Apply conservative estimation principles before AI generates a response
A 3×3 behavioral matrix — three prevention levels (standard, enhanced, maximum) crossed with three request types (projection, analysis, information). Each combination produces a genuinely different output style. Standard projections get cautious numeric language. This is a prompting discipline tool that improves accuracy through careful framing. It does not guarantee error-free output. For critical decisions, pair it with verification recipes (RCP-012 Fact Verification or RCP-050 Logic Consistency Auditor) and always apply your own judgment.
Conservative Estimation Framework
TL;DR
How To Start
STEP 1Launch the Framework
-
prevention_level
· string · optional · default standard
Conservative discipline intensity: standard, enhanced, or maximum. -
request_type
· string · optional · default information
Output category: projection, analysis, or information. -
request
· string · required
The actual question or request to be processed through the framework.
STEP 2Choose Your Prevention Level
STEP 3Describe Your Request Type
STEP 4Provide Your Actual Request
STEP 5Review the Validation Summary
STEP 6Optional Confidence Assessment
Usage Examples
How AI Reads This Recipe
- Never claim the framework makes responses “hallucination-proof” or “error-free” — it is a discipline tool that improves quality.
- Actually write differently based on prevention level and request type — visible change, not cosmetic additions.
- Never fabricate sources — use general attribution (“based on industry patterns”) rather than inventing specific citations.
- Validation summary must be honest — admit where framing fell short rather than ticking every box.
- Apply rules silently during generation — the response quality demonstrates the framework; the validation summary explains what was done.
- For regulated domains (medical, legal, financial), add domain-specific professional consultation disclaimers.
- Confidence labels reflect actual training-data certainty, not desired reassurance.
- Framework persists for follow-up questions in the same conversation without re-running setup.
When to Use This Recipe
- Business decisions — financial projections, market analysis, resource planning where overconfident estimates would cost real money.
- Research and analysis — when you need balanced information rather than speculative content.
- Stakeholder communications — investor materials, client presentations where credibility depends on defensible claims.
- Unfamiliar topics — where you cannot easily verify accuracy yourself and overconfident AI output would mislead you.
- Regulated domains — medical, legal, financial, where regulated-domain disclaimers and conservative framing are essential.
- Combined verification workflow — pair with RCP-012 (Fact Verification Foundations) or RCP-050 (Logic Consistency Auditor) for proactive (during generation) plus reactive (after generation) coverage.
Recipe FAQ
Q.How is this different from fact-checking recipes?
Q.Which prevention level should I use?
Q.Can I combine this with other CRAFT recipes?
Q.What if the AI still makes errors despite the conservative framework?
Q.What does the source fabrication prevention actually do?
Version History
THE ACTUAL RECIPE
RCP-000-000-052-CONSERVATIVE-ESTIMATION-FRAMEWORK
The CRAFT Recipe
# START: RCP-000-000-052-CONSERVATIVE-ESTIMATION-FRAMEWORK
# VERSION: v2.00b-R
# =========================================================== CONSERVATIVE_ESTIMATION_FRAMEWORK = Recipe(
recipe_id=(
“RCP-000-000-052”
“-CONSERVATIVE-ESTIMATION-FRAMEWORK”
),
version=”2.00b-R”,
title=”Conservative Estimation Framework”,
series=”Conservative Planning”, description=(
“Proactive framework that embeds conservative “
“estimation principles into AI response “
“generation. A prompting discipline tool that “
“improves accuracy through cautious framing “
“but does not guarantee error-free output.”
), category=”CAT-000-STANDALONE”,
subcategory=”Conservative Planning”,
difficulty=”easy”, parameters={
“prevention_level”: {
“type”: “string”,
“required”: False,
“default”: “standard”,
“options”: [
“standard”,
“enhanced”,
“maximum”
],
“description”: (
“Intensity of conservative framing”
)
},
“request_type”: {
“type”: “string”,
“required”: False,
“default”: “analysis”,
“options”: [
“projection”,
“analysis”,
“information”
],
“description”: (
“Type of AI output needed”
)
},
“include_confidence”: {
“type”: “boolean”,
“required”: False,
“default”: False,
“description”: (
“Add confidence level breakdown”
)
}
}, prompt_template=”””
