# 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_01Noaeh8ADLgEtxy5cLQZWGH (armed, 07 Oct 16:29 Bangkok) ROT Auditor. Nudge per http://159.223.88.168/hx-c9ois/wake-howto.md
LIVE PLAN MD5 CHECK (15:49 task): PASS. rot-live-plan.md v1 and rot-live-orders.json v1 match my own run of frozen SZ.H1ALT (md5 8b74531af3f75e4f56804525cece9526): same 4 coins, same entry dates, same stops, weights within 0.6% of mine.

# ROT Auditor status (v13, 7 Oct 2026, 16:34 Bangkok)

## Read this first (v13, live plan check done 16:34, ETA was 17:30)
- **How checked:** my own engine run on rot-box (frozen rule md5 8b74531af3f75e4f56804525cece9526 FROZEN MATCH, engine md5 01054c4fc6a502c8809a1816aea97eb4, cache 63316cf964 to 6 Oct, live_ops so the last 90 days are not masked, daily rebalance, drift 0.05, 1% cost). Script /opt/rot/auditor/livecheck.py.

| Coin | Orders file weight | My run target | Entry date (both) | CUT and TRAIL arithmetic |
|---|---|---|---|---|
| SHARE | 15.35% | 15.44% | 2026-09-06 | ok (0.435057 x 0.90, 0.720681 x 0.75) |
| BREW | 13.06% | 13.14% | 2026-09-30 | ok |
| FOLD | 11.81% | 11.89% | 2026-09-23 | ok |
| DOGO | 7.68% | 7.73% | 2026-09-26 | ok |
| MWXT | exit, no buy | not held | n/a | n/a |

- **Weights:** every orders weight is my weight x 0.994 (same factor for all four). That is a book-value base difference, not a different rule. Effect: about $0.09 per slot. Not a blocker.
- **Contracts checked against DEX data:** BREW `0xfa6D9B504848606Eb9aeC04CCc161D169B3f2159` (BSC, PancakeSwap, DexScreener: BREW) ok. FOLD `0xe172e9b6cfbeeb5593bdce3f077356fdb33af904` (Ethereum, Uniswap, DexScreener: Interfold) ok. SHARE `0x5FCA51aff213bFBEAB0b711b93c3374252fD6aC3` (BSC): DexScreener has no pair (the PancakeSwap Infinity pool is not indexed there), GeckoTerminal confirms "ShareX Token" on BSC, ok. DOGO not checked (paper only per 16:11).
- **Three things the team should know before the first buy:**
  1. **Live prices already ran past the sizing close.** At 16:31 Bangkok: SHARE $0.781 vs 6 Oct close $0.639 (+22%), BREW $0.0139 vs $0.0108 (+29%), FOLD $0.0655 vs $0.0700 (-6%). Live buys pay that gap; the fill log will show it as haircut. The stops stay on the PAPER entry price (mirror rule), so for a live SHARE bought near $0.78 the nearest stop is the TRAIL at $0.54 (31% under the live buy), BREW near $0.0139 has its TRAIL at $0.0080 (42% under). Say this to Ben: the -10% cut does not protect the live dollars on day one.
  2. **Zombie slot in the engine book:** APTM (CMC 36215), entered 2025-07-27, still held at 0.2% at -97%, 437 days past the 90 day cap (dead coin with no price, never sold). The orders file correctly ignores it, but the engine's "dead" write-off did not remove it. Coder: small fix (dead coins past maxhold must leave the book).
  3. DOGO: the rule needs $500k volume on the signal day; CMC now shows $17 a day. Paper only is right.

## v12 block (15:52, 6 month windows)
H1-ALT 6 MONTH WINDOWS (15:44 task): beats hold BTC in 8 of 12 clean windows and the random book median in 9 of 12. But since 2023 ONE window (Jan to Jun 2024) carries it: without it the book is +67% vs BTC +133%. Fragility FAIL on concentration (66% of the 2023-26 gain from one half year, bar 40%). Verdict stays FIX: PAPER only.

## Read this first (v12, 15:44 task done 15:51, ETA was 17:15)
- **Frozen code, rerun by me on the Coder's engine:** SZ.H1ALT md5 8b74531af3f75e4f56804525cece9526, engine md5 01054c4fc6a502c8809a1816aea97eb4, rebalance daily, drift_tol 0.05, 1 day delay, 1% each way. It reproduces the Sizer exactly: 2020-22 +846.3%, 2023 to now +373.4% (out /opt/rot/auditor/runs/h1alt_paper_cell). Random book median from my clean build (same 1,000 draws as v11: random liquid coin on each H1-ALT entry day, same exits, 10% slots). Script /opt/rot/auditor/audit_windows.py md5 a0ad2340138553f7ddd842117660ba7e, table /opt/rot/auditor/out/h1alt_windows.csv. Logged in tries.csv (SZ_H1ALT, WINDOWS6M_AUDITOR).

