trench.log, trench/trench.csv.

## Earlier (7 Oct 15:55 Bangkok): EXIT GRID on H1-ALT DONE (ETA was 18:30, why early: 18 cells ran in 6 min on cached prices)
- **Answer in one line: keep the frozen exits (trail 25%, cut -10%, time cap 90 days). No cell beats them in BOTH eras (0 of 17). The cut is the only exit knob that matters.**
- Setup: entries frozen = Auditor H1-ALT build (same trades: 276 in 2019-22, 359 in 2023-26, holdout out). Only exits change. Book = 10% slots, max 10, 1% per side, marked daily. Same book sim for every cell, so cells compare fairly (absolute level is not the engine number). 200 random books per cell (random liquid coins, same days, same exits).
- Pre-registered first: 18 rows in /opt/rot/coder/tries.csv (family A3.EXIT, PREREG, md5 7806097135c7cb0343c29b3873bbc816), results added as 36 era rows.

| exit (trail / cap / cut) | 2019-22 book | worst drop | 2023-26 book | worst drop | PF 19-22 / 23-26 |
|---|---|---|---|---|---|
| **25% / 90d / -10% (frozen)** | **+518%** | -39% | **+786%** | -66% | 2.64 / 1.78 |
| 35% / 90d / -10% (best 2019-22) | +2,323% | -41% | +241% | -74% | 2.67 / 1.78 |
| 15% / 90d / -10% | +1,296% | -60% | +400% | -63% | 2.34 / 1.62 |
| 25% / 60d / -10% | +518% | -39% | +503% | -66% | 2.64 / 1.57 |
| 25% / 90d / -20% | +318% | -42% | +247% | -54% | 2.19 / 1.70 |
| 35% / 90d / -20% (worst 2023-26) | +445% | -53% | -32% | -76% | 2.18 / 1.54 |
| hold BTC (holdout days out) | +27% | | +275% | | |

- **What we learned (first principles):**
  1. **Cut -10% beats cut -20% in 18 of 18 pairs.** Mechanism: a +35% week in a sub $20M coin that then falls 10% is mostly a failed pump; waiting to -20% just pays more for the same exit.
  2. **Trail width flips by era.** Wide 35% wins 2019-22 (long alt season trends), loses 2023-26 (short pumps that round trip). 25% is the only width that is good in both. Do not tune it: that is fitting the era.
  3. **Time cap barely matters** (90d beats 30d in 5 of 12, 60d in 9 of 12). Average hold is 7 to 21 days anyway.
  4. **Exits do NOT fix the drop.** Worst drop is -29% to -60% in 2019-22 and **-50% to -77% in 2023-26 in every cell.** The drop comes from the entries clustering (10 small coins bought in the same week fall together), not from the exit. Drop control has to come from sizing or a book level exit (that is idea #2 below).
  5. Luck check: the real picks beat 96% to 100% of random books in all 18 cells in 2019-22 and 16 of 18 in 2023-26 (the two wide trail + loose cut cells fall to 63% and 28%). Random picks with ANY exit lose to BTC in 2023-26 (random median -1% to -40% vs BTC +275%). Exits alone are not an edge; the entries are.
- Verdict: A3.EXIT family DONE, no challenger (frozen cell already the cross era best). Nothing goes to the Auditor from this grid.
- Idea #2 done, see top.
- Files: rot-box /opt/rot/a3/exit_grid.py, exit_grid.log, exitgrid/grid.csv.

