# Paper test legitimacy audit (Ben, 9 Oct: "the hooked tests look a little too good to be true")

Your job: make sure every paper test is legit. Until a test passes, the tests page marks it "⚠ not verified yet" and sorts it below verified tests. Only you set `"verified": true` (plus `"verify_link"` to your audit file) in /opt/crew/paper/registry.json.

## Red flags right now
- **Hooked fee-100:** 7 of 7 winners, 0 losers, every arm positive. **H3000:** 3 of 3 winners, means of 18x to 37x. No real market does that. Analyst 2 already found that outside wallets never sold above 5x in 113 H3000 pools (#138). So the 20x and 30x arms are probably prices nobody can actually sell at.
- **Live-rule replay** (grid, LIVE variant): +$2,860 at $10 over 158 trades with a 22% win rate. The real live wallet over the same rule has 2 winners in 25 trades and is down 0.14 ETH since 8 Oct. Either live is unlucky or the replay fill model is too kind. This decides whether every backtest we run is trustworthy, so it comes first.

## Checks for every test (all five, written down per test)
1. **Sellability:** can this token actually be sold? For hooked pools: read the hook code (beforeSwap/afterSwap fees, sell blocks, owner switches), and do an eth_call sell from a fresh address at the exit block. If unclear, the $5 real sell test goes to Ben (#138).
2. **Outside sellers:** for each paper exit, did real wallets (not the deployer or fleet) sell within ±60 s at a price within 20% of ours? Report the share of exits that pass. A single print is not a price.
3. **No look-ahead:** the entry decision uses only data that existed at entry time (no pool filters built from later data, no "survived" filters).
4. **Fills:** our size against the pool depth at that block, fees, slippage and gas exactly as live. Rebuild 3 trades per test by hand from chain data (block, reserves, amounts out) and compare.
5. **Replay vs live:** on coins live also traded, put the paper or replay result next to the real fill. Explain every gap over 20%.

## Output
One file per test (fw/tm-audit-<id>.md): the 5 checks with numbers, then verdict: verified / fix the test (what) / kill. Post a one-line summary to manager. Order: grid replay vs live first, then h100, h3k, p204. Same audit for every new test before it can be verified.