| Half year | Engine book | Hold BTC | Random median | Beats BTC | Beats random |
|---|---|---|---|---|---|
| 2019 H2 | -18% | -32% | -24% | yes | yes |
| 2020 H1 | +272% | +27% | +52% | yes | yes |
| 2020 H2 | holdout (Oct to Dec locked) | | | skip | skip |
| 2021 H1 | holdout (Jan to May locked) | | | skip | skip |
| 2021 H2 | +96% | +38% | +36% | yes | yes |
| 2022 H1 | 0% | -59% | -18% | yes | yes |
| 2022 H2 | -20% | -14% | -15% | **no** | **no** |
| 2023 H1 | +6% | +83% | +1% | **no** | yes |
| 2023 H2 | +74% | +38% | +43% | yes | yes |
| 2024 H1 | +168% | +42% | -4% | yes | yes |
| 2024 H2 | -23% | +49% | +5% | **no** | **no** |
| 2025 H1 | -17% | +14% | -11% | **no** | **no** |
| 2025 H2 | +65% | -17% | -18% | yes | yes |
| 2026 H1 | -15% | -34% | -24% | yes | yes |

- **Count (12 clean windows):** beats hold BTC **8/12**, beats random median **9/12**. My own build: 9/12 and 9/12. Two builds agree.
- **The fragile part:** 2023 to Jun 2026 compounded = engine +347% vs BTC +231%. Drop the single best half year (2024 H1) and it is **+67% vs BTC +133%**: it LOSES. That one window is 66% of the 2023-26 gain (log terms), over my 40% concentration bar. Same story in 2019-22: 2020 H1 is 84% of the gain.
- **Losing streak to watch:** 2024 H2 and 2025 H1 lost to BTC AND to random back to back (12 months). A live paper book could sit through a year like that.
- **Plain read:** it wins most half years, but the big lead over BTC comes from one or two monster half years per cycle (alt season bursts). It is a burst catcher, not a steady beater. Verdict unchanged: FIX, PAPER only, not CHAMPION.
- **What would change it:** a version whose 2023-26 lead over BTC survives removing its best half year. Candidate for A3's exit grid: hold winners longer in burst windows (trail 35%, 90 day cap), test if the lead stops depending on 2024 H1.

## v11 block (14:41, season check)
SEASON CHECK (14:35 task 1): NO. Random coins on the same gate days with the same exits do NOT beat hold BTC. Median random book 2023-26 -9% vs BTC +274% (1% of random books beat BTC); 2019-22 +22% vs BTC +27% (holdout days removed). Season + exits alone is not the candidate. In 2019-26 the PICKS matter: the real H1-ALT book beats 99% of random books in both eras.

## Read this first (v11, 14:35 task 1 done)
- **Push back on the 14:35 lesson "our picks are no better than a dart".** That is true in 2017-18 only (67% vs random). In 2019-22 and 2023-26 the picks beat random coins in 99% of books. The season and exits alone LOSE to hold BTC. So the SEASON family as written (equal weight every liquid young small coin on gate days) will most likely look like the random book below: flat to negative since 2023. Sizer can still run it (cheap), but expect a kill.
- **Method:** same 1,000 random draws as my H1-ALT Gate 4 (seed 11, first use, so identical draws): on every day H1-ALT entered, one random liquid coin (CMC top list, $500k volume, 20 days of history), same exits (cut -10%, trail 25%, 90 day cap), 1 day delay, 1% each way. Each draw turned into a book like SZ.H1ALT: 10% of equity per trade, max 10 open. Trades touching the holdout are dropped (both eras), so BTC is also shown with holdout days removed. Script /opt/rot/auditor/audit_season.py (rot-box) md5 300d4680ecc52f7776abf7be3c45cfea = my clean H1-ALT build (b855fb55...) up to the trade list + the book tail. Draws saved: /opt/rot/auditor/out/season_random_books.csv.

| Era (tradable days) | Real H1-ALT book | Random book median (p10 to p90) | Random avg trade (median) | Hold BTC | Random books above BTC | H1-ALT beats random books |
|---|---|---|---|---|---|---|
| 2019-22 (holdout Oct 2020 to May 2021 removed) | +402% | **+22%** (-26% to +144%) | +5.5% | +27% (full span +331%) | 47% | 99% |
| 2023 to 9 Jul 2026 | +437% | **-9%** (-47% to +89%) | -1.1% | +274% | **1%** | 99% |

- **What it means, plain:** being in small alts on alt season days with good exits is about a coin flip against BTC in 2019-22 and a clear loss since 2023. The H1-ALT entry filter (7 day jump of +35%, cap under $20M, young) is what made the money in 2019-26. It did NOT help in 2017-18 (one shot), so the edge is real in the recent cycle but not proven across cycles.
- **Remaining risk for H1-ALT (unchanged FIX, PAPER only):** Deflated Sharpe on the engine book 0.69 (2023-26, N=142) and 0.91 (2020-22), both under 0.95; one shot 2017-18 failed the random test; only the centre setting beats BTC. Forward paper calls decide.
- My H1-ALT book here (+437% since 2023) is close to Sizer's engine book (+373%): two builds agree.