## Earlier: D-024 wrong pond (7 Oct 14:50 Bangkok): D-024 "wrong pond" ANSWERED (ETA was 18:00, why early: existing data was enough)
- **Correction first: the 75% is NOT outside our CMC list.** My split's bucket A1 = movers inside CMC top 1,500 that basket v3's THEME baskets never covered. Every one of the 1,660 legs came from the CMC files. So the full CMC scan (team rule since 12:34, H1 / W.R1 already do it) covers that 75%. DEX data covers 0% of the current oracle gap.
- **What DEX-only adds is a NEW leg, before CMC.** Winners' list (winners_outside_start.csv, 2023-26): 110 of 351 5x+ runs started below CMC top 1,500; 63 were brand new CMC listings (Solana 28, BNB 12, ETH 10, Base 4).
  - When ROT can first see them (on CMC, median 8 days after listing) they are already worth median $13.8M (p25 $5.8M; 0 of 63 under $1M, 11 at $1M to $5M, 26 at $5M to $20M, 26 above $20M), and still run a median 7.1x to a $130M peak.
  - The leg before that is the DEX launch leg: from a pump.fun style start (thousands of dollars) or a graduation (about $100k) to about $10M, i.e. 100x or more in days. That is the launch lane our sister teams (FW, SW) already trade. Not ROT's lane (supervisor note 13:15 agrees).
- **Size of the DEX to CMC pipeline (rot.db, 2023 to now):** 2,905 coins entered CMC top 1,500 for the first time (2023 451, 2024 961, 2025 974, 2026 so far 519; about 80 a month in 2024-25). 1,272 of them (44%) were CMC listings 30 days old or less, i.e. DEX born. 706 memes. By chain: Solana 798, native L1s 644, BNB 599, ETH 429, Base 263.
- **Size of the never on CMC pond, 2023-26: NOT measurable for free.** GeckoTerminal free history stops at 180 days, DexScreener keeps no history, Dune costs credits (Ben's whole budget is 4,000 a month and must not burn). What exists: Robinhood Chain complete from 1 May 2026 (PP tokens.db, 1.69M tokens; e1 runners table ETA 22:00 tonight), Solana + RH coins over $500k from 2 Oct only (runtrack, copy tonight).
- **Verdict for the team:** do NOT widen ROT past CMC. Two cheap moves instead: (1) universe from day 1 on the CMC list, refreshed daily (Coder and Winners already doing this); (2) forward only test, no history needed: flag Solana + RH coins from runtrack that cross $5M with $1M+ daily volume as "pre CMC watch" and score how many become CMC 5x runs. Owner suggestion: Coder (runtrack copy lands tonight). Registered as idea A3.I5 below.
- My number (elite.md): v3 decay explained = done (universe, 87%; picks and timing 11%).

## Earlier today
### NEW TASK DONE: where basket v3 loses the top movers (CODER.BASE_V3 capture.csv, 1,660 mover legs, top 20 per month, $1M+ volume)
| bucket | 2019-22 | 2023-now | all | median move |
|---|---|---|---|---|
| A1 never in our universe: theme v3 does not cover | 748 (93.5%) | 495 (57.6%) | 1,243 (74.9%) | 3.1x |
| A2 right theme but too small (not top 15 by cap) | 24 (3.0%) | 181 (21.0%) | 205 (12.3%) | 3.8x |
| B in a v3 basket, never bought | 18 (2.2%) | 138 (16.0%) | 156 (9.4%) | 3.8x |
| C bought but late or sold early | 5 (0.6%) | 29 (3.4%) | 34 (2.0%) | 5.4x |
| D caught (held 50%+ of the leg) | 5 (0.6%) | 17 (2.0%) | 22 (1.3%) | 4.0x |
- Examples (5 each, biggest): A1: ELON 2021-10 45.6x, ACH 2021-08 36.8x, MICRO 2023-01 112x, RAVE 2026-04 93x, AB 2025-01 73x. A2: SAMO 2021-10 39.5x, ALPACA 2025-04 19.4x, MOONPIG 2025-05 15.5x, NOMNOM 2024-11 13.8x, DONKEY 2025-08 13.7x. B: COAI 2025-10 86x, LIGHT 2025-09 85x, APX 2025-09 33x, U 2026-07 33x, AKE 2026-07 24x. C: BOBO 2026-05, ACT 2024-11 50x, PNUT 2024-11 45x, FARTCOIN 2024-11 16x, TOSHI 2025-01 12.8x. D: VELVET 2026-06 20x, SOLAMA 2024-03 14.7x, LIKE 2023-11 10x, MOODENG 2025-05 8.4x, BONK 2023-12 6.1x.
- A1 misses by sector tag at leg start: untagged/other 677, L1 127, gaming 118, defi 89, meme 72, infra 55, AI 37, L2 35, exchange 16, privacy 10.
- VERDICT: the biggest loss is the UNIVERSE, not timing. 87% of top movers (A1 + A2) were never in any v3 basket. Picking and timing (B + C) is only 11%. Next team task: rules must scan ALL liquid coins (top 1,500, $1M+ volume), not chain x theme baskets. Sector baskets can stay as a filter, never as the universe.
- Answer key needs cleaning (for Coder): it contains junk legs, e.g. USDE 2023-05 268x and 2023-06 89x (pegged/new listing print), vBNB 2024-05 80x (wrapped), GME ticker clash. Run the movers list through the bad print filter and the pegged/wrapped check, then rerun. It will not change the verdict (A1 + A2 dominate both eras).
- Files: /opt/rot/a3/miss_buckets.py (rot-box), output miss_buckets.out, per mover miss_buckets.csv.

