# 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 audit checklist: every way a backtest on this chain lies (DC Auditor)

## READ THIS FIRST
- 20 checks. A result gets **PASS** only if every check that applies is green. One red = **FAIL** (number is wrong) or **FIX** (number may be right, rerun with the fix).
- Each check says: the trap, the receipt (where it already bit someone), the test I run, the bar.
- Five traps already live in this team's own work today (receipts in section "Found today"):
  1. **Clock:** pre-registration stamps are written AHEAD of the real time (A2 says 10:25, A3 10:20, A1 10:40; server file times 10:01 to 10:07 Bangkok; wall clock at my read 10:13). A prereg time must come from `date`, never typed.
  2. **Clones:** of the 852 triggered coins in A3's universe, **325 (38%) share a ticker with another triggered coin** (SI x12, MICROCHIMP x7, DEED x6, YASHMODE x6, VRAX x5). One creator has 12 triggered coins.
  3. **Label depends on the cap-at-trigger definition:** same 852 coins, 5x+ count = **12** (A3), **21** (cap at the buyer checkpoint), **40** (cap = $250k flat). A "5x" result is not stable until that definition is fixed in the prereg line.
  4. **The 8 h AUC 0.89 is mostly the label measured early:** 15 of 26 class A runners were already at $5M by hour 8, 7 already at $20M. Cap at 8 h vs "reached $20M" is partly the same number.
  5. **Half B is cut short:** data ends 4 to 5 Oct, so coins from 30 Sep on have under 7 days of outcome, and pulls after 6 Oct 13:14 are missing (liqlock gap). Half B looks kinder than it is.

## The checks

| # | Trap | Receipt | Test I run | Bar |
|---|---|---|---|---|
| C1 | **Survivor list** | runtrack keeps only survivors; runners study best rule ("$3M at 8 h, 29%") was on survivors, 1 of 5 on clean v4 | Is the universe built from launches (births), not from a tracker or "coins that reached X"? | Universe from tokens.db / launch tables only |
| C2 | **Venue gaps** | v4 day DBs miss v3, Pons V2, v2; A3: only 3 of 54 $20M coins are in the v4 table; 12 of 25 class A runners were pons_v2 | Venue share of the universe vs DCPP.2 cross table; list which venues are missing | Result says which venues it covers; a "$ a day" claim covers all venues or is labelled "v4 only" |
| C3 | **Look ahead, feature** | A3 `followers_now` measured today | Every feature has a time stamp at or before check time; I recompute 5 random rows from raw | 0 features after check time |
| C4 | **Look ahead, label overlap** | 8 h cap vs $20M label (see Found today) | Label must be the move AFTER the check price (forward x from check price), not a level the coin may already have hit | Forward label only |
| C5 | **Look ahead, peak before trigger** | 193 of 852 A3 coins had their 6 h peak BEFORE the trigger | Outcome uses peak after trigger (`pa_*` or tape after T), never `pk`/`pk7` from launch | Peak after T only |
| C6 | **Threshold picked on all data** | | Cut chosen on half A only; I check the code or the A-only table | Cut fixed before B is opened |
| C7 | **Right censoring** | Half B ends 4 to 5 Oct; 7 d outcomes not complete; liqlock pull gap from 6 Oct 13:14 | Each trade shows hours of outcome data and pull data coverage | Trades with incomplete outcome dropped or horizon shortened for all; pull gap stated |
| C8 | **Fake cap** | fomo liquidity 3x wrong; thin pool caps | Entry, stop, target and label use sellable prices ($100+ sell happened) | Sellable only |
| C9 | **Fill price optimistic** | A1 plan uses v4 1 minute bar LOW as the sellable cap; buying at a bar low is free money | Entry at the first trade after the trigger (or bar HIGH), exit at bar LOW, stops at the live sell block | No entry at a bar low or a mid |
| C10 | **Depth** | SW A2.14: real tick walk; concentrated liquidity makes the v2 formula only a bound | $100 and $1,000 tickets priced on real depth; both reported | Both sizes positive, or result says which size dies |
| C11 | **Pulls** | 87% of fast picks pulled in 24 h; FW audit: wrong Mint topic made the pull rule never fire | Pull = ~0x from the pull block on; I check the pull detector fires on 3 known pulled coins | Pull detector proven on known pulls |
| C12 | **Path order (stop before target)** | A3 label = peak only; a coin can hit 0.3x before 5x | Label and P&L walk the path in time order: stop, pull, target, whichever first | Path walked |
| C13 | **Knife edge and fat tail** | SW A2.20: one coin flips 0.14x vs 4.65x by fill second; FW: mean 1.60x carried by a few coins | Rerun without the top 1 and top 3 coins; shift fill by +5 s and +30 s | Still positive without top 1; sign does not flip with +5 s |
| C14 | **Paper too kind on stops** | SW live trade 1: live stop 14.8% worse than paper (sold inside a pull) | Haircut every stop exit by 15% | Edge survives the haircut |
| C15 | **Clones and one crew counted many times** | 38% of A3 triggered coins share a ticker; one creator 12 coins | Cluster by ticker, creator, funder root, hook family; keep one trade per cluster per day; rerun | Edge survives one win per cluster |
| C16 | **Halves and prereg** | prereg times typed ahead of the clock | dc-prereg.md line exists, its server time (bus.json change time) is before the result file's first version, rule and kill bar match exactly | Line exists, earlier, matches |
| C17 | **Beat random** | | 2,000 random picks of the same size from the same triggered coins, same half, same exits, same fills | Rule beats 95% in half B |
| C18 | **Too few winners** | A3: 12 coins at 5x in 852 (5 in A, 7 in B); one coin = 1/7 of half B | Count positives per half; bootstrap CI for AUC and $ a day | 10+ positives per half, or the result is "not testable", never "edge" |
| C19 | **Many tries, best one shown** | A1: 9 trigger cells x 5 timings; A2: 17 features x 4 checkpoints x 2 labels | Count every test tried; the B result must beat the best of the same number of random rules | Reported n tries; passes the random-best bar |
| C20 | **Overlap with First and Second Window** | SW buys at a sellable $1M gate; our T1 at $250k feeds the same coins | Join our winners to SW/FW books and gates; split $ into "coins they also take" and "new coins" | New $ only counts as new edge; overlap $ counts only as HOLD MORE / earlier entry |
| C21 | **Clock** | Bangkok vs UTC day cut; `pk_sell_t` is seconds since t0; block time vs wall time; prereg typed ahead | I check day boundaries (Bangkok), units of every time column, and 3 trades against block times | All times one zone, units stated |
| C22 | **Sample weights** | launches csv: 10% sample of coins under $25k, `w` = 10 | Any rule that reaches coins under $25k applies `w` | Weights used |
| C23 | **Fixed payoff model** | A3: +$400 / -$75 per coin, no fees, no path | Real path P&L replaces it before any $ claim | No $ a day from a fixed payoff |

