a structured starting point that makes your own verification faster and more targeted
You give the AI a topic, and it produces a structured list of claims with confidence ratings, source suggestions, and flags for anything that looks contradictory or uncertain.
Fact Verification Foundation
Tags: fact-checking, source-citation, hallucination-prevention, verification, credibility, research
TL;DR
What This Recipe Does
It helps you check whether facts are reliable before you use them. You give the AI a topic, and it produces a structured list of claims with confidence ratings, source suggestions, and flags for anything that looks contradictory or uncertain. The output is designed as a verification checklist — the AI does the organizing and the initial confidence assessment, and you do the final confirmation.
Why It Exists
AI is good at surfacing information, but it can also present made-up facts with total confidence. This recipe flips that problem on its head. Instead of just giving you facts and hoping they are right, it forces the AI to tell you how confident it actually is about each claim, where it thinks the information came from, and where you should double-check. The result is not a list of verified facts — it is a structured starting point that makes your own verification faster and more targeted.
An Important Thing to Understand
The AI does not actually look things up in real-time. It is not searching databases or checking journal archives while you wait. It is assessing claims based on what it learned during training. That means its source suggestions are starting points, not confirmed citations. Its confidence ratings reflect how well-supported a claim is in its training data, not whether it checked today. This recipe is honest about that limitation — and that honesty is what makes it useful.
What You Give It
A topic you need facts about, a verification level (“basic” for a quick confidence check with source suggestions, or “systematic” for full ratings, verification guidance, and contradiction detection), and optionally what the facts are for (business plan, academic paper, presentation) so the AI can calibrate its rigor, plus your preferred output format (list, table, or narrative).
What You Get Back
For each factual claim: the claim itself stated clearly, a confidence rating (HIGH, MEDIUM, or LOW) based on how well-supported it is in the AI’s training data, source suggestions framed as starting points for you to verify, verification guidance in systematic mode, and contradiction flags if the AI’s training data contains conflicting information. At the end, a verification summary tells you how many claims fell into each confidence tier and gives you a prioritized list of what to verify first.
What Makes This Recipe Different
It is honest about what AI can and cannot do — every session starts with a clear disclosure. It separates the claim from the source so you can evaluate each independently. It prioritizes your effort by telling you exactly which claims need your attention most. It flags contradictions even if you did not ask for them. And it adjusts rigor to your use case — facts for a casual blog post need a different confidence threshold than facts for a legal filing.
Best Used For
Business plans and pitch decks when you need solid market stats but are short on research time. Content creation when you want to catch errors before publishing. Presentations when you need a quick confidence check on key data points. Market research when structuring a large number of claims for systematic verification. Technical documentation when flagging which specs and figures need independent confirmation.
Tips for Best Results
Be specific about your topic — “US residential solar panel installation costs 2020–2024” gives the AI a focused target with verifiable claims. Tell it what the facts are for so the AI can calibrate its rigor. Start with basic mode for a quick confidence check; use systematic mode when you need a full verification trail. Always verify the sources independently — the AI’s source suggestions are leads, not citations. Focus your verification effort on MEDIUM and LOW claims first; HIGH claims are likely accurate but spot-check them for anything critical.
How To Start
STEP 1Define Your Topic and Verification Level
Identify the specific topic requiring fact verification. Decide your verification depth: basic for a quick confidence check with source suggestions, or systematic for full confidence ratings, verification guidance, and contradiction detection. You can tune the output with optional parameters that shape rigor, format, and scope.
Available parameters
-
topic
· string · required
The subject requiring verification. -
verification_level
· string · optional · default “basic”
Options: basic (confidence check with sources) or systematic (full ratings, verification guidance, contradiction detection). -
include_confidence_ratings
· boolean · optional · default false
Add HIGH/MEDIUM/LOW rating per claim. -
document_verification_guidance
· boolean · optional · default false
Include guidance on how to verify each claim independently. -
flag_contradictions
· boolean · optional · default true
Note conflicting information in training data. The AI will flag contradictions regardless of this setting when safety-critical conflicts are detected. -
output_format
· string · optional · default “list”
Options: list, table, or narrative. How to present assessed claims. -
use_context
· string · optional
What the facts will be used for (business plan, academic paper, presentation, etc.). The AI calibrates its rigor based on your intended use.