## v10 block (14:32, one shot)
ONE SHOT 2017-18: FAIL (H1-ALT stays PAPER, never CHAMPION on this evidence). Killing number: random coins on the same days with the same exits beat it in 32.7% of 1,000 draws (bar: under 5%). Book +349% vs hold BTC +275% (passes), avg trade +15.5% PF 2.2 (passes). 2 of 3 parts pass, the bar needed all 3.

## Read this first (v10, one shot done 14:31, Coder lock landed 14:25)
- **Moved to rot-box (168.144.252.100 (rot-box2)).** One shot run there, output only in /opt/rot/auditor/oneshot/.
- **What ran, once each, same sitting (14:30 Bangkok):**
  - Sizer's frozen SZ.H1ALT (md5 8b74531af3f75e4f56804525cece9526, verified) on the Coder's engine (md5 01054c4fc6a502c8809a1816aea97eb4, imported unchanged). Centre params, rebalance daily, drift_tol 0.05, 1 day delay, 1% each way. 2017-18 cache built by an exact copy of build_cache.py with only RAW, OUT and FLOOR changed (cache a3e60b6794: 730 days, 2,636 coins, 1,013 dead, 2,477 bad prints dropped). No holdout mask inside 2017-18 (the whole cycle IS the holdout).
  - My independent rebuild audit_h1alt_1718.py (md5 ce5af8529379744e7e4e5098da813f54, unchanged; a wrapper only swapped the folder path because the files moved to /opt/rot/locked/cmc_2017_2018/).
  - Both opens logged in /opt/rot/db/holdout_log.csv (07:30:25 and 07:31:38 UTC). Row added to /opt/rot/coder/tries.csv (family SZ_H1ALT, variant ONESHOT_1718, era E0_2017_18). Coder: keep that row if tries_collect.py rewrites the file.

| Bar part (set at 13:57, before any run) | Engine (frozen SZ.H1ALT) | My rebuild | Result |
|---|---|---|---|
| 1. Avg trade above 0 after cost, PF 1.3+ | 94 trades, +15.5%, PF 2.20 | 109 trades, +17.6%, PF 2.30 | PASS |
| 2. Beats random coins, same days and exits, 95%+ of 1,000 draws | n/a | **67.3%** | **FAIL** |
| 3. Book above hold BTC, same dates | **+349.3%** (worst drop -38.7%) vs BTC +274.9% (worst drop -83.4%) | n/a | PASS |

- **Other gates (information, not in the bar):** Deflated Sharpe 0.18 (N=340) FAIL. Fake search 5/20 FAIL. Fragility: all 17 moves positive, PF 1.84 to 4.08.
- **Why it fails, in one line:** in 2017-18 the money came from the season (small young coins in an alt season), not from the pick. Random liquid coins bought on the same days with the same exits did almost as well.
  - Plain H1 with NO alt gate did better here: +29.9% per trade, PF 3.20, 208 trades. The gate cut good trades.
  - Without its 5 best trades: engine -1.9% per trade, my rebuild +2.7%. The median trade loses 9% to 11%. It beats BTC on the same days in only 37% of trades.
  - The book only started 23 Nov 2017: the gate needs 180 days of history, so it was in cash for the first 11 months. The +349% is mostly Dec 2017 to Jun 2018 ($14.7k at end 2017, $43.0k at mid 2018, $44.9k at end 2018).
- **Steelman:** it made money in a clean cycle nobody fitted it to, beat hold BTC, and its worst drop (-39%) was half of BTC's (-83%). The exits (cut, trail, 90 day cap) protect well in a crash. That is real, but it is the EXITS plus the SEASON, not proof the picks beat a dart.
- **Deflated Sharpe on the true count (14:20 task 3):** tries.csv holds 142 unique H1 family variants (SZ_H1 123 rows of variants, SZ_H1ALT 18, SZ_H1ALTB 18; Sizer runs outside the time machine still missing). On the ENGINE book of the paper cell (v013, 2019-now run): 2023-26 DSR **0.69** at N=142 (0.63 at N=340), 2020-22 DSR **0.91** at N=142. Both under the 0.95 bar. My own per-trade build had 0.90 / 0.999; the engine book is weaker.
- **What changes the verdict:** 20+ scored forward paper calls where the book beats hold BTC AND random coins with the same exits. A new picker that beats random coins in 2017-18 would need its own registration and cannot use this cycle again (it is now burned).
- **2017-18 is now BURNED.** Opened once, logged. No role may tune on it.

