prediction-market-live-ops

Prediction-Market Trading Doctrine

Hard-won operational knowledge from a live Kalshi campaign (Aug–Sep 2026: crypto/commodity maker ladders, regime tilts, tennis dust vacuum — all falsified honestly, infrastructure open-sourced). Reference implementation: https://github.com/agiprolabs/kalshi-stack. Complements prediction-market-strategy (edge selection/sizing) — this skill covers the OPERATIONAL side: validating, executing, and killing strategies against a live venue without losing more than the lesson costs.

The prime directive

Every strategy decision is evidence-gated: backtest → paper → live micro-probe → ratchet. Each stage has a PRE-DECLARED kill condition with correctly calibrated probability math, written down BEFORE the data arrives. When a threshold turns out miscalibrated, correct it openly and re-declare — never silently move a goalpost in either direction.

Backtest hygiene (each rule was violated once, expensively)

• Print-replay EV of passive-fill strategies is an UPPER BOUND, never an estimate. The prints a resting order actually captures are an adversely-selected sample of the tape (sweeps deep enough to reach your queue skew toward true collapses). A tennis dust book that measured +125%/$ on prints went 0-for-78 live (P<2%). Only fill-conditioned live outcomes validate a maker strategy.
• Time-series joins use the last bar ending ≤ the decision instant. A candle containing the instant is look-ahead and will fabricate covariates that vanish out-of-sample.
• Taker backtests need print-confirmed executability, not quoted-ask edges — stale quotes create phantom fills.
• Any bucketed-PnL review includes a top-N-share column. Lottery-payoff books concentrate EV; a "pattern" that is two elephant rounds is noise. Same rule when a holdout "wins": check concentration before believing it.
• Early settlements are selection-biased toward losers (collapses end fast; comebacks are long matches/rounds). Never read a verdict off the fast-settling prefix.
• Static tilts that flip sign between tape halves are regime bets, not edges. A conditioning scheme must beat baseline on BOTH halves and on each individual day; pre-register the rules, no per-day fitting.
• Mirror/complement markets are the SAME bet — deduplicate before computing independence-based probabilities.

Live execution rules (the incident ledger)

• Smoke-test every live order path with a real 1-lot round trip (place, verify resting, cancel, verify gone) before any strategy trades through it. Paper cannot exercise venue routing parameters. Cost: $0.01.
• Cancels must be truthful: distinguish cancelled / already-gone (filled!) / failed. A 404 is not a cancel — it means wrong routing shard or a fill-race. Swallowed cancel failures + re-anchor loops = runaway position stacking (cost us $84 in one evening).
• Route the exchange shard (exchange_index) consistently on EVERY call — placement, reads, cancels, collateral. Collateral must be pre-positioned on the shard before orders.
• Measure rate limits empirically with dust orders (burst ladder + sustained test). Advertised budgets can be 10× off effective per-order cost. Pace bursts to the measured rate with retry-requeue on 429.
• post_only for any maker strategy — an "aggressive maker" price on a market trading inside your band silently becomes a taker fill of a different (unvalidated) trade. Tag and cohort maker vs taker fills separately in all realized-EV accounting.
• Rails, always: halt-file checked before every order batch (cancel-all on appearance), independent scheduler-driven drawdown breaker with deposit-jump detection, per-order and per-market size caps with anomaly ledger events, equity floor stand-down. The breaker baseline is set at session start per the operator's stated loss tolerance.
• Queue priority is the product in penny/maker books. Synchronized-open venues: NTP + fire at boundary+0 with a retry probe (measure the venue's open-transition latency). Event-driven venues: push-based discovery (lifecycle WebSocket channels) beats polling; a fresh market's book is empty, so first-to-rest owns the level for its lifetime.
• Restart daemons at activity boundaries; mid-round restarts orphan in-memory state (untracked orders, missed settlements).
• Geofencing is real: order-origin IP matters, categories differ (sports/elections vs financials), and enforcement can change overnight. Datacenter location is part of your compliance posture.

Scaling doctrine

• House-money ratchet: double size only when realized profit at the current step covers the next step's incremental risk AND round-count and EV-continuity gates pass. Drop back a step on trailing degradation.
• Measure the capacity curve (EV/$ by per-print cap) before believing any scale projection; then the binding constraint is capture rate × measured pool, and capture rate is only measurable live.
• Fill quality is the untested link at every new size — the paper/live EV-per-$ ratio at each step gates the next.

Experiment operations

• Cohort every fill (maker/taker, band, source) at write time — post-hoc attribution is what lets a bug contaminate a verdict.
• Ship one variable at a time; changing size and placement logic in the same window destroys attribution.
• Run A/B variants in paper against a frozen baseline (small-footprint variant, shared rounds, EV/premium-$ as the scale-free judge).
• Log everything to append-only ledgers; reconcile realized PnL to the venue's own fills/settlements, never to internal marks.
• Post results (wins AND falsifications) to the operator's channel; write a decision record with evidence and reversal conditions for every ship, retire, and kill.