Example invocations
Basic (topic only)
#H->AI::Directive: (Run the Fact Verification Foundation recipe for topic: farm-to-table dining trends in the US. Use context: business plan.)
Systematic (full protocol)
#H->AI::Directive: (Run the Fact Verification Foundation with verification_level=systematic and output_format=table for topic: environmental consulting market growth. Use context: client presentation.)
STEP 2Request AI-Assisted Fact Assessment
Use the prompt template to request facts about your topic. The AI will present claims it can support from its training data, tag each with a confidence level, and surface known sources as starting points for your own verification.
STEP 3Review Source Suggestions
The AI will suggest sources it associates with each claim from its training data. These are starting points, not confirmed citations. Verify that suggested sources exist, are current, and actually support the claim. Reputable sources include peer-reviewed journals, government databases, industry reports, and established news organizations.
STEP 4Use Confidence Ratings to Prioritize Verification
For systematic verification, the AI provides confidence ratings (HIGH, MEDIUM, LOW) for each claim. Focus your independent verification effort on MEDIUM and LOW claims first — HIGH claims are likely accurate but should still be spot-checked for critical content.
STEP 5Review Contradiction Flags
If the AI’s training data contains conflicting information, it will flag the discrepancy and present both perspectives. These contradictions are your highest-priority items for independent verification.
STEP 6Complete Human Verification
Use the AI’s structured output as a verification checklist. Confirm sources exist and are current. Check that facts are accurately represented. Verify any claim tagged MEDIUM or LOW before including it in final content.
STEP 7Integrate Verified Facts
Include only facts that meet your verification threshold in your final content. Replace AI-suggested source references with your own confirmed citations.
Usage Examples
Two scenarios showing how the Fact Verification Foundation works at different verification levels. The AI structures claims, rates confidence, suggests sources, and hands you a prioritized verification checklist.
Example 1 · Basic verification for a business plan
User:#H->AI::Directive: (Run the Fact Verification Foundation for topic: farm-to-table dining trends in the US. Use context: business plan.)
AI:I will assess these claims based on my training data. I am not accessing live databases or searching the internet. My source suggestions are starting points for your own verification. Here are the key claims I can assess for farm-to-table dining trends…
AI:Claim 1 (HIGH confidence): The farm-to-table movement has grown significantly in the US restaurant industry since the mid-2000s. This claim is commonly associated with National Restaurant Association trend reports — verify this source exists and supports this specific claim.
[Assessment continues with additional claims, source suggestions, and a verification summary]
Example 2 · Systematic verification for a client presentation
User:#H->AI::Directive: (Run the Fact Verification Foundation with verification_level=systematic and output_format=table for topic: environmental consulting market growth. Use context: client presentation.)
AI:Important disclosure: I will assess these claims based on my training data. I am not accessing live databases, searching the internet, or cross-referencing publications in real-time. Claims rated HIGH are well-supported in my training data but should still be independently verified for presentation-level content.
AI:Here is the systematic verification table with confidence ratings, source suggestions, and verification guidance for each claim…
[Full systematic output with table format, contradiction flags, and priority verification list]
How AI Reads This Recipe
When this recipe is triggered, the AI assists with fact verification by structuring claims, assessing confidence, and surfacing sources as starting points for human verification. The AI should:
- CONFIRM the topic, verification level, output format, and intended use with the user before producing any output.
- DISCLOSE clearly at the start that it is assessing claims based on training data, not performing live fact-checking.
- PRESENT each claim clearly, then confidence assessment, then source suggestions — separated so the user can verify each independently.
- RATE confidence honestly: HIGH means specific authoritative sources across multiple independent types; MEDIUM means reasonable basis but no pinpointable sources; LOW means limited basis or frequently changing data.
- NEVER fabricate source names, publication dates, journal names, author names, or DOI numbers. If uncertain, suggest a source category instead.
- FLAG contradictions regardless of the flag_contradictions parameter setting — conflicting information is safety-critical.
- CALIBRATE rigor to the stated use context (academic paper demands maximum conservatism; casual content uses standard rigor).
- STATE explicitly when it has no basis for assessing a claim, rather than stretching thin knowledge into a confidence rating.
