# Audit of dc-plan.md v2: the 10 weakest assumptions (DC Auditor)

## READ THIS FIRST
- 10 assumptions the plan leans on that I think are weakest, ranked by how much $ a day they could fake. Each has a cheap test (under 3 h) that would break it. Numbers marked **measured** I ran today from dc/data; the rest are tests to queue.
- Top 3 to queue first: **W1 (half B cut short), W2 (sellable means public router), W3 (every T1 winner is also a Second Window coin).**

| # | Assumption in the plan | Why I doubt it | Cheap test that breaks it | Hours |
|---|---|---|---|---|
| W1 | "Honest 7 d outcome" can be scored on half B now | Data ends 4 to 5 Oct; coins from 30 Sep on have under 7 days; liqlock pulls missing after 6 Oct 13:14. Missing pulls = losers scored as holders | Fix one evaluation cut: horizon 48 h for every coin, outcome data must cover it, pull coverage flag per trade. Rerun DC0 baseline both ways; if $ a day moves 20%+, every result so far is on the wrong ruler | 1 |
| W2 | "Sellable cap" ($100+ sell happened) means WE can sell there | **Measured:** the skipped campaign family (5059_7adb2ec7) shows 30 of 40 triggered coins at a sellable 5x, yet normal apps cannot sell (99.6% of trades through its own router). So "sellable" counts sells we cannot make | Recompute sellable using only public router sells (Uniswap Universal Router, known aggregators). Count how many T1 triggers and 5x labels survive | 2 |
| W3 | The deep check finds NEW money | **Measured:** every T1 coin that does 5x from $250k must cross $1M, the Second Window's gate. Only 76 of 852 kept T1 coins ever reach a sellable $1M. So T1 winners are a subset of SW gate coins; our only new $ is the $250k to $1M leg (minus the losers SW never touches) | Split DC0 P&L into (a) leg before the coin hits $1M, (b) after. Report (a) alone as Third Window's new $. Also: does T1 at $250k beat SW's $1M gate on the same coins per $ risked? | 1 |
| W4 | The 8 h signal (AUC 0.89) is a real deep check signal | **Measured:** 15 of 26 class A runners were at $5M by hour 8, 7 at $20M. The AUC mostly rereads the label | Redo with a forward label: buy at the 8 h sellable price, outcome = 3x from THAT price. Expect AUC to fall a lot; if under 0.65 the "wait to 8 h" idea loses its main receipt | 1 |
| W5 | 20 to 60 triggers a day give enough winners to judge a module | **Measured:** A3's universe has 12 coins at 5x in 20 days (5 in A, 7 in B). With 7 positives, one coin is 14% of B. Module tests at 5x have almost no power | Power check: at 7 positives, the AUC 95% band is about plus or minus 0.2. Switch the module label to path P&L (continuous) or 3x, and require 10+ positives per half | 0.5 |
| W6 | Coins are independent trades | **Measured:** 325 of 852 triggered coins (38%) share a ticker with another triggered coin (SI x12); creator 0xe0609fc152f22ccac765c306b6b14bddb75a00a0 has 12. Three of the 5x winners come from multi-coin creators | Cluster by ticker, creator, funder root; one trade per cluster per day; rerun DC0 and every module | 1 |
| W7 | The skip list only removes losers | **Measured:** the bundle family (341_6de2bb44) has 80 of 97 triggered coins at a sellable 5x (on peak, not path). Skipping it may delete the richest lane, or it may be fake (pull speed, W2) | Run the honest path book on the 153 skipped coins alone, public router sells only, pulls at 0x. If positive in both halves, the "fast flip" lane is an edge, not a skip | 2 |
| W8 | The Second Window's exits (0.3x stop, 5x sell) fit a $250k entry | They were tuned on $1M gate kit coins. A $250k coin is earlier and wilder; 0.3x stop vs pull timing differs; live stop was 14.8% worse than paper (SW A2.30) | Grid of 3 stops x 3 targets on half A only, applied to B; plus 15% haircut on stop exits. If the baseline sign flips with the exit, no module result is stable until exits are fixed | 2 |
| W9 | L2 nightly champion and challenger, swap after 3 forward days | ~0.6 winners a day: 3 days hold about 2 winners. A random challenger will "beat" the champion often | Replay 20 past days: 100 random challengers vs a fixed champion, count how often a random one wins 3 days in a row. If over 10%, L2 swaps on noise; require 14 forward days or a bigger sample | 1 |
| W10 | Unique buyers and buy counts mean real people | A1 uses buy swap count as the buyer proxy; SW A1.18 found most new buyers are fresh disperse funded wallets | On 21 Sep to 4 Oct (wash.db), compare T1 by buy count vs T1 by unique wallets vs unique non-farm wallets. If more than 20% of triggers differ, the trigger is defined by bots | 1.5 |

## Smaller holes (no test needed, just fix the text)
- **Prereg timestamps are typed, not stamped.** Plan should say: time from `date`; the Auditor uses bus.json change times as the truth.
- **Analysts' preregs live only in their own files.** dc-prereg.md has one line (DC0). A1, A2, A3 must each append their line there, or the plan's "no line, no edge" fails all three.
- **"Beat random 95%"** must say: same size, same half, same exits, same fills, and best of as many random rules as rules tried (C19).
- **"$ a day"** must say which venues (v4 only vs all) and which ticket ($100 or $1,000).
- **Fixed payoff money models** (A3: +$400 / -$75) are fine for a first look, never for a $ claim.

## Cost
- 0 X credits, $0 API, 0.8 h. All "measured" numbers rerun from dc/data (launches_15sep_4oct.csv.gz, crossers500k_7d.csv.gz, coins54.csv, runner_paths_310.csv.gz) with A3's trigger rule (852 coins, same count as A3).

Version: v1, 7 Oct 2026 10:20 Bangkok (stamped from `date`), DC Auditor. Latest: http://159.223.88.168/hx-c9ois/dc/dc-audit-plan.md