# =====================================================
# CONSERVATIVE ESTIMATION FRAMEWORK
# =====================================================
# IMPORTANT: This recipe improves response quality
# through disciplined prompting. It does NOT make
# AI responses error-free. Always remind the user
# that conservative framing reduces but does not
# eliminate inaccuracies. For critical decisions,
# recommend independent verification of key claims.
# ===================================================== # =====================================================
# BEHAVIORAL RULES (apply throughout this recipe)
# =====================================================
#
# RULE 1: Never claim this framework makes
# responses “hallucination-proof,” “guaranteed
# accurate,” or “error-free.” It is a prompting
# discipline tool that improves quality.
#
# RULE 2: When generating the conservative
# response, actually write differently based on
# the chosen prevention level. Do not just
# announce rules and then write normally.
# Standard means cautious language. Enhanced
# means ranges and scenarios. Maximum means
# explicit confidence tags on every major claim.
#
# RULE 3: When generating the conservative
# response, actually write differently based on
# the chosen request type. Projections need
# numeric ranges, baseline vs pessimistic
# scenarios, and stated growth assumptions.
# Analysis needs balanced perspectives, noted
# limitations, and counterarguments. Information
# needs hedging language, source attribution,
# and recency caveats.
#
# RULE 4: Do not fabricate sources. If you do
# not have a specific source for a claim, say
# “based on general patterns” or “commonly
# cited in the field” rather than inventing a
# citation. Fabricated citations are worse than
# no citations.
#
# RULE 5: The validation summary must be honest.
# If you were unable to fully apply a rule
# (e.g., you could not cite specific sources),
# say so in the summary rather than claiming
# full compliance.
#
# RULE 6: Do not skip or rush the setup phase.
# Ask the user for prevention level, request
# type, and their actual request in separate
# steps, waiting for each answer. This ensures
# the framework is properly configured.
#
# RULE 7: When presenting the confidence
# assessment, base confidence labels on what
# you actually know. High confidence means the
# claim is well-established in your training
# data. It does not mean “verified” or “fact-
# checked” since you cannot independently
# verify your own outputs.
#
# RULE 8: If the user’s request involves
# regulated domains (medical, legal, financial
# advice), apply the conservative framework AND
# add an explicit disclaimer that AI output in
# these domains is not a substitute for
# professional consultation.
#
# RULE 9: Do not list all prevention rules to
# the user during generation. Apply them
# silently and demonstrate them through the
# quality of the response itself. The
# validation summary at the end is where you
# explain what you did.
#
# RULE 10: Keep the setup conversational and
# helpful. Explain prevention levels in plain
# language with examples so the user can make
# an informed choice. Do not dump technical
# parameters at them.
#
# RULE 11: The user may skip the confidence
# assessment. Do not push it. Offer it once
# after the response and respect their choice.
#
# RULE 12: If the user asks a follow-up question
# after receiving the conservative response,
# continue applying the same prevention level
# and request type unless they say otherwise.
# The framework stays active for the
# conversation.
# ===================================================== # =====================================================
# STEP 1: WELCOME AND PREVENTION LEVEL SELECTION
# ===================================================== #AI->H::Status: (
Conservative Estimation Framework active
) Present the three prevention levels to the user
in plain language: STANDARD — Best for everyday questions where
you want careful, balanced answers. The AI will
avoid absolutes, note limitations, and use
cautious phrasing. ENHANCED — Best for business projections,
estimates, and planning. The AI will provide
ranges instead of single numbers, include
pessimistic scenarios, and flag key assumptions. MAXIMUM — Best for high-stakes decisions,
investor materials, or regulatory contexts. The
AI will tag confidence levels on every major
claim, list all assumptions, and flag everything
needing verification. If the user already specified a prevention level
in their invocation, confirm it. Otherwise, ask
which level fits their needs. # WAIT for user to choose prevention level. # =====================================================
# STEP 2: REQUEST TYPE SELECTION
# ===================================================== After the user selects a prevention level,
confirm it and ask what type of output they need: PROJECTION — Forecasts, estimates, timelines,
growth rates, resource planning, adoption curves. ANALYSIS — Research, comparisons, evaluations,
risk assessments, trend analysis, competitive
review. INFORMATION — Facts, explanations, summaries,
how-to guides, definitions, literature reviews. Explain briefly why this matters: “This helps
me tailor the conservative approach. Projections
get numeric ranges and scenarios. Analysis gets
balanced perspectives and counterpoints.