The AI should NOT present its source suggestions as confirmed citations, rate everything MEDIUM as a safe hedge, or let the user’s apparent confidence in a claim influence the rating. The value of this recipe is in the structure and the honesty — not in pretending the AI is a fact-checking database.
When to Use This Recipe
Use this recipe when you:
- Need accurate, verifiable information for content creation, presentations, reports, or any context where factual errors could damage credibility.
- Are doing market research and need to structure a large number of claims for systematic verification.
- Want a quick confidence check on key data points before a presentation or pitch.
- Are writing technical documentation and need to flag which specs and figures require independent confirmation.
- Want to catch factual errors before publishing any public-facing content.
Do not use this recipe when:
You need real-time fact-checking against live databases, current publications, or breaking news. This recipe assesses claims based on AI training data — it cannot verify information that postdates its training cutoff or access external sources during execution. For academic or legal use cases requiring verified citations, treat all AI output as requiring independent confirmation regardless of confidence rating.
Version History
Changes to this recipe over time. Most recent first.
v2.00b
2026-02-15
QA audit revision (P051f-H005). Fundamentally reframed: AI assists verification, does not perform it. Added behavioral rules R1–R8, I-DON’T-KNOW handling protocol, interactive parameter confirmation with mandatory AI limitations disclosure. Replaced source_count with verification_level. Added use_context parameter for rigor calibration. Added verification summary with claim distribution and priority verification list.
v1.00
2025-12-29
Initial release. Consolidated from four source prompts (ChatGPT V1/V2, Claude V1, Gemini V1). Basic fact assessment with source suggestions.
Run this recipe with your AI from a single attachment
Every section on this page, packaged as one .txt file your AI can consume directly. Download and attach.
THE ACTUAL RECIPE
RCP-000-000-012-FACT-VERIFICATION-FOUNDATION
Structure factual claims for verification using AI-assistedsource surfacing, confidence assessment, and contradictiondetection. This recipe reduces AI hallucination by requiringthe AI to distinguish between well-supported claims anduncertain ones, surface known sources as starting pointsfor human verification, and flag where independent checkingis essential. Adapts from basic claim assessment tosystematic verification protocol based on your needs.
The CRAFT Recipe
# ===========================================================
# RECIPE-ID: RCP-000-000-012-FACT-VERIFICATION-FOUNDATION
# =========================================================== FACT_VERIFICATION_FOUNDATION = Recipe(
recipe_id=”RCP-000-000-012-FACT-VERIFICATION-FOUNDATION”,
title=”Fact Verification Foundation”,
description=”AI-assisted fact verification through
confidence assessment, source surfacing, and
structured output for human verification.”,
category=”CAT-000″,
subcategory=”Standalone”,
difficulty=”easy”,
version=”2.00b”, parameters={
“topic”: {
“type”: “string”,
“required”: True,
“default”: None,
“description”: “Subject requiring verification”
},
“verification_level”: {
“type”: “string”,
“required”: False,
“default”: “basic”,
“options”: [“basic”, “systematic”],
“description”: “Basic = confidence check with sources. Systematic = full ratings, verification guidance, contradiction detection.”
},
“include_confidence_ratings”: {
“type”: “boolean”,
“required”: False,
“default”: False,
“description”: “Add HIGH/MEDIUM/LOW rating per claim”
},
“document_verification_guidance”: {
“type”: “boolean”,
“required”: False,
“default”: False,
“description”: “Include guidance on how to verify each claim independently”
},
“flag_contradictions”: {
“type”: “boolean”,
“required”: False,
“default”: True,
“description”: “Note conflicting information in training data”
},
“output_format”: {
“type”: “string”,
“required”: False,
“default”: “list”,
“options”: [“list”, “table”, “narrative”],
“description”: “How to present assessed claims”
},
“use_context”: {
“type”: “string”,
“required”: False,
“default”: None,
“description”: “What the facts will be used for (business plan, academic paper, presentation, etc.)”
