Xpulse — Real-Time Social Signal Scanner

Overview

Xpulse is a real-time X/Twitter signal scanner optimized for prediction market traders. It detects market-relevant signals on configurable topics using a three-stage intelligence pipeline:

1. 1. Stage 1: Signal Detection — Scans X via DuckDuckGo (site:x.com queries), analyzes posts with local Qwen for tradeable signals
2. Stage 2: Materiality Gate — Compares against 48h signal history, filters out noise and repeat signals (fail-closed)
3. Stage 3: Position Matching — Only alerts on signals that match active Kalshi positions (keyword overlap, 2+ keywords required)

Key philosophy: Zero API costs (DuckDuckGo + local Qwen), fail-closed design (silence over noise), position-aware filtering (no irrelevant alerts).

When to Use This Skill

• - You're actively trading prediction markets on Kalshi
• You want real-time social signals for your active positions
• You prefer local LLM analysis over expensive API calls
• You need to prevent notification fatigue from generic social signals
• You want signals cached for morning brief consumption

Requirements

Python & Dependencies

• - Python 3.10 or higher
• Required packages:

CODEBLOCK0

Local Infrastructure

• - Ollama (https://ollama.ai) with Qwen model

- Install from https://ollama.ai (macOS, Linux, Windows) - Then: ollama pull qwen3:latest - Must be running: ollama serve (background process)

Kalshi API (for position matching)

• - Free Kalshi account with API credentials (https://kalshi.com)

- API key ID - Private key file (ES256 format) - Cost: Free (read-only access to positions)

Configuration

Create or update your config.yaml:

CODEBLOCK1

Topics Configuration

INLINECODE3 list should contain:

• - Market-relevant keywords you're interested in
• DuckDuckGo will search: INLINECODE4
• Examples: "Trump tariff", "inflation CPI", "Fed rate", "crypto SEC", "presidential election"

Thresholds

Parameter	Default	Meaning
INLINECODE5	0.7	Stage 1: Only keep signals with ≥70% AI confidence of being tradeable
INLINECODE6

30 | Minimum minutes between scans (prevent API hammering) | | materiality_gate | true | Stage 2: Enable/disable deduplication filter | | position_gate | true | Stage 3: Enable/disable Kalshi position matching | | max_history_entries | 200 | Max entries in 48h signal history (auto-trimmed) |

The Three-Stage Pipeline

Stage 1: Signal Detection

Input: List of topics (e.g., ["tariff", "inflation", "Ukraine"])

Process:

1. 1. For each topic, search X/Twitter via DuckDuckGo: INLINECODE10
2. Fetch up to 5 recent posts
3. Feed posts to local Qwen with this prompt:

CODEBLOCK2

Thresholding: Only keep signals where:

• - INLINECODE11
• INLINECODE12 (default 0.7)

Output: List of candidate signals with direction + confidence

---

Stage 2: Materiality Gate (Fail-Closed)

Input: Candidate signals from Stage 1 + 48h signal history

Purpose: Prevent alert fatigue by filtering out:

• - Repeat signals about the same story (different wording)
• Ongoing background noise (tariffs in news for days)
• Speculation with no concrete new event
• Minor updates to previously alerted developments

Process:

1. 1. Load recently sent signal history (last 20 alerts, 48h window)
2. Build prompt with:

- Candidate new signals
- What the user already knows (recent alerts)

1. 3. Send to Qwen:

CODEBLOCK3

Fail-Closed Design:

• - If Qwen fails (timeout, error, unparseable): drop all signals
• This prevents spam of uncertain signals when the filter breaks
• Better to miss a signal than spam the user

Output: Filtered list (subset of Stage 1 candidates)

---

Stage 3: Position Matching Gate

Input: Filtered signals from Stage 2 + active Kalshi positions

Purpose: Only alert the user about signals that match their active positions

Process:

1. 1. Fetch active Kalshi positions (unsettled)
2. Extract keywords from ticker + market title (>2 char words)

- Example: "POTUS-2028-DEM" → keywords: {potus, 2028, dem}

1. 3. For each signal, extract keywords from topic + summary
2. Match if 2+ keyword overlap

Example:
CODEBLOCK4

Suppressed Signals:

• - Signals without matching positions are logged silently
• Available in cache (.x_signal_cache.json) for morning brief
• Not sent to user

Output: Critical signals matched to positions → iMessage alert

---

Signal Caching

All signals (pre-filter) are cached in .x_signal_cache.json:

CODEBLOCK5

Used for:

• - Morning brief consumption (all signals, even suppressed ones)
• On-demand signal review (what was detected but not alerted)
• Debugging signal pipeline

---

Running Xpulse

Command Line

CODEBLOCK6

Scheduled (Every 30 Minutes)