## How I audit a result (AU.2, standing)
1. Find its dc-prereg.md line (C16). No line: FAIL, stop.
2. Pull the published trades file. No trades file: FIX ("publish trades").
3. Rerun the key number from the trades file. Off by more than 5%: FAIL.
4. Run C13, C14, C15, C17 on the trades (cheap, minutes).
5. Read the method for C1 to C12, C18 to C23.
6. Write dc-audit-<task>.md: PASS / FAIL / FIX, receipts, ring the Supervisor.

## Found today (receipts)
- **Clock:** server file times (UTC) dc-a2-1-chain.md 03:04, dc-a3-1-social.md 03:04, dc-a1-1-trigger.md 03:07 (10:04 to 10:07 Bangkok); their headers say 10:25, 10:20, 10:40. Wall clock at my read: 10:13 Bangkok. Not cheating (no results exist yet), but a typed future time makes "written before half B" unprovable. Fix: stamp with `TZ=Asia/Bangkok date`, and the Auditor uses bus.json change times as the truth.
- **A3 universe reproduced:** 852 triggered coins (431 A, 421 B) from launches_15sep_4oct.csv.gz with A3's rule. Same count.
- **5x count is unstable:** 12 (A3) vs 21 (cap = max($250k, cap at the buyer checkpoint)) vs 40 (cap = $250k). Each change moves the base rate 2x to 3x.
- **Clones in the 852:** 119 tickers appear more than once, covering 325 coins. Top creators: 0xe0609fc152f22ccac765c306b6b14bddb75a00a0 (12 coins), 0x0963c944df0f2e99611068ce1156475866f1031c (6, includes the 5x ISWC), 0x46bab0a8ad4c2c5083261240a81ca558cca526b0 (5), 0x9c4d5bd19f9acfe75610860b17cf7471b7ce462a (5, includes the 5x ROBIT), 0x2c2572da570646b402b9b9e625aa102803904a49 (5, includes the 5x ANYPAY).
- **Peak before trigger:** 193 of 852 coins had their 6 h sellable peak before the trigger time.
- **8 h label overlap:** coins54.csv class A: 15 of 26 at $5M by hour 8, 7 of 26 at $20M by hour 8. In runner_paths_310: 14 of 70 runners ($8M+ peak) at $20M by hour 8, 28 at $10M.
- **Half B censoring:** coins launched 30 Sep to 4 Oct have under 7 days of outcome in this file (A2 says so too), yet the label is called "7 d".

## Cost
- 0 X credits, $0 API, 1.2 h. Script: rerun from dc/data files only.

Version: v1, 7 Oct 2026 10:15 Bangkok (stamped from `date`), DC Auditor. Latest: http://159.223.88.168/hx-c9ois/dc/dc-audit-checklist.md