## v9 block (13:57)
- Doorbell: trig_01SMPstqrupkTnWwwyTqBnpT (Doorbell: ROT Auditor)
- **CORRECTION: my H1 (12:42) and H1-ALT (13:35) runs were contaminated.** A1's 2017-18 files landed in the same CMC folder at 12:17 to 12:23 Bangkok. My scripts read every file in that folder, and a 2017-18 date gave a negative day number, so those prices were written into the LAST two years (Oct 2024 to Oct 2026) of each coin. Receipt: 7,880 coins and 1,316 bad prints in the bad runs, 7,016 and 497 in the clean runs. No 2017-18 result was ever computed or printed, so the clean cycle is NOT burned. But junk prices entered the 2023-26 numbers, mostly the random-coin pools.
- **Corrected numbers (scripts now read 2019 onward only):**
  - **H1:** unchanged verdict, KILLED for 2023-26. Random coins on the same days beat it 9.3% of draws (bar 5%), fake search 12/20 (bar 20/20), Deflated Sharpe 0.51.
  - **H1-ALT, 2023-26 is now stronger:** +10.1% per trade, PF 1.78. Random coins **100% PASS**, fake search **20/20 PASS** (was 18/20), Deflated Sharpe **0.90 FAIL** (bar 0.95). 2019-22: random 99.1% PASS, fake search 16/20 FAIL, Deflated Sharpe 0.999 PASS.
  - Verdict stays **FIX (PAPER yes, CHAMPION no)**, but it is now one gate short in each era. The 2017-18 one shot decides it.
- **ONE SHOT on 2017-18, set BEFORE any run (nobody has looked):**
  - My independent rebuild: /opt/rot/auditor/audit_h1alt_1718.py, md5 ce5af8529379744e7e4e5098da813f54 (same frozen gate: share of liquid coins beating BTC over 30 days above its own past 1 year median; W.R2 picks; cut -10%; trail 25%; 90 day cap; 1 day delay; 1% each way). It reads only 2017-01-01 to 2018-12-31.
  - Plus Sizer's frozen SZ.H1ALT centre (md5 8b74531af3f75e4f56804525cece9526) on the Coder's engine, same sitting.
  - **Pass bar (all three):** (1) average trade above 0 after cost and PF 1.3 or more; (2) beats random coins on the same days with the same exits in 95% or more of 1,000 draws; (3) the SZ.H1ALT book ends the window above hold BTC over the same dates. Any miss: H1-ALT stays a PAPER test only.
- **Blocker for Coder (needed for part 3):** I cannot reach rot-box (10.104.0.5) from my side. Either run the frozen SZ.H1ALT centre on 2017-18 with the Auditor flag and write the output only to research-box /opt/rot/auditor/oneshot/ (do not print it), or give the Auditor a key to rot-box. I run my part in the same sitting when yours is ready, and log both opens in /opt/rot/db/holdout_log.csv.
- **For every role:** anything that globs /opt/rot/analyst/raw/cmc/ now reads 2017-18 too. Filter by date (2019 onward) or you have the same bug and you also break the 2017-18 lock.

## H1-ALT first verdict (v7 13:35, v8 13:41; numbers before the correction above)
- **H1-ALT early verdict (due 17:00, done 13:35): NOT YET. Best rule the team has, but it misses 2 of my 3 luck gates in 2023-26.** Run on the gate exactly as frozen in the bus at 13:15 (share of liquid CMC coins beating BTC over 30 days above its own past 1 year median), before Sizer's run, so this is an independent check.
- **What the gate does:** 2023-26 goes from +3.7% to **+9.3% per trade, PF 1.27 to 1.72**. Gate is ON about half the days in both eras (46% / 52%).
- **Gates, 2023-26:** random coins on the same days with the same exits beat it only 0.4% of the time (**PASS**, 99.6%). But a fake search of 160 exit combos on random picks beats it in 2 of 20 runs (**FAIL**, need 0 of 20), and Deflated Sharpe is 0.87 (**FAIL**, bar 0.95). 2019-22: random 98.6% PASS, fake search 15 of 20 FAIL, Deflated Sharpe 0.999 PASS.
- **Robust to settings:** all 17 nearby settings stay positive in both eras (2023-26 PF 1.44 to 1.89). Gate window 292 / 438 days and quantile 40 / 60% all hold. Not curve fit.
- **The weak spot:** without its 5 best trades it is -1.8% per trade in 2023-26 and +2.7% in 2019-22. The median trade loses 13%. Two of the top 5 are the same week (22 to 23 Jul 2025). It is a lottery-ticket rule: many small losses, few big wins.
- **Verdict: FIX, not FAIL.** Put it on PAPER now (forward days are a clean test). It cannot be CHAMPION yet. Two ways to pass: (1) A1's 2017 to 2018 history as a clean test (the gate was never fit there), or (2) 20+ scored paper calls. I will re-audit on either.
- **13:41 match check with Sizer's book (sizer-h1alt.md): same answer, verdict stays FIX.** Book level 2023 to 8 Jul 2026: centre +373% vs hold BTC +275%, worst drop -65%. But only the centre beats BTC: its nearest neighbours (gate 0.9: +138%, gate 1.1: +261%, history 7: +237%) all LOSE to hold BTC. Beating BTC rests on one exact setting, so on the money bar that matters most it is fragile. Worst drop -65% also misses the -60% bar.
- Registry: SZ.H1ALT registered 13:32 Bangkok, md5 8b74531af3f75e4f56804525cece9526, 18 variants counted. PASS on check 8 (ledger time and file match the bus spec).
- Clock: Sizer's report says 13:50, the file landed at 13:40 on the bus. Times in reports still run ahead of the real clock.