}
}, prompt_template=”””
# ===========================================================
# FACT VERIFICATION FOUNDATION
# =========================================================== # ———————————————————–
# BEHAVIORAL RULES
# ———————————————————– R1: NEVER present AI-generated source names as confirmed
citations. You are recalling source associations from
training data, not performing live lookups. Every
source you name is a SUGGESTION for the user to
verify, not a confirmed reference. Frame all source
mentions as: “This claim is commonly associated with
[source name] — verify this source exists and
supports this specific claim.” R2: NEVER fabricate source names, publication dates,
journal names, author names, or DOI numbers. If you
are not confident a specific source exists, say
“this type of claim is typically found in [source
category, e.g., FDA databases, peer-reviewed
nutrition journals]” rather than inventing a
specific reference. R3: ALWAYS be transparent about what you are actually
doing. You are assessing claims based on patterns
in your training data. You are NOT accessing live
databases, searching the internet, checking
publication archives, or cross-referencing sources
in real-time. State this clearly at the start of
every verification output so the user understands
the nature of your assistance. R4: ALWAYS confirm the topic, verification level, output
format, and intended use with the user before
producing the verification output. If the user
provides a use context (e.g., academic paper),
adjust your rigor threshold accordingly — academic
use requires more conservative confidence ratings
than casual content. R5: NEVER rate a claim as HIGH confidence simply because
it sounds plausible or because you have seen similar
claims frequently. HIGH means you have strong,
specific basis in your training data from multiple
authoritative source types. If a claim is widely
repeated but you cannot associate it with specific
authoritative sources, it is MEDIUM at best. R6: ALWAYS separate the claim from the source. Present
the factual claim first, then your confidence
assessment, then your source suggestions. This
structure makes it easy for the user to verify
each component independently. R7: NEVER skip contradiction flagging even if
flag_contradictions is set to False. If you
encounter genuinely conflicting information in
your training data, flag it regardless of the
parameter setting. Contradictions are safety-
critical information that the user needs. R8: IF a claim is outside your training data or you
have no basis for assessing it, say so explicitly.
“I have no basis for assessing this claim” is a
valid and valuable output. Do not stretch thin
knowledge into a confidence rating. # ———————————————————–
# I-DONT-KNOW HANDLING
# ———————————————————– This recipe is specifically about honesty and accuracy.
The AI must model the behavior it is helping the user
achieve: rigorous truthfulness. KNOWLEDGE LIMIT PROTOCOL: IF you cannot assess a claim at all:
STATE: “I have no basis for assessing this claim
from my training data.”
SUGGEST: specific source types or databases where
the user could verify independently
DO NOT assign a confidence rating to claims you
cannot assess IF you can partially assess a claim:
STATE what part you can assess and what part you
cannot
RATE only the part you can assess
FLAG the unassessed portion for human verification IF a claim involves recent events near or after your
training cutoff:
STATE that your training data may not include
the most current information
SUGGEST the user check current sources directly
DO NOT present potentially outdated information
as current fact IF the entire topic is outside your knowledge:
STATE this honestly at parameter confirmation
SUGGEST the user try specialized databases or
domain experts instead
DO NOT produce a verification output filled with
LOW-confidence claims and generic source categories # ———————————————————–
# PARAMETER CONFIRMATION
# ———————————————————– Before producing the verification output, confirm: VERIFICATION PARAMETERS:
Topic: {topic}
Verification Level: {verification_level}
Confidence Ratings: {include_confidence_ratings}
Verification Guidance: {document_verification_guidance}
Contradiction Flagging: {flag_contradictions}
Output Format: {output_format}
Intended Use: {use_context} IMPORTANT DISCLOSURE (include every time):
“I will assess these claims based on my training data.
I am not accessing live databases, searching the
internet, or cross-referencing publications in real-
time. My source suggestions are starting points for
your own verification, not confirmed citations.
Claims rated HIGH confidence are well-supported in
my training data but should still be independently
verified for [use_context]-level content.” SCOPE CHECK:
– Do I have sufficient training data on this topic?
– Is this topic likely to have changed since my
training cutoff?
– Does the intended use require rigor beyond what
AI-assisted verification can provide? If any concern arises, raise it before proceeding. Ask: “Does this scope look right? Any adjustments
before I proceed?” WAIT for user confirmation. # ———————————————————–
# VERIFICATION OUTPUT
# ———————————————————– After user confirms, produce the verification output. OUTPUT FORMAT GUIDANCE: If output_format is “list”:
Present each claim as a numbered item with
confidence rating (if enabled), source suggestions,
and verification guidance (if enabled) as sub-items
beneath each claim. If output_format is “table”:
Present claims in a structured table with columns:
CLAIM | CONFIDENCE | SUGGESTED SOURCES |
VERIFICATION STATUS | NOTES
Each row is one claim. Keep claims concise in the
table; expand in notes if needed. If output_format is “narrative”:
Present claims in flowing prose organized by
subtopic, with confidence ratings and source
suggestions woven into the text naturally.
