We ran the late-favourite sweep on 8 Polymarket crypto books for 7 days with live...
The daily research thread from our internal kill ledger.
Posted first to X.
We ran the late-favourite sweep on 8 Polymarket crypto books for 7 days with live money. It lost on every book that traded meaningfully. The strategy is refuted.
Prior simulation showed the rule clearing a pre-registered bar on a queue-realistic replayer — 198 maker fills, net positive per bar, t-stat above 6. A 34-hour single-book probe didn’t lose. We expected that to hold.
The test: post-only buys in the last 60 seconds of a 5-minute bar, at bids already in a tight band near the top of the price range, hold to settlement. 8 books, 8 accounts, 7 days, 615 fills.
Result: −$170 across 7 days, −5.77% of capital. Every book that traded a meaningful number of bars finished negative. The arithmetic is the problem — at these prices, breakeven requires losing fewer than 1 in 44 fills. We ran roughly twice that rate.
A free cross-asset tape screen from before the test had flagged most of these books as negative or sign-unstable. Live confirmed it. The screen cost nothing and was more correct than the $170 test that was supposed to validate it.
Three API failure modes cost real money or real measurement: fills silently dropped from the trades endpoint biased PnL estimates upward; a stale positions endpoint fired a safety stop spuriously; order submission returned 503 while the API reported the market open.
A strong win rate at the top of the price range is not evidence of edge — it’s the arithmetic of the band, and the band eats you on the losses.