Polymarket Coherence Arb

Detects logically-linked Polymarket markets that price incoherently (a mutually-exclusive, exhaustive set whose YES prices don't sum to ~1) and trades the cheap/rich legs back toward coherence. No external data — arbitrages Polymarket against itself. Exclusivity is confirmed from market text; unconfirmed sets are alerted, never arbed. Sim by default.

Finds sets of related Polymarket markets that should satisfy a probability constraint but don't, and trades the cheap/rich legs back toward coherence — arbitraging Polymarket against itself.

The edge

No sharp book to outrun, no external data, low latency pressure. If the YES prices of a complete, mutually-exclusive set sum materially above 1, the set is collectively overpriced (buy NO on the richest leg); if materially below 1, it's underpriced (buy YES on the cheapest leg).

What it actually trades (v0.1)

1. Confirmed sets only. The only relationship v0.1 confirms from market text is a complete FIFA "win Group X" set — exactly 4 legs, same group letter (2026 format). Mutual exclusivity is read from the text itself; if a set can't be confirmed (partial, mixed letters, advance/qualify wording) it is alerted, never arbed.
2. Bring-your-own sets. Set COHERENCE_GROUPS to trade explicit market-id groups you assert are mutually exclusive (e.g. a tournament/final ladder). Format: "id1,id2,id3,id4; idA,idB,idC" — ; separates groups, , separates legs.
3. Remix point. discover_sets() in discovery.py is where you add other relationships (win-tournament ≤ reach-final, parlay consistency, etc.). The skill handles pricing, the coherence test, the context/exposure gates, and execution — you supply which markets are linked.

This is not the WC Group Repricer: there's no repricing-timing mechanism and no Elo tiebreak — it's pure price-coherence, leg picked by price alone, generalizable beyond the World Cup.

Safety rails

• Sim-first. Default venue is $SIM; real trading requires --live and --venue polymarket and a claimed, wallet-linked Simmer agent. --live and --dry-run are mutually exclusive.
• Budget = open exposure, not daily spend: new trades are rejected if open_exposure + cost > DAILY_BUDGET_USD. State updates are lock-protected (fcntl.flock), and live vs dry-run state live in separate files (state_live.json / state_dry.json).
• Dry-run never touches live state and routes through the SDK paper engine (live=False) rather than stubbing success.
• Prices are passed raw to the SDK (no pre-rounding); edges are computed vs the ask when the venue exposes one (falls back to mid on $SIM — see Known limitations).
• One position per market; every trade carries skill_slug, a public reasoning string, and signal_data (Autoresearch backtest-ready).

Configuration (env — all knobs Autoresearch-mutable)

Var	Default	Meaning
TRADING_VENUE	sim	sim or polymarket (with --live)
MAX_TRADE_USD	20	per-leg cap
DAILY_BUDGET_USD	100	open-exposure cap
MIN_COHERENCE_GAP	0.05	min `
MAX_SLIPPAGE_PCT	0.03	skip legs whose context slippage exceeds this
COHERENCE_GROUPS	(unset)	optional explicit market-id groups (see above)

Usage

python coherence_arb.py --dry-run            # default; sim paper pass
python coherence_arb.py --status             # show positions/exposure
python coherence_arb.py --live --venue polymarket   # real money (after sim record)

Requires SIMMER_API_KEY.

Known limitations (v0.1)

1. $SIM (LMSR) has no order book — ask/spread/depth gates only bind on the real venue; sim validates logic, not microstructure. Capturing this edge cleanly often needs near-simultaneous fills, which $SIM cannot model — validate in live=False paper mode against real spreads.
2. Auto-discovery confirms only the WC "win Group X" set from text; the canonical WC tag/series slug is unverified upstream, so discovery may need the paginated sports-markets workaround at launch. Other relationships require COHERENCE_GROUPS or a discover_sets() remix.
3. Legs are entered one at a time, not atomically; partial-fill unwind is basic.
4. Inconsistencies are often small and short-lived — MIN_COHERENCE_GAP ships at a reasonable default and is meant to be tuned in sim / by Autoresearch, not trusted blindly.
5. Real-venue depth (L2) checks are a TODO hook; v0.1 uses the context slippage gate.

No performance claims are made or implied. This skill describes what it does, not what it returns.