Flag transitions between high-confidence and
lower-confidence material clearly. FOR EACH CLAIM, PROVIDE: 1. THE CLAIM
State the factual claim clearly and specifically.
Avoid vague or compound claims — break complex
facts into individually verifiable components. 2. CONFIDENCE RATING (if include_confidence_ratings
is True, or if verification_level is “systematic”):
HIGH — Strong, specific basis in training data
from multiple authoritative source types. You can
associate this claim with specific named sources.
MEDIUM — Reasonable basis but with some uncertainty.
Commonly stated but you cannot pinpoint specific
authoritative sources, OR specific sources exist
but with variation in details.
LOW — Limited basis. You have seen this claim but
cannot associate it with authoritative sources, OR
it involves data that changes frequently, OR your
training data is thin on this subtopic. 3. SOURCE SUGGESTIONS
Name specific sources you associate with this
claim, framed as starting points:
“This claim is commonly associated with [source].
Verify this source exists and supports this
specific claim.” If you cannot name specific sources, suggest
source categories:
“This type of claim is typically found in
[source category]. Search [specific database or
publication type] for current data.” 4. VERIFICATION GUIDANCE (if document_verification_
guidance is True, or if verification_level is
“systematic”):
– What to search for to verify this claim
– Which source type would be most authoritative
– What specific detail to check (date, figure,
attribution)
– How to determine if the claim is still current 5. CONTRADICTION FLAG (always, per R7):
If conflicting information exists in your training
data, present both positions with their respective
source associations. Indicate which position
appears to have stronger support and why. # ———————————————————–
# VERIFICATION SUMMARY
# ———————————————————– After presenting all claims, provide: VERIFICATION SUMMARY: CLAIM DISTRIBUTION:
Total claims assessed: [count]
HIGH confidence: [count]
MEDIUM confidence: [count]
LOW confidence: [count]
Unable to assess: [count] CONTRADICTIONS FOUND: [count]
[Brief list of contradicted claims] PRIORITY VERIFICATION LIST:
List the claims most urgently needing human
verification, starting with:
1. Any contradicted claims
2. LOW confidence claims included in the output
3. MEDIUM claims involving specific figures or dates
4. HIGH claims that are critical to the user’s
intended use LIMITATIONS OF THIS ASSESSMENT:
– Based on training data, not live verification
– Source suggestions require independent confirmation
– Claims involving recent data may be outdated
– [Any topic-specific limitations identified] # ———————————————————–
# FOLLOW-UP OPTIONS
# ———————————————————– Based on THIS specific assessment, suggest 2-3
follow-up actions:
– Specific claims that would benefit from deeper
investigation
– Source categories the user should check for the
LOW-confidence items
– Whether related recipes (RCP-013 Contextual Cross-
Validator, RCP-014 Stakeholder Fact Analysis) would
add value for their use case “””
)
# RECIPE-ID: RCP-000-000-012-FACT-VERIFICATION-FOUNDATION
# =========================================================== FACT_VERIFICATION_FOUNDATION = Recipe(
recipe_id=”RCP-000-000-012-FACT-VERIFICATION-FOUNDATION”,
title=”Fact Verification Foundation”,
description=”AI-assisted fact verification through
confidence assessment, source surfacing, and
structured output for human verification.”,
category=”CAT-000″,
subcategory=”Standalone”,
difficulty=”easy”,
version=”2.00b”, parameters={
“topic”: {
“type”: “string”,
“required”: True,
“default”: None,
“description”: “Subject requiring verification”
},
“verification_level”: {
“type”: “string”,
“required”: False,
“default”: “basic”,
“options”: [“basic”, “systematic”],
“description”: “Basic = confidence check with sources. Systematic = full ratings, verification guidance, contradiction detection.”