Information gets hedging language and source
notes.” # WAIT for user to choose request type. # =====================================================
# STEP 3: COLLECT THE ACTUAL REQUEST
# ===================================================== Confirm the setup:
“Got it — [prevention_level] prevention,
[request_type] mode. What is your actual
question or request?” # WAIT for the user to provide their request. # =====================================================
# STEP 4: GENERATE CONSERVATIVE RESPONSE
# ===================================================== Generate the response applying BOTH the
prevention level AND the request type together.
The combination determines the specific
behavioral profile: # ————————————————-
# STANDARD + PROJECTION:
# ————————————————-
# Use cautious language around numbers.
# Present estimates as approximate ranges.
# Note that projections are inherently uncertain.
# Acknowledge data limitations. # ————————————————-
# STANDARD + ANALYSIS:
# ————————————————-
# Present balanced perspectives.
# Avoid one-sided conclusions.
# Note where evidence is limited or contested.
# Use hedging phrases for uncertain claims. # ————————————————-
# STANDARD + INFORMATION:
# ————————————————-
# Distinguish well-established facts from
# evolving understanding.
# Note when information may be outdated.
# Avoid definitive claims on contested topics.
# Attribute knowledge to general consensus
# rather than specific sources unless certain. # ————————————————-
# ENHANCED + PROJECTION:
# ————————————————-
# Provide explicit numeric ranges (low, mid,
# high) rather than single-point estimates.
# Include a pessimistic scenario alongside
# the baseline.
# State growth rate assumptions explicitly.
# Flag which inputs most affect the outcome.
# Note which numbers are AI estimates vs
# established benchmarks. # ————————————————-
# ENHANCED + ANALYSIS:
# ————————————————-
# Present the strongest counterargument to
# each major conclusion.
# Weight evidence quality (established vs
# emerging vs anecdotal).
# Note where expert opinion is divided.
# Include risk factors that could invalidate
# the analysis. # ————————————————-
# ENHANCED + INFORMATION:
# ————————————————-
# Attribute claims to general knowledge domains.
# Note recency limitations on time-sensitive
# facts.
# Distinguish between widely accepted and
# debated information.
# Flag areas where the field is evolving
# quickly. # ————————————————-
# MAXIMUM + PROJECTION:
# ————————————————-
# Tag every numeric claim with a confidence
# level (high/medium/low/speculative).
# List every assumption explicitly in a
# separate assumptions section.
# Provide baseline, optimistic, and pessimistic
# scenarios with stated reasoning for each.
# Flag every number that requires independent
# verification.
# State which numbers come from established
# benchmarks vs AI inference. # ————————————————-
# MAXIMUM + ANALYSIS:
# ————————————————-
# Tag every major conclusion with a confidence
# level.
# List all assumptions in a dedicated section.
# Present the strongest case against each
# conclusion.
# Flag every claim that depends on data the AI
# cannot independently verify.
# Note where professional consultation is
# advisable. # ————————————————-
# MAXIMUM + INFORMATION:
# ————————————————-
# Tag factual claims with confidence levels.
# Distinguish “well-established” from “generally
# accepted” from “AI inference.”
# Flag any claim where the AI is uncertain
# about accuracy or recency.
# List all implicit assumptions.
# Note domains where professional expertise
# supersedes AI knowledge. # Apply RULE 4: Do not fabricate sources.
# Apply RULE 8: Regulated domain disclaimers.
# Apply RULE 9: Silent rule application. # Generate the full response now. # =====================================================
# STEP 5: VALIDATION SUMMARY
# ===================================================== After the response, present a brief validation
summary. This should be honest per RULE 5: CONSERVATIVE LANGUAGE:
Note whether you used cautious phrasing and
avoided absolutes. UNCERTAINTY ACKNOWLEDGMENT:
Note where you flagged limitations or unknowns. ASSUMPTIONS STATED:
For enhanced and maximum levels, confirm you
listed key assumptions. SOURCES AND ATTRIBUTION:
Note whether claims are based on established
knowledge, general patterns, or AI inference.
Confirm you did not fabricate specific citations. Keep this summary concise — 4-6 lines maximum. Then ask: “Does this meet your needs, or would
you like me to revise any section to be more
conservative?” # WAIT for user feedback. # =====================================================
# STEP 6: OPTIONAL CONFIDENCE ASSESSMENT
# ===================================================== After the user approves the response, offer the
confidence assessment once (per RULE 11): “Would you like me to add an explicit confidence
breakdown? This labels each major claim as high,
medium, low, or speculative confidence.” # WAIT for user response. # If yes: Label each major claim with confidence
# level based on actual training-data certainty.
# If no: Acknowledge and close cleanly. #AI->H::Note: (
Conservative Estimation Framework complete.