Via OpenClaw config:
CODEBLOCK7

Via crontab:
CODEBLOCK8

Via launchd (macOS):

<!-- ~/.launchd/com.xpulse.scanner.plist -->
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
  <key>Label</key>
  <string>com.xpulse.scanner</string>
  <key>ProgramArguments</key>
  <array>
    <string>/usr/bin/python3</string>
    <string>/path/to/xpulse.py</string>
  </array>
  <key>StartInterval</key>
  <integer>1800</integer>
  <key>StandardOutPath</key>
  <string>/tmp/xpulse.log</string>
  <key>StandardErrorPath</key>
  <string>/tmp/xpulse.log</string>
</dict>
</plist>

---

Example Output

Alert to iMessage (Stage 3 Pass)

CODEBLOCK10

Signal Cache (Stage 1 → Morning Brief)

CODEBLOCK11

---

Materiality Gate — Technical Details

See references/materiality-gate.md for detailed explanation of:

• - How the 48h rolling window works
• Example filtering scenarios (signal vs noise)
• Fail-closed behavior

---

Position Matching — Technical Details

See references/position-matching.md for:

• - Keyword extraction algorithm (>2 char threshold)
• Overlap matching logic (2+ keyword requirement)
• Example position-signal pairs

---

Troubleshooting

Ollama Not Running

Error: Connection refused or INLINECODE18

Fix:
CODEBLOCK12

Qwen Model Not Found

Error: ollama run qwen:latest fails

Fix:
CODEBLOCK13

DuckDuckGo Returning No Results

Likely causes:

• - Topic is too specific or niche
• No recent posts on X about that topic
• DuckDuckGo rate limiting (try again later)

Debug:
CODEBLOCK14

No Kalshi Positions (Stage 3 Suppresses All)

Expected behavior: If you have no active Kalshi positions, all signals are suppressed (logged silently). This is intentional — position-aware alerting is the goal.

To test: Either:

1. 1. Create a test Kalshi position matching your topic
2. Disable Stage 3: set position_gate: false in config (alerts all non-materiality signals)

Signal History Growing Too Large

Automatic: History is capped at 200 entries (oldest trimmed first)

Manual reset:

rm ~/.openclaw/state/x_signal_history.json

---

Qwen Prompts (Core IP)

Stage 1 Prompt: Signal Detection

CODEBLOCK16

Stage 2 Prompt: Materiality Gate

CODEBLOCK17

---

OpenClaw Ecosystem Integration

Social signal intelligence layer for the Prediction Market Trading Stack.

Connected Skill	How It Connects
Market Morning Brief	X signals appear in your daily morning digest
Kalshalyst

Social signals validate or challenge contrarian edges |
| Prediction Market Arbiter | Correlate social sentiment with cross-platform divergences |
| Kalshi Command Center | Act on position-matched signals via trade execution |

Install the complete stack:
CODEBLOCK18

Implementation Notes

Battle-tested in production trading environments. Design principles:

1. 1. Standalone implementation — zero external dependencies beyond listed packages
2. Direct alerts — sends signals straight to you via your OpenClaw agent
3. All thresholds, filters, and prompts — refined through live social signal monitoring
4. Fail-closed design — no signal is better than a false positive
5. Scripts are standalone — works with any OpenClaw setup

---

Performance & Cost

Typical Run

• - Runtime: 1-2 minutes (5 topics, Qwen Stage 1+2)
• API calls:

- DuckDuckGo: 5 calls (1 per topic) - Kalshi: 1 call (fetch positions) - Ollama (local): 2 calls (Stage 1 + 2)

• - Cost: $0 (DuckDuckGo free, Ollama local, Kalshi free read-only)

Scaling

10 runs per day = ~10-20 minutes total runtime + zero API cost. Minimal resource usage.

---

Further Reading

• - references/materiality-gate.md — Signal deduplication + 48h window
• INLINECODE22 — Keyword overlap logic

---

API Reference

Main Function

CODEBLOCK19

State Management

CODEBLOCK20

---

Support & Iteration

Xpulse is designed for iteration:

1. 1. Topic Tuning: Add/remove topics as your trading focus changes
2. Confidence Threshold: Lower min_confidence to catch more signals (noisier), raise to reduce false positives
3. Materiality Gate: Disable if you're missing important signals, enable if too noisy
4. Position Gate: Disable position_gate: false if you want ALL signals (not recommended)
5. Qwen Prompts: Customize Stage 1/2 prompts for your trading style

---

License & Attribution

Author: KingMadeLLC

---

Feedback & Issues

Found a bug? Have a feature request? Want to share results?

• - GitHub Issues: github.com/kingmadellc/openclaw-prediction-stack/issues
• X/Twitter: @KingMadeLLC

Part of the OpenClaw Prediction Stack — the first prediction market skill suite on ClawHub.