We make two distinct claims about the cheap band. Conflating them is the failure mode of every "follow the whales" product we've audited, so we keep them apart here.

Claim 1: cheap-band BUYs have a structural (1−P)/P advantage

This is arithmetic. A 34% hit rate on BUYs at avg price 0.30 yields:

EV = 0.34 × (1 − 0.30)/0.30  −  0.66 × 1
   = 0.34 × 2.333 − 0.66
   = +13.3% per trade

That positive expectancy is true regardless of who clicks Buy. It falls out of payoff structure: cheap entries pay 3–5× when right, full loss when wrong, and the cheap band has historically resolved in the buyer's favor often enough to clear the threshold. The mirror claim — buys at price >0.60 being net-negative even at 70%+ hit rates — is the same math in reverse. This is the part of the thesis we have receipts for.

Claim 2: wallet reputation — tested and failed

The wave-34 cheap_conviction gate ( resolved_trades ≥ 10 AND pnl_per_dollar ≥ 0.05) was our operationalization of the "informed capital" claim: only fire (and only auto-trade) on cheap-band BUYs from wallets with proven historical PnL.

The wave-57 audit tested it properly — point-in-time rep, no survivorship — on every historical cheap-band BUY $25K+ in 0.20–0.30. The result, surfaced on /transparency:

A_proven   (passes gate)        n=11  hit_pct=9.1%   net_roi=-68%
B_unproven (rep, below threshold) n=26  hit_pct=53.9%  net_roi=+99%
C_no_rep   (no prior resolved)    n=11  hit_pct=45.5%  net_roi=+76%

The gate was filtering AGAINST alpha. Proven wallets — most of whom built their PnL on different bet types (favorites at 0.7+ or knowledge-specific markets) — bring no edge to cheap-band BUYs and often bring negative edge when they bet outside their domain. The cheap-band positive expectancy comes from the structural (1−P)/P math regardless of who clicks Buy.

Wave 61 dropped the gate. The cheap_conviction detector now fires on every $25K+ BUY in [0.20, 0.30] regardless of wallet rep. Paper portfolios A and B no longer filter by allowed_gate_reasons. Wallet rep is captured in detector_context as descriptive metadata for display only. Full receipt at the rep-alpha audit.

The wave-62 follow-up: the U-shape

Re-running the audit with six rep tiers (instead of binary A/B/C) exposed structure the headline was hiding. Cheap-band hit rate is non-monotonic in prior PnL/$:

Rep tier (prior PnL/$)             | n  | hit %  | 95% CI
no_rep                              | 11 |  45.5% | [21.3, 72.0]
known_loser (<0%)                   | 18 |  44.4% | [24.6, 66.3]
mild_winner (0-5%)                  |  4 |  75.0% | [30.1, 95.4]  ← sweet spot
good (5-20%) [wave-34 gate band]    |  4 |  25.0% | [4.6,  69.9]
high (20-50%)                       |  2 |   0.0% | [0,    65.8]
elite (>=50%)                       |  8 |  37.5% | [13.7, 69.4]

Wave-67b tightened these numbers: backfilled markets_cache.resolved_at
from signal_outcomes (1,195 markets) replacing end_date + 24h heuristic
on the rows the audit reads. U-shape persists; mild_winner shifted
86% -> 75% as some wallets re-classified into lower tiers.

The wave-34 threshold landed on the worst slice — wallets with a comfortable winning record entering a band they don't usually trade. Wave 63 added a forward-test paper portfolio (D) that auto- trades only the mild_winner tier. Wave 64.3 backfilledgateReason on historical fires so live cohort attribution on /transparency is comparable across tiers — and the U-shape is replicating in paper PnL with each new closed position. The audit's mild_winner prediction was 75% on n=4 (Wilson CI [30%, 95%]); see the transparency tables for the live forward-confirmed win rate, realized PnL, and current sample size.

Live fire badges on /cheap-conviction tag every signal with its trigger cohort so users can mentally filter at the moment of decision.

Live narrowing of the band

Live data narrowed the band twice. Wave 29 killed 0.30–0.40 after it hit 17.6% / −16pp on n=20. Wave 31 killed 0.05–0.20 after it hit 0% / −17pp on n=2 with the longshot sample trending the same direction. Current live band is 0.20–0.30.

Every detector below either targets that cheap-band directly or has been disabled because calibration showed it doesn't have edge. Every fire is graded against actual market resolution where applicable.

Two grading paths depending on signal type:

kelly_fraction  = min(0.25, edge / (edge + 0.5))
position_size   = min(kelly_fraction × bankroll,
                      bankroll × max_position_pct)

Each portfolio sets its own max_position_pct (5–10%) and overall risk caps (max open notional, max single-market exposure, max drawdown auto-pause, per-position stop-loss). See the public strategies leaderboard for the live numbers.

Polymarket charges roughly 2% on exits. Every PnL number on this site subtracts 200 bps for fees before showing the net. If you hold to resolution and win, the fee doesn't apply, but we treat every trade as a round-trip to stay conservative — the transparency page replays every signal at 200 bps to keep us honest.