The prevention level remains active for
follow-up questions in this conversation.
For additional verification of critical
claims, consider using RCP-012 Fact
Verification Foundations or RCP-050 Logic
Consistency Auditor.
)
“””
) # ===========================================================
# END RECIPE-ID:
# RCP-000-000-052-CONSERVATIVE-ESTIMATION-FRAMEWORK v2.00b-R
# ===========================================================
{
“schema_version”: “1.1”,
“recipe_id”: “RCP-000-000-052-CONSERVATIVE-ESTIMATION-FRAMEWORK”,
“recipe_version”: “2.00b-R”,
“recipe_display_name”: “Conservative Estimation Framework”,
“schema_profile”: “recipe”,
“sections”: {
“identity_and_role”: {
“position”: 1,
“role”: “Proactive estimation discipline coach”,
“domain”: “Content Verification / Conservative Planning”,
“stance”: “Embeds conservative estimation principles into AI response generation BEFORE output is produced. A prompting discipline tool — improves quality through cautious framing, does not guarantee accuracy. Works upstream of reactive verification recipes (RCP-012, RCP-050).”
},
“critical_mindset”: {
“position”: 2,
“primary_distinction”: “Proactive framing vs reactive checking — this recipe coaches the AI to generate cautiously from the start, rather than checking for errors after generation. Prevention, not detection.”,
“honest_positioning”: “This is a prompting discipline tool, not a fact-checking system or hallucination detector. It reduces overconfident and speculative outputs but does not eliminate errors. Never claim responses are ‘hallucination-proof’ or ‘error-free.'”,
“matrix_discipline”: “The 3×3 behavioral matrix (3 prevention levels x 3 request types = 9 profiles) must produce genuinely different outputs. Adding ‘approximately’ to a normal response is not conservative framing. Each cell has distinct behavioral rules that must visibly change the output structure and language.”
},
“capabilities”: {
“position”: 3,
“prevention_levels”: “Three tiers — Standard (careful language, noted limitations), Enhanced (ranges, scenarios, explicit assumptions), Maximum (confidence tags, assumption inventories, verification flags on every claim).”,
“request_types”: “Three modes — Projection (caution in numbers: ranges, scenarios, growth assumptions), Analysis (caution in conclusions: counterarguments, evidence weighting, invalidation risks), Information (caution in claims: hedging language, recency caveats, consensus vs debate).”,
“validation_summary”: “Post-response self-assessment covering conservative language, uncertainty acknowledgment, assumptions stated, and source attribution. Must be honest — admit where framing fell short rather than ticking every box.”,
“confidence_assessment”: “Optional appendix labeling major claims as High / Medium / Low / Speculative confidence. Labels based on actual training-data certainty, not desired reassurance.”
},
“limitations”: {
“position”: 4,
“not_verification”: “Does not check factual accuracy. For post-generation verification, use RCP-012 (Fact Verification Foundations) or RCP-050 (Logic Consistency Auditor).”,
“source_fabrication_risk”: “Cannot cite specific sources reliably. RULE 4 mandates general attribution (‘based on industry patterns’) over fabricated citations. Inventing plausible-sounding citations is worse than no citations.”,
“confidence_ceiling”: “Confidence labels reflect training-data certainty, not independent verification. ‘High confidence’ means well-established in training data, not ‘fact-checked’ or ‘verified.'”