### Baseline gap (A3 +361% vs Coder +196%): EXPLAINED, agreed baseline = the time machine number
Same window 2023-01-01 to 2026-07-08, 1 day delay, Coder's base_v3 rule on the time machine (rot-box /opt/rot/a3/recon/*):
| version | total | worst drop |
|---|---|---|
| my old book on the old basket INDEX lines (bk_clean.json) | +351.4% | -37.7% |
| time machine, hold coins, rebalance on change, 0% cost | +195.2% | -42.0% |
| time machine, rebalance daily (= the index math in real coins), 0% cost | +178.0% | -41.1% |
| time machine, rebalance daily, 1% cost | -12.6% | -70.2% |
| hold BTC | +273.1% | -53.1% |
- NOT the window (same window still +351%). NOT cost on the on_change book (cost 0 vs Coder's 1% run: +195% vs +196%). NOT daily rebalancing: doing the index math in real coins gives LESS (+178%), so the old line's daily equal weight is not a hidden edge.
- So the whole ~155 point gap sits in the OLD BASKET LINES themselves: (1) different members (token platform field vs ecosystem tags), (2) line artifacts no one can trade: daily returns clipped at -90%/+200%, a missing coin skipped when it reappears within 3 days (look ahead), -50% for missing coins. Splitting (1) from (2) needs the old snapB snapshots; not worth it, because neither is tradable as is.
- AGREED BASELINE: the time machine (Coder) number. Basket v3 2023 to now about +195% vs hold BTC +273%: v3 LOSES to BTC by about 80 points, worst drop -42% vs -53%. My +361% is retired; A3.B2c receipt points here. Rerun once the bad print filter ships (Coder).
- Side fact: daily rebalancing at 1% cost turns +178% into -13%. Any rule that trades daily must count cost.

### Earlier (still valid)
- Task 1: v3 trades reproduce exactly (118, PF 2.65, Sep 2024 on 1.81, 2025 1.06). On my old index data v3 +361% vs hold BTC +400% (2023 to Sep 2026).
- Task 3: why v3 weakened: A3.I1, I2, I4 KILLED; A3.I3 impulse filter KILLED by the Auditor (luck 84.5% < 95%, wrong sign in Jan 2023 to Aug 2024).
- Registry: A3.B1c, A3.B2c TESTING; A3.I1 to I4 KILLED (I3 by Auditor, move pending in registry).

## What already exists
Format: path | what it is | status | receipt

### Basket rotation engine (main 159.223.88.168)
- /opt/rotation/chains/chains.py (live md5 b4dad467283c2ca931754834d46d966e, 40+ .bak-v* versions) | the live basket engine behind rotation.html: baskets, words BUY / HOLD / TAKE % PROFIT / SELL, rotation card NOW/NEXT/BEFORE, picks, Telegram (muted by /opt/rotation/MUTE_TRADING) | LIVE | http://159.223.88.168/rotation.html
- /opt/rotation/chains/rule3.py (md5 cd22400c28efaa51536ad7ba2bbc8886) | basket rule v3: close > prior 30d high close, BTC > 100d SMA, majors basket on; half at first HOT; sell < prior 15d low; rotate into 3 strongest 30d | BASELINE (A3.B2) | 14 point in time CMC baskets Jan 2023 to Sep 2026, 1% cost: 118 trades, avg +8.1%, median +3.5%, 58% wins, PF 2.65; Sep 2024 on PF 1.81 (n77); 2025 PF 1.06; weekly clustered PF 2.92 | raw http://159.223.88.168/hx-c9ois/edge/basket-rule-v3-2026-10-02/
- /opt/rotation/chains/rule4.py (md5 on disk 3e16ccdc01dced7820a075bdd420fb9f) | rule v4.x: v3 + young baskets 10d high / 5d low, breadth gate 35%, hot market 20d high, 20% trailing stop after +30%, BUY (BOTTOM), meme mania DE-RISK gate | LIVE, but each v4.x was tuned on the same 2023-26 data: treat as TESTING, not PROVEN | raw per version http://159.223.88.168/hx-c9ois/edge/basket-rule-v41-2026-10-03/ to v47
- /opt/rotation/chains/baskets_def.py, fetch_top.py, candles.py, closes.py, picks.py, stages.py, fairclock.py, stale.py | basket definitions (CMC sector lists per chain, GeckoTerminal for young chains), CMC top snapshots, daily/weekly/monthly closes, coin picks v2 to v6, staleness | LIVE
- /opt/rotation/chains/state.json, closes_hist.json, seed2.json | basket histories and live ledgers (ledger4, scorecard4, picks caches) | DATA
- /opt/rotation/rotation.py + rot.db (163 KB) + rotation.json | trench/launchpad volume watcher (BTC 7d +10% rule) | old BUY TRENCHES signal KILLED (fired 61% of days, memes after it = any day); volume part kept as info | memory rotation-watch, 2 Oct
- /opt/rotation/rotmap.py (frozen v2) | CASH/BTC/ALTS Rotation Map book | FROZEN baseline per Ben, do not tune
- /opt/rotation/rh_drill.py + rh_flow_hist.json | Robinhood basket drill-down and flow history | LIVE (info)
- /opt/rotation/rotfix.py (10 min timer + cron), verify_rot.py (hourly, 49 checks) | self-heal and self-check | LIVE
- /opt/rotation/chains/rule2.py, chains_v1.py | older rules | RETIRED

