{"sources": [
 {"url": "https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/filter-brief.md", "quote": "a retail fill is a weak observation of the rival's marginal price through the router split (the rival is always within 30 bps of p at retail time)."},
 {"url": "https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/filter-brief.md", "quote": "Inside that structure the quality of the price estimate is worth about 13 points. RS2 (520 board) has about 8 of them left to take. Knowing the rival pool's position exactly is worth +0.9; knowing the run's sigma, arrival rate and order size, +0.04."},
 {"url": "https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/filter-brief.md", "quote": "98% of arb loss is retail-induced. A large lognormal order knocks the spot out of the fee band and the arbitrageur unwinds it at once, taking about twice the impact gain the order paid."},
 {"url": "https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/oracle_table.json", "quote": "askFee = 1 - exp(ln(spot/p) - ask*0.003); bidFee = 1 - exp(-bid*0.003 - ln(spot/p)); clamp to [0, 0.10]"},
 {"url": "https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/oracle_table.json", "quote": "\"replica_score_1000_seeds\": 532.45"},
 {"url": "https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/replica.py", "quote": "RNG streams differ from the Rust engine, so absolute scores move by the seed set;"},
 {"url": "https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/replica.py", "quote": "sigma = cr.uniform(0.000882, 0.001008)"}
]}