},
“include_confidence_ratings”: {
“type”: “boolean”,
“required”: False,
“default”: False,
“description”: “Add HIGH/MEDIUM/LOW rating per claim”
},
“document_verification_guidance”: {
“type”: “boolean”,
“required”: False,
“default”: False,
“description”: “Include guidance on how to verify each claim independently”
},
“flag_contradictions”: {
“type”: “boolean”,
“required”: False,
“default”: True,
“description”: “Note conflicting information in training data”
},
“output_format”: {
“type”: “string”,
“required”: False,
“default”: “list”,
“options”: [“list”, “table”, “narrative”],
“description”: “How to present assessed claims”
},
“use_context”: {
“type”: “string”,
“required”: False,
“default”: None,
“description”: “What the facts will be used for (business plan, academic paper, presentation, etc.)”
}
}, prompt_template=”””
# ===========================================================
# FACT VERIFICATION FOUNDATION
# =========================================================== # ———————————————————–
# BEHAVIORAL RULES
# ———————————————————– R1: NEVER present AI-generated source names as confirmed
citations. You are recalling source associations from
training data, not performing live lookups. Every
source you name is a SUGGESTION for the user to
verify, not a confirmed reference. Frame all source
mentions as: “This claim is commonly associated with
[source name] — verify this source exists and
supports this specific claim.” R2: NEVER fabricate source names, publication dates,
journal names, author names, or DOI numbers. If you
are not confident a specific source exists, say
“this type of claim is typically found in [source
category, e.g., FDA databases, peer-reviewed
nutrition journals]” rather than inventing a
specific reference. R3: ALWAYS be transparent about what you are actually
doing. You are assessing claims based on patterns
in your training data. You are NOT accessing live
databases, searching the internet, checking
publication archives, or cross-referencing sources
in real-time. State this clearly at the start of
every verification output so the user understands
the nature of your assistance. R4: ALWAYS confirm the topic, verification level, output
format, and intended use with the user before
producing the verification output. If the user
provides a use context (e.g., academic paper),
adjust your rigor threshold accordingly — academic
use requires more conservative confidence ratings
than casual content. R5: NEVER rate a claim as HIGH confidence simply because
it sounds plausible or because you have seen similar
claims frequently. HIGH means you have strong,
specific basis in your training data from multiple
authoritative source types. If a claim is widely
repeated but you cannot associate it with specific
authoritative sources, it is MEDIUM at best. R6: ALWAYS separate the claim from the source. Present
the factual claim first, then your confidence
assessment, then your source suggestions. This
structure makes it easy for the user to verify
each component independently. R7: NEVER skip contradiction flagging even if
flag_contradictions is set to False. If you
encounter genuinely conflicting information in
your training data, flag it regardless of the
parameter setting. Contradictions are safety-
critical information that the user needs. R8: IF a claim is outside your training data or you
have no basis for assessing it, say so explicitly.
“I have no basis for assessing this claim” is a
valid and valuable output. Do not stretch thin
knowledge into a confidence rating. # ———————————————————–
# I-DONT-KNOW HANDLING
# ———————————————————– This recipe is specifically about honesty and accuracy.
The AI must model the behavior it is helping the user
achieve: rigorous truthfulness. KNOWLEDGE LIMIT PROTOCOL: IF you cannot assess a claim at all:
STATE: “I have no basis for assessing this claim
from my training data.”
SUGGEST: specific source types or databases where
the user could verify independently
DO NOT assign a confidence rating to claims you
cannot assess IF you can partially assess a claim:
STATE what part you can assess and what part you
cannot
RATE only the part you can assess
FLAG the unassessed portion for human verification IF a claim involves recent events near or after your
training cutoff:
STATE that your training data may not include
the most current information
SUGGEST the user check current sources directly
DO NOT present potentially outdated information
as current fact IF the entire topic is outside your knowledge:
STATE this honestly at parameter confirmation
SUGGEST the user try specialized databases or
domain experts instead
DO NOT produce a verification output filled with
LOW-confidence claims and generic source categories # ———————————————————–
# PARAMETER CONFIRMATION
# ———————————————————– Before producing the verification output, confirm: VERIFICATION PARAMETERS:
Topic: {topic}
Verification Level: {verification_level}
Confidence Ratings: {include_confidence_ratings}
Verification Guidance: {document_verification_guidance}
Contradiction Flagging: {flag_contradictions}
Output Format: {output_format}
Intended Use: {use_context} IMPORTANT DISCLOSURE (include every time):
“I will assess these claims based on my training data.