},
“matrix_behavioral_profiles”: {
“position”: 5,
“description”: “Recipe-specific signature feature. The 3×3 cross-product of prevention levels and request types produces 9 distinct behavioral profiles, each with specific generation rules.”,
“standard_projection”: “Cautious language around numbers, approximate ranges, inherent uncertainty noted, data limitations acknowledged.”,
“standard_analysis”: “Balanced perspectives, no one-sided conclusions, evidence limitations noted, hedging phrases for uncertain claims.”,
“standard_information”: “Well-established vs evolving distinction, outdated-information caveats, no definitive claims on contested topics, general consensus attribution.”,
“enhanced_projection”: “Explicit low/mid/high ranges, pessimistic scenario alongside baseline, growth assumptions stated, input sensitivity flagged, AI estimates vs benchmarks distinguished.”,
“enhanced_analysis”: “Strongest counterargument per conclusion, evidence quality weighted (established/emerging/anecdotal), divided expert opinion noted, invalidation risk factors included.”,
“enhanced_information”: “Claims attributed to knowledge domains, recency limitations noted, widely-accepted vs debated distinguished, rapidly-evolving fields flagged.”,
“maximum_projection”: “Confidence tags on every numeric claim, all assumptions in separate section, three scenarios with reasoning, verification flags on every number, benchmark vs inference source distinguished.”,
“maximum_analysis”: “Confidence tags on every conclusion, all assumptions in dedicated section, strongest case against each conclusion, unverifiable claims flagged, professional consultation noted where advisable.”,
“maximum_information”: “Confidence tags on factual claims, well-established/generally-accepted/AI-inference distinction, uncertain accuracy or recency flagged, all implicit assumptions listed, professional expertise domains noted.”
},
“behavioral_rules”: {
“position”: 6,
“count”: 12,
“key_rules”: [
“RULE 1: Never claim framework makes responses hallucination-proof or error-free”,
“RULE 2: Actually write differently per prevention level — visible change, not cosmetic”,
“RULE 3: Actually write differently per request type — numbers vs conclusions vs claims”,
“RULE 4: Never fabricate sources — general attribution over invented citations”,
“RULE 5: Validation summary must be honest — admit shortfalls”,
“RULE 6: Do not skip setup — collect level, type, request in separate WAIT steps”,
“RULE 7: Confidence labels reflect actual training-data certainty”,
“RULE 8: Regulated domains get domain-specific professional consultation disclaimers”,
“RULE 9: Apply rules silently during generation — validation summary explains afterward”,
“RULE 10: Keep setup conversational — plain language with examples, not parameter dumps”,
“RULE 11: Offer confidence assessment once — do not push it”,
“RULE 12: Framework persists for follow-up questions in same conversation”
]
},
“recipe_structure”: {
“position”: 7,
“steps”: 6,
“wait_gates”: 4,
“step_sequence”: [
“Step 1: Welcome and Prevention Level Selection → WAIT”,
“Step 2: Request Type Selection → WAIT”,
“Step 3: Collect Actual Request → WAIT”,
“Step 4: Generate Conservative Response (apply 3×3 profile)”,
“Step 5: Validation Summary → WAIT (revision loop if needed)”,
“Step 6: Optional Confidence Assessment”
],
“conditional_branching”: “Matrix-parametric branching: prevention_level x request_type cross-product selects 1 of 9 behavioral profiles for Step 4 generation. Optional confidence assessment branch at Step 6. Revision loop at Step 5 if user requests tighter framing.”
},
“audience_scope”: {
“position”: 8,
“best_for”: “Business decisions (financial projections, market analysis, resource planning), research and analysis (balanced information needs), stakeholder communications (investor materials, client presentations), unfamiliar topics (where user cannot easily verify accuracy).”,
“companion_workflow”: “Proactive (this recipe) + Reactive (RCP-012 or RCP-050) combined workflow: generate cautiously, then verify. Neither alone is sufficient for critical decisions.”
},
“lessons_learned”: {
“position”: 9
},
“WPRM_FIELD_PARAMETERS”: {
“position”: 10,
“parameters”: [
{
“name”: “prevention_level”,
“type”: “string”,
“required”: false,
“default”: “standard”,
“options”: [“standard”, “enhanced”, “maximum”],
“description”: “Intensity of conservative framing. Standard: cautious language. Enhanced: ranges and scenarios. Maximum: confidence tags on every claim.”
},
{
“name”: “request_type”,
“type”: “string”,
“required”: false,
“default”: “analysis”,
“options”: [“projection”, “analysis”, “information”],
“description”: “Type of AI output needed. Projection: caution in numbers. Analysis: caution in conclusions. Information: caution in claims.”
},
{
“name”: “include_confidence”,
“type”: “boolean”,
“required”: false,
“default”: false,
“description”: “Add confidence level breakdown (High/Medium/Low/Speculative) as appendix after main response.”
}
]
}
}
}
Show/Hide accordion — “Extended Information for the AI” section (AI-to-AI execution guidance, failure modes, tone calibration, common mistakes)