## H1-ALT receipts (CONTAMINATED run, see correction at top) (/opt/rot/auditor/audit_h1alt.py md5 6c76d2a57f175b62adb0c9800c5e2107, log h1alt.log, trades out/h1alt_trades.csv)
| Check | 2019-22 | 2023-26 |
|---|---|---|
| trades, avg, median, PF | 275, +18.2%, -10.2%, 2.65 | 371, +9.3%, -13.1%, 1.72 |
| plain H1 (no alt gate), same build | 463, +14.7%, PF 2.24 | 631, +3.8%, PF 1.29 |
| without the 5 best trades | +2.7% | -1.8% |
| beats BTC on the same days | 32% of trades (excess +17.7%) | 25% of trades (excess +8.3%) |
| Gate 2 Deflated Sharpe (N=340) | 0.999 PASS | 0.867 FAIL |
| Gate 3 fake search, 20 runs | 15/20 FAIL | 18/20 FAIL |
| Gate 4 random coins, same days and exits | 98.6% PASS | 99.6% PASS |
| Fragility (17 moves) | all positive, PF 1.90 to 2.94 | all positive, PF 1.44 to 1.89 |
| Profit by year (sum of trades) | 2019 -1.2, 2020 +35.4, 2021 +16.2, 2022 -0.3 | 2023 +11.0, 2024 +16.2, 2025 +9.5, 2026 -2.2 |
- History gate 7 days (W.R1b) instead of 20: 2019-22 +20.9% PF 2.79, 2023-26 +7.5% PF 1.56. 20 days is the better setting in 2023-26.

## H1 verdict (v5 12:42, v6 13:16)
- **H1 first verdict (deadline 16:00, done 12:42): FAIL in 2023-26, PASS in 2019-22.** H1 = W.R2 picks + cut -10% + trail 25%. Rebuilt independently on research-box from A1's CMC files with my own bad print filter (1,316 removed) and the age fix. No holdout used.
- **Killing number (2023-26):** random liquid coins bought on the SAME days with the SAME exits beat H1 in 18% of 1,000 draws (bar 5%). Let the random picks choose their best of the same 160 exit combos Sizer searched, and they beat H1 in 10 of 20 runs (bar 0 of 20). Deflated Sharpe 0.51 (bar 0.95). In 2023-26 H1 is luck.
- **2019-22 is real:** beats 99.3% of random draws, beats the best fake search in 19 of 20, Deflated Sharpe 0.999. But most of the profit is 2020 and 2021 (Jun to Sep 2020 and the 2021 runs).
- **Money, 2023-26:** +3.7% per trade, PF 1.27, median -13%, beats BTC on only 23% of trades, and -1.6% per trade without the 5 best trades. 2025: -11.4 total, 2026 so far: -6.1. It is getting worse, not better.
- **What this means:** H1 is an alt-season rule. It works when small coins are in a broad bull (2020, 2021) and is a coin flip outside it. Next test worth doing (register first): H1 only when an alt-season gate is on (for example: share of top 1,500 coins beating BTC over 30 days above its 1 year median). And A3 found 75% of the movers are outside our CMC universe: the 2023-26 runners live on DEX, so H1 on the CMC list may be fishing in the wrong pond.
- **13:16 match check: Sizer agrees, H1 FAILS.** Sizer's time machine (sizer-h1.md): 2023 to now -9.5%, worst drop -93%, 13 of 27 nearby settings negative. My trade-level rebuild says the same. Two independent builds, one answer: KILLED for 2023-26.
- **Registry: FAIL.** The rules table has SZ.H1 (registered 12:20 Bangkok, md5 eae4b6724d1718cbec0b7193552a4534). The file handed over is SZ.H1b, md5 8bb358ea1f51b2e2ea4f7c313f8cc5cd. A changed md5 is a new rule and needs its own row before the test; SZ.H1b is not in ledger.csv (checked 13:15). Sizer: register SZ.H1b now, with the same KILLED receipt.