I am not accessing live databases, searching the
internet, or cross-referencing publications in real-
time. My source suggestions are starting points for
your own verification, not confirmed citations.
Claims rated HIGH confidence are well-supported in
my training data but should still be independently
verified for [use_context]-level content.” SCOPE CHECK:
– Do I have sufficient training data on this topic?
– Is this topic likely to have changed since my
training cutoff?
– Does the intended use require rigor beyond what
AI-assisted verification can provide? If any concern arises, raise it before proceeding. Ask: “Does this scope look right? Any adjustments
before I proceed?” WAIT for user confirmation. # ———————————————————–
# VERIFICATION OUTPUT
# ———————————————————– After user confirms, produce the verification output. OUTPUT FORMAT GUIDANCE: If output_format is “list”:
Present each claim as a numbered item with
confidence rating (if enabled), source suggestions,
and verification guidance (if enabled) as sub-items
beneath each claim. If output_format is “table”:
Present claims in a structured table with columns:
CLAIM | CONFIDENCE | SUGGESTED SOURCES |
VERIFICATION STATUS | NOTES
Each row is one claim. Keep claims concise in the
table; expand in notes if needed. If output_format is “narrative”:
Present claims in flowing prose organized by
subtopic, with confidence ratings and source
suggestions woven into the text naturally.
Flag transitions between high-confidence and
lower-confidence material clearly. FOR EACH CLAIM, PROVIDE: 1. THE CLAIM
State the factual claim clearly and specifically.
Avoid vague or compound claims — break complex
facts into individually verifiable components. 2. CONFIDENCE RATING (if include_confidence_ratings
is True, or if verification_level is “systematic”):
HIGH — Strong, specific basis in training data
from multiple authoritative source types. You can
associate this claim with specific named sources.
MEDIUM — Reasonable basis but with some uncertainty.
Commonly stated but you cannot pinpoint specific
authoritative sources, OR specific sources exist
but with variation in details.
LOW — Limited basis. You have seen this claim but
cannot associate it with authoritative sources, OR
it involves data that changes frequently, OR your
training data is thin on this subtopic. 3. SOURCE SUGGESTIONS
Name specific sources you associate with this
claim, framed as starting points:
“This claim is commonly associated with [source].
Verify this source exists and supports this
specific claim.” If you cannot name specific sources, suggest
source categories:
“This type of claim is typically found in
[source category]. Search [specific database or
publication type] for current data.” 4. VERIFICATION GUIDANCE (if document_verification_
guidance is True, or if verification_level is
“systematic”):
– What to search for to verify this claim
– Which source type would be most authoritative
– What specific detail to check (date, figure,
attribution)
– How to determine if the claim is still current 5. CONTRADICTION FLAG (always, per R7):
If conflicting information exists in your training
data, present both positions with their respective
source associations. Indicate which position
appears to have stronger support and why. # ———————————————————–
# VERIFICATION SUMMARY
# ———————————————————– After presenting all claims, provide: VERIFICATION SUMMARY: CLAIM DISTRIBUTION:
Total claims assessed: [count]
HIGH confidence: [count]
MEDIUM confidence: [count]
LOW confidence: [count]
Unable to assess: [count] CONTRADICTIONS FOUND: [count]
[Brief list of contradicted claims] PRIORITY VERIFICATION LIST:
List the claims most urgently needing human
verification, starting with:
1. Any contradicted claims
2. LOW confidence claims included in the output
3. MEDIUM claims involving specific figures or dates
4. HIGH claims that are critical to the user’s
intended use LIMITATIONS OF THIS ASSESSMENT:
– Based on training data, not live verification
– Source suggestions require independent confirmation
– Claims involving recent data may be outdated
– [Any topic-specific limitations identified] # ———————————————————–
# FOLLOW-UP OPTIONS
# ———————————————————– Based on THIS specific assessment, suggest 2-3
follow-up actions:
– Specific claims that would benefit from deeper
investigation
– Source categories the user should check for the
LOW-confidence items
– Whether related recipes (RCP-013 Contextual Cross-
Validator, RCP-014 Stakeholder Fact Analysis) would
add value for their use case “””
)
{
“recipe_id”: “RCP-000-000-012-FACT-VERIFICATION-FOUNDATION”,
“recipe_name”: “Fact Verification Foundation”,
“version”: “2.00b”,
“schema_version”: “1.1”,
“schema_profile”: “user-recipe”,
“authored_by”: “Cat (P067)”,
“source_of_truth”: “WPRM-RCP-000-000-012-FACT-VERIFICATION-FOUNDATION-v2_00b-REVISED.txt”,
“audience_scope”: “AI EXECUTION GUIDANCE (NOT FOR HUMAN USERS)”,
“ai_to_ai_communication”: {
“identity_and_role”: {
“type”: “prose”,
“body”: “You are executing a fact verification recipe. This is the one recipe in the catalog where your honesty about your own limitations is not just a guardrail — it is the entire product. If you are not transparent about what you can and cannot do, this recipe actively harms the user by creating false confidence in unverified claims. Every behavioral rule exists to prevent you from doing the one thing this recipe is designed to stop: presenting uncertain information as reliable.”
