signal-vs-noise（信号与噪声）

SKILL: signal-vs-noise

Purpose

Filter relevant information from noise while preserving evidence and decision-impact.

When to Use

• Many messages/news/items arrive at once
• A decision must be made and inputs are overwhelming
• You need a ranked list of what matters

Inputs

• dataset (required): list of items (news, messages, metrics, notes)
• decision_context (optional): what decision this supports
• time_window (optional): timeframe considered relevant

Steps

1. Normalize the dataset into items with source, timestamp (if present), and content.
2. Extract key claims per item (1–3 claims max).
3. Remove redundancy:

- merge duplicates - group near-duplicates by same claim

1. Identify high-impact signals:

- changes in constraints (governance, deadlines, outages) - verified facts that shift probability - actionable next steps

1. Rank signals by:

- impact on the decision - credibility/verifiability - urgency (only if real)

1. Output:

- ranked signals with evidence - discarded noise (with brief reason)

Validation

• No duplicated signals in the ranked list.
• Each signal includes at least one evidence pointer (source/item id).
• Novelty is not treated as importance by default.

Output

• ranked_signals: ordered list with claim, why_it_matters, evidence
• discarded_noise: list with item + reason

Safety Rules

• Avoid bias toward novelty: “new” is not automatically “important”.
• Do not delete dissent; label it as low-confidence when evidence is weak.

Example

Input: 30 chat messages + 5 news headlines about a protocol. Output: top 5 signals (governance vote date, confirmed exploit, liquidity change) + noise bucket (memes, repeated hype).