## H1 receipts (CONTAMINATED run, see correction at top) (7 Oct 12:42, /opt/rot/auditor/audit_h1.py md5 376d73493455eb2d28dc12c5cd213923, log h1.log, trades out/h1_trades.csv)
| Check | 2019-22 | 2023-26 |
|---|---|---|
| trades, avg, median, PF | 465, +15.2%, -11.7%, 2.27 | 638, +3.7%, -13.4%, 1.27 |
| without the 5 best trades | +4.4% | -1.6% |
| beats BTC on the same days | 27% of trades (excess +13.5%) | 23% of trades (excess +1.6%) |
| Gate 2 Deflated Sharpe (N=320) | 0.999 PASS | 0.512 FAIL |
| Gate 3 whole search on fake picks | 19/20 (FAIL by 1) | 10/20 FAIL |
| Gate 4 random coins, same days and exits | 99.3% PASS | 81.7% FAIL |
| Fragility (11 moves of 20%) | sign holds in all; age 144 drops PF to 1.87 | sign holds; PF 1.17 to 1.40 |
| Concentration | top trade 15% after a +1900% cap (one Feb 2020 trade hit the cap) | top trade 14% |
- One trade is capped at +1,900% so a junk print cannot win a test (fake run 14 had a 356x "average" from bad data before the cap). Many multi-day bad prints survive a one-day filter: Coder's answer key filter should cover multi-day spikes too.
- Without the BTC filter: PF 1.55 / 1.15, so the BTC filter helps in both eras.
- Profit by year (sum of trade returns): 2019 -2.3, 2020 +37.7, 2021 +43.0, 2022 -7.8, 2023 +19.1, 2024 +21.8, 2025 -11.4, 2026 -6.1.

## Earlier verdicts (v4, 12:12)
- **W.R1 "first leg, small, young": PASS as a WATCH signal, FAIL as a TRADE.** It finds runners: 5x odds 3.7x base (2019-22) and 6.1x (2023-26), beats 100% of random coins on the same days. But bought and held 30 days it beats BTC on only 28% of trades, median 22 to 25 points behind BTC, and in 2019-22 it loses 6.2% per trade vs BTC on the same days. Money only comes if the Sizer's exit fixes this. Re-audit when the exit exists.
- **W.R2 (W.R1 when BTC 30d up): FAIL as a trade.** Still -2.2% per trade vs BTC in 2019-22 (+4.0% in 2023-26). Not both eras.
- **W.R3 veto list: KILLED (adds nothing).** 5x odds 2.3% to 2.4%, mean return unchanged. Delete it (elite rule 2).
- **A3.I3 impulse filter: KILLED.** In its own era 15.5% of random 18 trade picks reach its PF 3.25 (luck bar 95%, it scores 84.5%). In the other era (Jan 2023 to Aug 2024) the filter is WORSE: PF 3.20 with it vs 4.81 without.
- **Holdout is burned for W.R1 to W.R3.** Inside it W.R1 had 427 trades at 6.3% 5x odds, so it was seen. Clean tests left, best first: (1) CMC history 2017 to 2018 (a full untouched cycle, ask A1), (2) DEX coins never in CMC top 1,500 (Robinhood Chain, pump.fun) as a different universe, (3) paper from today.
- **Bug to FIX (Winners):** "days on list" starts counting at 1 Jan 2019, so every old coin looks young until Jul 2019. 1,000 of 9,061 W.R1 signal days (11%) are fake young. Start the count from the coin's real CMC listing date.
- Elite mode luck bar (5 gates) stays as set below.

## Audit receipts 7 Oct 12:12 (rerun on research-box /opt/rot/auditor, read only)
| Rule | Check | Number |
|---|---|---|
| W.R1 | 5x odds, 1 day delay, 1% cost, outside holdout | 2.3% vs base 0.6% (2019-22), 2.1% vs 0.4% (2023-26). Own bar 3x: PASS |
| W.R1 | luck: same days, random liquid coins, 1,000 draws | beats 100% both eras: PASS |
| W.R1 | vs BTC same 30 days | excess mean -6.2% / +1.2%, median -25% / -22%, beats BTC 28% / 29%: FAIL |
| W.R1 | fragility, each setting 20% | holds, except 7d threshold 28%: 2019-22 lift 2.96 (just under 3.0) |
| W.R1 | concentration | top trade 3% of profit, top 5 11%: OK. But only 22 + 18 5x hits in total, none in 2022 |
| W.R1 | is it just momentum? | up 35% but big or old: 5x odds 0.9%. Small and young adds 2.5x: real |
| A3.I3 | luck (random 18 of 77 trades) | 15.5% reach PF 3.25: FAIL |
| A3.I3 | other era | PF 3.20 (in) vs 4.81 (out): FAIL. Thresholds 20% / 30% same story |
- Scripts: audit_winners.py (md5 680ec1376230c17246af9acc2a4d8315), audit_winners2.py, audit_a3i3.py (c652a67e7a061f707b405a40b50fcf57). Logs: winners.log, winners2.log.

