WC Group Repricer

A Groups-category skill for the 2026 World Cup built on market dynamics, not winner-picking:

1. Repricing timing — enter group favorites at pre-tournament asks; exit after qualification becomes "obvious" and retail money reprices the leg up. "Obvious" is read from the market itself (the leg crossing a configured price threshold) — no sports model required. (Strategy family credit: group-stage repricing ideas circulating on X.)
2. Group-set coherence — within a group's winner set (mutually exclusive, exhaustive), trade prices that sum incoherently back toward 1. Mutual exclusivity is confirmed from market text (a complete 4-leg "win Group X" set) — if it can't be confirmed, the set is alerted, never arbed.
3. Elo anchor (tiebreak only) — a static Elo table decides which leg of an incoherent set is the mispriced one. It never picks winners standalone.

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).
• Dry-run never touches live state (separate state_dry.json) and routes through the SDK paper engine 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.

Configuration (env — all knobs Autoresearch-mutable)

Usage

python scripts/group_repricer.py --dry-run            # default; sim paper pass
python scripts/group_repricer.py --status             # show positions/exposure
python scripts/group_repricer.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.
2. Market discovery uses a text search for "win Group X" sets; the canonical WC tag/series slug is unverified upstream — discovery may need the paginated sports-markets workaround at launch.
3. Coherence legs are entered sequentially, not atomically; partial-fill unwind is basic.
4. The exit threshold (EXIT_REPRICE) is the strategy's most sensitive knob — it 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.
6. Platform per-leg stop-losses can fire on temporarily-down legs of a coherent set — review auto-risk settings before running live.

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