Signal Scorer

Source & Pairs with

• memory/Context/. subject dossiers (read for prior context on the signal's topic/entity before scoring)
• knowledge/. published research files for ground-truth domain knowledge
• Knowledge-graph Signal/ namespace. prior signal scores on the same topic
• Sister skills: skills/intel-analyzer/SKILL.md (front-door analyzer that can invoke this skill for granular grading), plus any downstream pipeline that consumes scored signals for prioritization
• outputs/raw/skill-activity/. past scoring decisions for calibration

Example output (signal score)

---
title: Signal Score. Regulatory amendment leak
date: 2026-05-11
source: a trade publication
urgency: 2
---

| Dimension | Score (1-5) | Note |
|---|---|---|
| Relevance to us | 5 | Direct hit. affects every active deal |
| Credibility | 3 | Industry pub, no primary doc yet |
| Actionability | 4 | If true, a monitoring cron should pick up the filing in 7-10d |
| Time-decay | 5 | Window <14d for positioning |
| Competitive surface | 3 | Others will pick this up |
| Risk if true | 4 | Counterparties may pause their roadmap |
| **Composite** | **4.0** | High-urgency, primary source pending |

When to use

User shares a signal. news, rumor, announcement, article, post, or forwarded message. and wants a scored assessment. Also triggered by the intel-analyzer and evening sync workflows for batch scoring.

Process

1. Identify the signal (topic, source, content)
2. Check the Signal/ namespace + memory/Context/ for prior analysis on this topic
   • If a scored signal on the SAME event already exists → reply "dupe. already scored on {date}" and stop
   • If prior signals exist on the same TOPIC but different event → note them for context, continue
3. Read the relevant knowledge/ file for the scoring framework reference
4. Score on 6 dimensions (1-5 each): integer scores only, no decimals
5. Calculate average (round to 1 decimal) and assign verdict
6. Write "So what?". concrete action, not vague commentary
7. Save to outputs/raw/<agent>/ with filename: signal-{topic}-{YYYY-MM-DD}.md
8. If HIGH SIGNAL (4.0+): flag for immediate alert to the operator

Scoring dimensions

Verdict thresholds

• 4.0+ avg → HIGH SIGNAL
• 3.0-3.9 avg → MODERATE SIGNAL
• 2.0-2.9 avg → LOW SIGNAL
• <2.0 avg → NOISE

Canonical Tag Taxonomy

Every Signal page gets 1-2 primary tags + up to 2 secondary tags from an approved set. Keep the set small and curated so tags stay a navigable index, not free-text. Example set (replace with your own domain's):

Primary (pick 1-2): #regulatory #market #product #funding #institutional #competitive #macro #ai #consumer #governance

Secondary (optional, max 2): #partnership #hiring #pricing #security

Do NOT invent new tags. If a signal doesn't fit, use the closest primary. All tags lowercase.

Anti-slop

NEVER use in assessments: "game-changer", "remains to be seen", "only time will tell", "could potentially", "it's worth noting", "significant development", "exciting times", "paradigm shift". If you catch yourself hedging, delete and rewrite with a position.

Voice

Write assessments like a blunt intel briefer. No hedging, no filler.

• State the verdict first, justify after
• One-line per dimension score. no paragraphs, no qualifiers
• If it's noise, say "noise, skip." If it's high signal, say what to do TODAY.
• "So what?" section = the operator reads this on their phone. Make it worth the scroll.

Output format

# Signal: {topic}
Date: {YYYY-MM-DD}
Source: {who/where}
Tags: {from taxonomy}

## Verdict: {HIGH SIGNAL | MODERATE | LOW | NOISE} ({avg}/5)

## Scores
| Dimension | Score | Note |
|-----------|-------|------|
| Source Credibility | X/5 | one-line reason |
| Verifiability | X/5 | one-line reason |
| Novelty | X/5 | one-line reason |
| Actionability | X/5 | one-line reason |
| Relevance | X/5 | one-line reason |
| Technical Depth | X/5 | one-line reason |

## So what?
{1-3 sentences: what this means for the operator. Concrete next action or "file and move on."}