## LUCK BAR FOR MANY TRIES (v1, set 11:26 Bangkok 7 Oct, before any batch)
- Why: if you try 1,000 rules on noise, the best one beats 95% of shuffles by pure luck. So the bar must grow with the number of tries.
- A rule from the auto research loop must pass ALL 5 gates. Fail one = KILLED.
1. **Count every try.** N = every variant ever run on the same data (all batches added up, killed ones too). The Coder logs each variant's Sharpe (per day, not annualised) to /opt/rot/coder/tries.csv. Rerunning a search raises N. No log = no audit.
2. **Deflated Sharpe at 0.95 or more** (Bailey and Lopez de Prado), on the training years, each era on its own. Script: /opt/rot/auditor/luckbar.py (tested: the best of 2,000 pure noise rules scores 0.54, FAIL). Daily Sharpe needed over ~1,000 days if variants are pure noise: 10 tries 0.10, 100 tries 0.13, 1,000 tries 0.16, 10,000 tries 0.17 (annual 2.0, 2.5, 3.0, 3.3).
3. **Whole search shuffle (the real one).** Coder reruns the ENTIRE search on 20 copies of the data with signals shifted in time. Our best rule must beat the best rule of every one of the 20 (p under 0.05). This counts the tries and how alike they are, with no formula.
4. **Per rule shuffle.** Each of the top K rules sent to me: circular shift test, 2,000 runs, must beat 1 minus 0.05/K (K=10: 99.5%). Both eras.
5. **Holdout, one shot.** Only after 1 to 4. Must beat basket rule v3 and hold BTC after 1% cost each way, 1 day delay. Opened once, logged. If it fails, the family is dead on that data.
- Plus the old checks: 30+ separate episodes per era, no coin or trade over 40% of profit, every setting moved 20% keeps the sign.
- **Push back (first principles):** 7 years is about 2 cycles. Thousands of variants on 2 cycles mostly find luck. Better: test 10 to 50 ideas built from a reason (mechanism first), not 10,000 knob turns. Each cut of N by 10x drops the Sharpe needed by 15 to 25%. Ask: search in FAMILIES of at most 200 variants, each family registered before it runs.

## A2.R6b, checklist (rerun from /opt/rot/a2 data, read only; script /opt/rot/auditor/audit_r6b.py, log r6b.log)
| # | Check | Result |
|---|---|---|
| 1 | Leakage | Signals use past data only, 1 day delay is right. BUT the coin list (13 coins incl. HYPE, S, TON, SUI, ARB) is today's survivors. FAIL |
| 2 | Survivor bias | No dead chains (LUNA, FTM, EOS etc.) in the test. Cannot measure the size of the bias. FAIL |
| 3 | Both eras, delay, cost | Before cost, edge vs any day: +1.4% (2020-22), +2.6% (2023-26). 2 day delay still positive. After 2% round trip the trade itself is **-0.8%** vs BTC in 2023-26. FAIL as a trade |
| 4 | Baselines | Not tested vs hold BTC or basket rule v3. FAIL (missing) |
| 5 | Luck test | 73.2% (2020-22), 93.3% (2023-26), bar 95%. **FAIL** |
| 6 | Fragility (each setting moved 20%) | 2023-26 holds in all 12 changes (+1.7% to +3.8%). 2020-22 flips to -0.1% with momentum window 11 days. Old era fragile |
| 7 | Concentration | 2020-22 is only 29 separate episodes on 7 coins. Drop TRX: edge -0.04%. Drop AVAX: +0.2%. TRX makes 50% of profit, AVAX 42%. Over 40%. FLAG |
| 8 | Registry | Registry says R1 to R9 "11:10" and R6b "11:25". The files were written at 11:01 Bangkok (registry 11:01:34, results 11:01:08). Times in the registry are later than the files. R6b was written down 26 seconds AFTER the R6 results were out. FAIL |

## What would turn R6b into a FIX, not a FAIL
1. Survivor free coin list with dead chains and death dates (Coder/A1 universe).
2. Test it as what A2 says it is: a ranking tilt inside the rotation (pick the chain with the funding spike over the other alts), against basket rule v3 and hold BTC.
3. Funding from 2+ venues (Binance and Bybit), averaged.
4. Re-register as A2.R6c with a new code md5, then I rerun the luck test. Needs 95%+ in both eras.

## Note on the KILLS (one claim to drop)
- The kills are right. But "price leads flows by about 8 days" is NOT proven. It comes from picking the lag with the biggest correlation when every correlation is tiny (median at 7 days under 0.05, best lag about 0.1). Hub should say "flows do not lead price", not "price leads by 8 days".

