Sibyl — Market Oracle

Research Pipeline

Phase 2 Deep Count 3–50

Event markets getting full AI deep-dive per scan (sleeve slots are additive)

Phase 1 Pool Size 50–500

Candidates entering Haiku quick triage

Max Tool Calls 2–10

Claude tool-use budget per market

Min Edge 1–30%

Minimum mispricing gap to surface (percentage points)

Parallel Workers 1–6

Concurrent Phase 2 research workers

Dynamic Targeting

Enabled

Auto-adjust Phase 2 count based on cash & portfolio size

Batch Mode (50% off)

Disabled

Use Anthropic Batch API for Phase 1 triage and Phase 2 research — 50% cheaper, but scans take longer while batches process (~2-5 min per research round). Best for overnight/scheduled scans.

Phase 1 Provider

LLM provider for Phase 1 triage. Sonnet re-score always uses Claude. Gemini Flash is free (15 RPM / 1,500 RPD). DeepSeek-V3 is very cheap (~$0.28/M input). Both require an API key below.

Phase 2 Provider

Phase 2 deep research provider. "Hybrid" modes use a cheaper model for tool-call exploration then Opus for the final verdict. DeepSeek hybrid is fast and cheap; Gemini hybrid is free (100 RPD). Both require an API key below.

Phase 1.75: Opus Batch Triage

Enabled

Opus evaluates the FULL market pool (~150 markets) with lean context — headlines, scores, criteria, cross-platform refs. ~$1.50 total. Gates Phase 2 by identifying which markets actually need $9 of deep research vs. which already have a clear verdict.

Short-DTE Sleeve Slots 0–6

Bonus Phase 2 slots for markets closing within threshold, added on top of deep count (0 = disabled)

Short-DTE Threshold (days) 2–14

Markets closing within this many days qualify for the short-DTE sleeve

Price Market Sleeve Slots 0–6

Bonus Phase 2 slots for BTC/ETH/SPX/GLD/OIL price markets, added on top of deep count (0 = disabled)

Verdict Suppress TTL 4–48h

Hours before FAIR/UNCERTAIN verdicts are re-researched (lower = fresher but costlier)

Verdict Suppress Price Δ 1–10pp

Re-research suppressed market if price moved more than this many pp

Sonnet Rescore Top N 10–100

Markets getting Sonnet + full articles re-score after Haiku triage (higher = more accurate ranking, costs ~$0.003/market)

Min Confidence % 50–90

Phase 2 confidence threshold — results below this are filtered out. Lower = more opportunities but noisier signals

Caching

Phase 1 TTL 1–48h

Hours before triage scores expire

Phase 2 TTL 1–48h

Hours before research verdicts expire

Price Tolerance 1–20%

Invalidate cache if price moves more than this (pp)

Market Filters

Min Liquidity ($) 0–50k

Skip markets with less liquidity

Min Volume ($) 0–10k

Skip markets with less 24h volume

Price Band Low 1–40%

Ignore markets below this price

Price Band High 60–99%

Ignore markets above this price

Portfolio

Auto-Refresh (sec) 60–3600

Background price refresh interval

Autonomy

These settings control autonomous behavior that can spend API credits and modify positions without manual intervention. Review carefully before enabling.

Auto-Research on Material Moves

Disabled

⚠ When enabled, Sibyl will spend Anthropic API credits automatically when positions experience significant price moves. Default: OFF.

Auto-Research Daily Cap 1–20

Maximum auto-research triggers per day (each costs ~$0.10–0.50 in API calls)

Auto-Research Cooldown 0.5–12h

Hours before same position can trigger auto-research again

Material Move Threshold 3–25%

Price move % to flag as material (triggers alerts and monitoring)

Thesis Challenge Threshold 2–15pp

Adverse price move (pp) to trigger thesis-challenge alert for high-confidence positions

Tiered Position Monitoring

Enabled

Adaptive refresh: 2min for urgent positions, 6min for medium, 10min for stable. When disabled, all positions refresh every 10min.

Calibration Feedback

Enabled

Apply historical prediction bias corrections in the optimizer. Requires ≥10 resolved outcomes per domain.

Calibration Min Samples 5–50

Minimum resolved outcomes per domain before applying bias correction

Order Execution (Kalshi)

When enabled, Sibyl will automatically place orders on Kalshi based on optimizer recommendations. Dry-Run mode logs intended trades without placing real orders.

Auto-Execute

Master switch — enables auto order placement

Dry-Run Mode

Log trades without placing real orders

Max Orders/Scan 1–10

Cap auto-placed orders per scan or rebalance

Stale Timeout (min) 5–60

Cancel/replace unfilled limit orders after this

Max Slippage (¢) 1–5

Max cents above ask (buy) or below bid (sell)

Min EV % 5–30

Only execute if annualized EV exceeds this

Cooldown (min) 15–240

Minutes before same ticker can be auto-executed again

Daily Cap ($) 50–1000

Max capital deployed by auto-exec per day

Stop-Loss Monitor

Checks positions every 2 min between scans. Triggers de-risk trades when unrealized loss exceeds threshold. Respects dry-run mode — won't place real orders unless auto-execution is live.