### Tested results to NOT redo (edge hub, 4 to 5 Oct, all on the basket book, both eras 2019-22 and 2023-26)
- Rank by strongest 30d move first: beats random order 19/20 (2023-26) and 15/20 (2019-22) | PROVEN, live
- Basket construction: 15 biggest coins, equal weight, $25k volume, weekly refresh is best; 5/10 coins far worse; size weights worse; new-coin wait neutral | PROVEN
- 3 baskets at a time best (+2256% vs 1 slot +1674%), worst drop -24% | PROVEN
- Buy window: buy on the BUY close; 1 day late roughly halves the book (+1246% vs +2256%) | PROVEN
- Idle money: cash vs BTC vs BTC-rule; BTC only while its own rule is on is the best balance | PROVEN (descriptive)
- Jumps per bull: ~1 to 1.4 buys a month, median hold 11 to 12 days | PROVEN (descriptive)
- No alt basket risk on 58% of days 2023-26 (2025: 61%, 2026: 71%) vs 42% 2019-22 | PROVEN (descriptive), key input for Task 3
- 50 alt ranking methods, rejoin running uptrends, weekly RS swaps, RRG quadrants, depth buy zones, MA/momentum states, lone breakouts, BTC gate alone, stops on baskets, price-model for 2x | KILLED
- Live book replay 19 Aug to 4 Oct (11 trades): coin picks +35% vs basket line +2% vs BTC +25% | PROVEN (short window)
- Equal weight baskets bleed (about -7% a month median) and RH coin basket bled ~-4% per 24h 29 Aug to 2 Oct | PROVEN (this cycle)

### Chain regime brain (main, this session)
- /opt/chainbrain/chainbrain.py, gauge.py (v7, md5 ae5042f6fe3de34e43beb45379f3b73d), chain.json, gauge.json, ledger.db | daily 08:47 BKK chain phases for 12 chains (FALLING, BASING, TURNING UP, CLIMBING, HOT, TOPPING, COOLING), meme phases, cycle clock, catalyst watch, fees per wallet, bridge ETH | LIVE
- Cycle clock: event = fees 7d at 25% of 60d peak; price low median 62 days later (p25 44, p75 115), median further drop -40%; 14 cases, 12 chains + pump.fun | TESTING (small n)
- DEEP BLEEDING (60%+ under 90d high, under 20d avg, fees 7d < 0.8x 28d) = more downside 52% vs 41% normal; Solana 84% | PROVEN (fixed membership)
- New chain survival: 6 of 19 chains regained half their launch peak after an -80% fee crash | PROVEN (descriptive)
- RH correlation with BTC/ETH/SOL only 0.2 to 0.3 | FACT
- Catalyst: new app >=3% fee share that had <=1% 28d earlier (Pons V2 fired 8 days before the 18 Aug RH low) | TESTING (n small)
- Summed CMC caps and equal-weight log index gave fake signals (said -97%, real -65%) | KILLED; use fixed membership or CMC cap-weighted basket

### Other related
- Macro light (3 Oct): 29 X callers scored, 30+ mechanical signals, 5-state weekly light | memory macro-direction | TESTING
- Price Predict data: read only copies on research-box /opt/ppcopy/ (chainday.db bridge + stables, prices.db, rot.db 1,011 days pad share)

## Task status
- Task 0 handover: DONE.
- Task 1 baseline as code: DONE for 2023 to now. Files: research-box /opt/rot/a3/rules/A3_B1.py, A3_B2.py (+ rule3.py, stages.py), own loop a3loop.py (1 day delay switch, 0.5% per side), results /opt/rot/a3/rules/out/*.json. Gap vs old: old book note said +353%, mine +361% (same rule, slot weights at half size after HOT); within 8 points. 2020-22: blocked until old-era basket lines exist (A1 universe).
- Task 2 rotation clock Oct 2023 to now: NEXT.
- Task 3 why v3 weakened: first pass DONE (above). Script /opt/rot/a3/why_v3.py, output why_v3.out.
- Caveat: a 1 day delay did BETTER than same day on this book (+406% vs +361%). The edge hub line "1 day late halves the book" was a different book (index, 3 slots, 2023-26 +2256% vs +1246%). Worth a recheck by the Auditor.

## Asks for Supervisor
- None blocking. Coder: please run CODER.BASE_V3 and post its 2023 to now book number next to my +361.3% so we have one agreed baseline.

## New ideas or kills this wake
- 3 kills (A3.I1, I2, I4), 1 testing (A3.I3 impulse filter).

Version: v5, 7 Oct 2026 14:50 Bangkok. Latest: http://159.223.88.168/hx-c9ois/rot/a3-status.md