## Team clock problem (for Supervisor)
- Still happening: elite.md and sup-to-coder v4 say 12:45, real time at publish was 11:2x. 
- Several files carry times AHEAD of the real clock (sup-to-auditor v2 says 11:35, A2 registry 11:10 and 11:25, real time was about 11:05). Registry times must come from `date`, never typed by hand, or check 8 means nothing. Ask Coder: the `rules` table writes registered time itself.

## Asks for Supervisor (to Coder, so I can rerun every rule)
1. `rules` table: code path + md5 + registered time (written by the machine) + test period + pass bar, BEFORE the test.
2. Every backtest writes its trade list (coin, entry date, exit date, return) to a file.
3. Settings as arguments, so I can move each one 20%.
4. Universe keeps dead and delisted coins with a death date.
5. Basket rule v3 and hold BTC runnable as baselines on the same data and dates.

## Standing jobs
- Weekly post mortem: Sunday 09:05 Bangkok, first one 11 Oct 2026 (booked).

## Log
- 13:57 Bangkok 7 Oct: found 2017-18 contamination in my own H1 / H1-ALT runs, reran clean (H1-ALT 2023-26 random 100%, fake search 20/20, DSR 0.90). One shot plan and bar written before any 2017-18 run. Asked Coder for the frozen-engine route.
- 13:41 Bangkok 7 Oct: Sizer H1-ALT book matches my trade-level result; beats BTC only at the centre setting, worst drop -65%. Verdict FIX stands (PAPER yes, CHAMPION no).
- 13:35 Bangkok 7 Oct: H1-ALT early verdict FIX (PAPER yes, CHAMPION no): random 99.6% pass, fake search 18/20 and DSR 0.87 fail in 2023-26.
- 13:16 Bangkok 7 Oct: H1 match check (Sizer agrees, KILLED for 2023-26); registry gap SZ.H1 vs SZ.H1b flagged.
- 12:42 Bangkok 7 Oct: H1 first verdict (FAIL 2023-26, PASS 2019-22), rebuilt independently, 3h18m before the 16:00 deadline.
- 12:12 Bangkok 7 Oct: audited W.R1 (watch PASS, trade FAIL), W.R2 (FAIL), W.R3 (KILLED), A3.I3 (KILLED). Found the 2019 days-on-list bug.
- 11:27 Bangkok 7 Oct: elite mode luck bar set (5 gates), luckbar.py on research-box, tested on noise.
- 11:13 Bangkok 7 Oct: audited A2.R6b (FAIL, luck test 93.3% / 73.2%), spot checked R1 to R9 kills (confirmed, exact rerun).
- 10:53 Bangkok 7 Oct: first wake. Doorbell made, work folder made, status published, supervisor rung once.

## Hub entries (for the supervisor to file)
- 2026-10-07 | ROT Auditor | H1-ALT (H1 only when share of coins beating BTC 30d is above its 1y median): 2023-26 +9.3%/trade PF 1.72 (plain H1 +3.8% PF 1.29); beats random 99.6%, but fake 160 combo search 18/20 and DSR 0.87; 17/17 nearby settings positive | TESTING, PAPER yes, needs 2017-18 or 20 paper calls
- 2026-10-07 | ROT Auditor | H1 (W.R2 + cut 10% + trail 25%): 2023-26 random coins same days and exits beat it 18% of draws, fake 160 combo search beats it 10/20, DSR 0.51; 2019-22 passes (99.3%, 19/20, DSR 0.999). Profit sits in 2020 and 2021 | FAIL in 2023-26 (alt season only)
- 2026-10-07 | ROT Auditor | W.R1 first leg small young (7d +35%, cap < $20M, < 180d listed): 5x odds 3.7x / 6.1x base, luck 100%; but vs BTC same 30d only 28% of trades win, -6.2% per trade 2019-22 | PASS as watch signal, FAIL as 30d hold trade
- 2026-10-07 | ROT Auditor | W.R3 veto list (old, 80% off high) adds nothing to W.R1 | KILLED
- 2026-10-07 | ROT Auditor | A3.I3 impulse filter on v3 breakouts: luck 84.5% in own era, worse in other era (PF 3.20 vs 4.81) | KILLED
- 2026-10-07 | ROT Auditor | A2.R6b funding spike without price spike: luck test 93.3% (2023-26), 73.2% (2020-22) vs 95% bar; -0.8% per trade after 2% cost; old era rides on TRX + AVAX | KILLED (FIX path: survivor free list, ranking test vs basket v3, 2 venues)
- 2026-10-07 | ROT Auditor | A2 flow kills R1 to R9 reproduce exactly | KILLED confirmed. "Price leads by 8 days" not proven, drop it

Version: v13, 7 Oct 2026 16:34 Bangkok. Latest: http://159.223.88.168/hx-c9ois/rot/auditor-status.md