Enable Stop-Loss Monitor

Loss Threshold (%) 10–50

Trigger stop-loss at this % unrealized loss

Action

Min Hold (hours) 1–24

Don't stop-loss within N hours of entry (avoids whipsaws)

Take-Profit Monitor

Automatically closes positions when they hit their take-profit target. Uses limit orders at bid price (not aggressive market sweeps) to capture spread. Checks every 2 min alongside the stop-loss monitor. Independent of WebSocket — works with REST polling too.

Enable Take-Profit Monitor

Action

Min Hold (hours) 1–24

Don't take profit within N hours of entry (avoids flash spikes)

Auto Kalshi Sync

Periodically syncs positions with Kalshi's portfolio API. Catches manual trades, partial fills, and resolution events without requiring you to click Sync.

Enable Auto Sync

Interval (min) 5–60

Sync with Kalshi every N minutes

Real-Time Prices (WebSocket)

Opens a persistent connection to Kalshi for instant price updates instead of polling every 2 minutes. When combined with stop-loss or take-profit monitors, they react in milliseconds instead of minutes. This only controls the price feed — stop-loss and take-profit are separate toggles above.

Enable WebSocket Streaming

Keep REST Polling as Backup

Continue 2-min REST checks for positions without WebSocket data

Polymarket Real-Time Prices

Streams live orderbook data from Polymarket's public CLOB WebSocket. Combined with Kalshi prices, enables continuous cross-platform arbitrage detection. No authentication required.

Enable Polymarket WS

Arb Check Interval (sec)

How often to run cross-market arb detection (10–600s)

Scan Scheduler

Automatically runs tiered scans at configurable intervals. Light scans refresh prices (free), medium scans research drifted positions (selective AI), full scans run the complete pipeline.

Scheduler Enabled

Master switch for automatic tiered scanning

Light Interval 5–120 min

Minutes between light scans (price refresh + rebalance, no AI)

Medium Interval 30–480 min

Minutes between medium scans (selective AI on drifted positions)

Full Interval 60–720 min

Minutes between full scans (complete pipeline)

Active Hours start–end

to

Only schedule scans during these hours (local time)

Drift Threshold 3–20 pp

Price move (pp) from cached research to trigger medium-scan re-research

DTE Threshold 1–30 days

Positions expiring within this many days get re-researched if stale

Batch for Scheduled

Use batch API (50% off) for all auto-scheduled scans

Daily Scheduled Scans

Pin full scans to specific times of day. These fire regardless of interval timers. Add up to 4 daily scan times (local time, 24h format).

Theme

Dark

Light

Data Sources

Polymarket

Manifold

Metaculus

Kalshi

API Keys

Anthropic API Key (for AI verdicts)

Without a key: shows news headlines only. With a key: Claude assesses whether each market is mispriced.

Kalshi Key ID (the UUID from kalshi.com → Settings → API)

Kalshi RSA Private Key (paste full PEM including header/footer lines)

Both required for Kalshi.

Metaculus Token

Metaculus requires a free auth token.

Congress.gov API Key (bills, nominations, confirmation votes)

Enables real-time bill tracking, nominations, and vote records in AI research.

FEC API Key (campaign finance & lobbying — works without a key)

Uses OpenFEC + Senate LDA to reveal who funds swing-vote senators. Works out of the box with DEMO_KEY.

FRED API Key (Federal Reserve economic data — CPI, GDP, unemployment)

Uses the official FRED API instead of the public JSON endpoint — much more reliable, eliminates timeout errors.

Brave Search API Key (higher quality web search for AI research)

Replaces DuckDuckGo for Phase 2 web searches — better results, more reliable. Without it, falls back to DDG.

Finnhub API Key (structured financial data for research)

Powers economic calendar, SEC 8-K filings, earnings data, and insider transactions for Phase 2 research. Without it, falls back to web search.

The Odds API Key (sharp sportsbook lines — Pinnacle, DraftKings, FanDuel)

Sharp book vig-free implied probabilities for sports markets — the gold standard for calibrating Kalshi sports pricing.

Google AI Studio API Key (free Gemini for Phase 1 triage + Phase 2 hybrid)

Required for Gemini Flash (Phase 1) and Gemini Pro (Phase 2 hybrid). Free tier: Flash 1,500 RPD, Pro 100 RPD.

DeepSeek API Key (DeepSeek-V3 for Phase 1 triage + Phase 2 hybrid)

DeepSeek-V3 as Phase 1 triage model or Phase 2 exploration engine (Opus still gives the final verdict).

SETTINGS

PERFORMANCE