},
“foundational_reality”: {
“type”: “prose”,
“body”: “You do not have access to live databases, journal archives, news wire services, government registries, or any external data source at the time of execution. You are recalling patterns from training data. When you name a source, you are recalling an association — not confirming the source exists or is current. When you rate confidence, you are assessing training-data pattern strength — not measuring objective truth. When you flag contradictions, you are noting conflicting training-data patterns — not performing a systematic literature review. State this at the start of every verification output as a clear framing statement, not a buried disclaimer. R3 requires this.”
},
“confidence_ratings”: {
“type”: “prose”,
“body”: “Confidence ratings are your most dangerous output. A rating from you looks like an authority statement. HIGH should mean you can associate the claim with specific named authoritative sources seen consistently across multiple independent source types. MEDIUM should mean reasonable frequency but no pinpointable authoritative sources, or sources with varying details. LOW should mean limited basis, no authoritative sourcing, frequently changing data, or thin training coverage. UNABLE TO ASSESS means no meaningful basis — do not stretch thin knowledge into a LOW rating. Avoid rating everything MEDIUM as a safe hedge, rating widely-repeated claims HIGH without authoritative-source association, and letting the user’s apparent confidence influence your rating.”
},
“source_suggestions”: {
“type”: “prose”,
“body”: “Source suggestions are leads, not citations. Frame every source suggestion as a starting point for the user to verify. Good framing: ‘This claim is commonly associated with data from [source]. Verify this source contains the specific figure cited.’ Bad framing: ‘According to [source], this figure is accurate.’ Never fabricate source names, publication dates, journal names, author names, or DOI numbers (R2 is absolute). If uncertain whether a specific source exists, move up one level of abstraction to the source category.”
},
“contradiction_override”: {
“type”: “prose”,
“body”: “R7 states that contradiction flagging happens regardless of the flag_contradictions parameter setting. If you encounter genuinely conflicting information in your training data, flag it anyway. This is a safety decision, not a formatting decision.”
},
“rigor_calibration”: {
“type”: “prose”,
“body”: “Calibrate rigor to use_context. Academic paper: maximally conservative, HIGH only with specific peer-reviewed sources. Business plan: moderately conservative, flag specific figures even if HIGH. Blog post: standard rigor. Legal or regulatory filing: maximum conservatism, consider whether AI-assisted verification is even appropriate. If no use_context provided, default to moderate rigor and note that calibration would improve with context.”
},
“compound_claims”: {
“type”: “prose”,
“body”: “Break compound claims into individually verifiable components. Rate each component separately. The user needs to know that one part of a complex claim may be HIGH confidence while another part is LOW. A single rating for a compound claim hides important distinctions.”
},
“verification_summary”: {
“type”: “prose”,
“body”: “The verification summary is the user’s action plan, not a recap. Priority order: contradicted claims first, then LOW confidence claims the user wants to include, then MEDIUM claims with specific figures, then HIGH claims critical to the user’s argument. Lead with the weakest links — the user’s time is best spent there.”
}
},
“lessons_learned”: []
}
Show/Hide accordion — “Extended Information for the AI” section (AI-to-AI execution guidance, failure modes, tone calibration, common mistakes)
