# A1 review #256: A2-FAST5 (u100 entry +5 s)
Verdict: AGREE, send to paper. Checked the write-up, not re-simulated.
- Sample is fixed (same 397 pools at every delay), halves agree, gate R>=1 ETH is known at entry: no look-ahead found. Effect is monotone in delay (1.87 -> 1.01), which is what a real ramp edge looks like.
- Consistent with my own side: hooked pools enter-early gain small (2.58 vs 2.37), v2 ramp buy <=5 s; early-in-the-ramp is a repeated pattern across families.
- Conditions for the paper test:
  1. Fill model: block-quantised price at first block after (signal + feed lag 1.5-2 s), not the +5 s price. Run arms at +5 s and +8 s so the lag cost is visible.
  2. Pull handling: pull must be a real liquidity-removal exit (>=50% of max), not 0 only; report share of trades ended by pull.
  3. Size cap: position / reserve impact at the actual size; report mean at R<1 ETH separately (1.4, weaker).
  4. Outside-money: 500/500 first swaps are the ramp bot; no tx.from yet. Count non-fleet buyers before entry once txfrom2 lands (risk is the same fleet-inflation as elsewhere).
  5. Arms: tp3 (main), tp5 (B); gate R>=1 ETH as arm C.
- Kill (A2's): forward on-time mean <1.2x over 40 trades. Add: if pull share >25%, stop.
- Live not proven: execution latency unknown. Ben decision only after forward n>=100.