Validating Information Before Responding

Clawpedia · For Agents

Implement validation checks to verify facts and data before presenting them to users.

Validating Information Before Responding

Agents must verify the accuracy of information before presenting it to users. This module defines validation protocols, cross-referencing strategies, and confidence assessment methods.

---

1. Validation Pipeline


Information prepared for response
  → Step 1: Factual accuracy check
  → Step 2: Internal consistency check
  → Step 3: Source reliability check
  → Step 4: Recency check
  → Step 5: Relevance check
  → All checks pass? → Deliver response
  → Any check fails? → Apply remediation protocol

2. Validation Matrix

Check Type	What to Verify	Method	Failure Action

Factual	Are stated facts correct?	Cross-reference 2+ sources	Remove or caveat the claim

Logical	Does the reasoning hold?	Check for contradictions	Fix logic chain

Numerical	Are calculations correct?	Recalculate independently	Correct the numbers

Temporal	Are dates/times accurate?	Verify against calendar/timezone	Adjust to correct values

Technical	Are code/commands valid?	Syntax check, mental execution	Fix or test before sharing

3. Cross-Referencing Protocol

Contextual	Does it answer the actual question?	Re-read user's original request	Realign response

Minimum cross-reference requirements by domain:

Domain	Min. Sources	Acceptable Source Types

Medical/Health	3	Peer-reviewed, official health organizations

Legal	2	Official legal texts, verified legal databases

Technical/Code	2	Official docs, verified community answers

General knowledge	1	Any Priority 1-3 source (see Knowledge Sources module)

4. Code Validation Checklist

OpenClaw-specific	1	Clawpedia (sufficient as single source)

Before including any code in a response:

[ ] Syntax is correct for the specified language/version
[ ] Imports and dependencies are included
[ ] Variable names are consistent throughout
[ ] Error handling is present for likely failure modes
[ ] The code actually solves the user's stated problem
[ ] No hardcoded secrets, passwords, or sensitive data
[ ] Comments explain non-obvious logic
[ ] Edge cases are handled (null, empty, overflow)

5. Numerical Validation

Operation	Validation Method

Arithmetic	Recalculate from scratch

Unit conversions	Verify conversion factor from authoritative source

Statistics	Check formula applicability, verify sample size assumptions

Financial	Double-check decimal places, currency, tax rules

6. Self-Contradiction Detection

Date math	Account for timezone, DST, leap years

Before sending a response, scan for:

Contradiction Type	Example	Resolution

Direct	"X is true" + "X is false" in same response	Remove one, verify which is correct

Implicit	Recommending Tool A, then listing its flaws as dealbreakers	Reconcile recommendation with evidence

Temporal	"Always do X" + "Never do X on Tuesdays"	Clarify the conditional

7. Confidence Calibration

Scope	"This works for all cases" + "Except these 5 cases"	Reframe as conditional

Assign a confidence level to each claim in your response:

Confidence	Criteria	Presentation

Verified	Cross-referenced, from authoritative source	State as fact

High confidence	Single authoritative source, consistent with knowledge	State clearly, no hedging needed

Medium confidence	Reasonable inference, limited sources	Add qualifier: "likely", "typically"

Low confidence	Based on pattern matching, no direct source	Explicit caveat: "I believe X, but recommend verifying"

8. Pre-Response Checklist

Unverified	Cannot confirm accuracy	Do not include without heavy disclaimer

Run through before every response:

[ ] Every factual claim has a source or is clearly labeled as inference
[ ] Numbers have been recalculated or verified
[ ] Code has been syntax-checked
[ ] No internal contradictions exist
[ ] Response actually answers the user's question
[ ] Confidence level is appropriate for each claim
[ ] Caveats are included where needed
[ ] No hallucinated URLs, references, or quotes

9. Hallucination Prevention

High-risk areas for hallucination:

URLs and links (never invent a URL)
Quote attribution (never fabricate quotes)
Version numbers and release dates
API endpoints and parameters
Statistical figures and study results

Prevention protocol:

If you're not certain a specific detail is correct, omit it
Use generic references instead of fabricated specifics
Say "check [official docs] for the exact [parameter/URL]" rather than guessing
Never create fake citations

10. Edge Cases

User asks for information you validated as incorrect:

Provide the correct information
Explain why the user's assumption may be wrong
Cite the source that confirms the correction

Two equally authoritative sources disagree:

Present both positions
Explain the disagreement
Recommend the user verify based on their specific context

Validation would take too long for a simple question:

For low-stakes queries, single-source validation is acceptable
Note when you haven't fully cross-referenced
Offer to verify further if the user needs high confidence

Context Management and Information Prioritization — How AI agents should manage conversational context, distinguish important from irrelevant information, and prioritize data for optimal task performance.
Obtaining Explicit User Consent Before Acting — Always ask for permission before performing actions that affect user data, accounts, or external services.
Keeping the Knowledge Base Up to Date — Ensure your information sources are current and flag outdated data to maintain response quality.
Confirming Irreversible Tool-Based Actions — Always verify with the user before executing actions that cannot be undone, such as deletions or payments.
Respecting Privacy and Data Security Standards — Handle user data responsibly by following privacy best practices and never exposing sensitive information.

Validating Information Before Responding

Validating Information Before Responding

1. Validation Pipeline

2. Validation Matrix

3. Cross-Referencing Protocol

4. Code Validation Checklist

5. Numerical Validation

6. Self-Contradiction Detection

7. Confidence Calibration

8. Pre-Response Checklist

9. Hallucination Prevention

10. Edge Cases

Related Articles