2nd/10_Wiki/Topics/Other/Assumptions-vs-Facts.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-assumptions-vs-facts	Assumptions vs Facts	10_Wiki/Topics	verified	self	Fact-Assumption Distinction Premise vs Evidence	none	A	0.9	applied	reasoning epistemology decision-making critical-thinking		2026-05-10	pending	language framework python na

Assumptions vs Facts

매 한 줄

"매 fact 는 매 verifiable observation, 매 assumption 은 매 unverified premise". 매 둘 의 conflation 매 most decision failure 의 root. 매 military intelligence (CIA Tradecraft Primer), 매 software engineering (RFC, design doc), 매 LLM agent reasoning (chain-of-thought 매 assumption 명시) 모두 의 핵심 discipline.

매 핵심

매 정의

• Fact: 매 currently verifiable claim — 매 measurement, 매 reproducible observation, 매 authoritative record.
• Assumption: 매 not verified, 매 taken as true 매 reasoning 진행 위해. 매 implicit / explicit.
• Inference: 매 fact + assumption → 매 conclusion.

매 Verification spectrum

• Hard fact: 매 measurement (e.g., latency = 142ms p95).
• Soft fact: 매 expert testimony / consensus (e.g., "FDA-approved").
• Reasonable assumption: 매 base rate / 매 prior (e.g., "user 매 attention < 10s").
• Speculative assumption: 매 untested premise (e.g., "competitor 매 Q4 launch").

매 응용

1. Design doc: 매 "Assumptions" section 별도 — 매 reviewer 검증.
2. Intelligence analysis: 매 ACH (Analysis of Competing Hypotheses).
3. Postmortem: 매 implicit assumption 적출 — 매 next-time fact 로 verify.
4. LLM CoT: 매 reasoning chain 에서 매 assumption 의 explicit tag.

💻 패턴

Pattern 1: 매 Design doc template

## Facts
- 매 current p95 latency: 240ms (verified via 매 grafana 2026-05-09).
- 매 user count: 1.2M MAU (analytics dashboard).

## Assumptions
- [A1] 매 traffic grow 30% YoY (prior: 2024-2025 trend).
- [A2] 매 redis cluster 매 horizontal scale 가능 (vendor docs, untested at our scale).

## Inferences
- A1 + Facts → 매 Q4 capacity = 1.56M MAU.
- A2 + Facts → 매 cache layer 매 bottleneck 의 X.

## Validation plan
- A1: 매 monthly reforecast.
- A2: 매 Q3 load-test 8x current.

Pattern 2: 매 ACH (Analysis of Competing Hypotheses)

import numpy as np
hypotheses = ["H1: 매 supply shock", "H2: 매 demand drop", "H3: 매 competitor"]
evidence   = ["E1: price up", "E2: query down", "E3: rival ad spike"]
# 매 매 evidence × hypothesis: consistent (+1), inconsistent (-1), N/A (0)
M = np.array([
    # E1, E2, E3
    [+1,  0,  0],   # H1
    [-1, +1,  0],   # H2
    [ 0, +1, +1],   # H3
])
scores = M.sum(axis=1)
for h, s in zip(hypotheses, scores):
    print(h, s)
# 매 lowest disconfirmed = 매 most likely (CIA tradecraft logic)

Pattern 3: 매 Assumption tagging in CoT

def reason_with_tags(query: str) -> str:
    return llm(f"""
Answer step by step. For every claim:
- Tag [FACT: source] if verifiable.
- Tag [ASSUMP: confidence 0-1] if untested.
- Tag [INFER] if derived.

Q: {query}
""")

Pattern 4: 매 Premortem (assumption stress-test)

Imagine the project failed in 6 months. List the 5 most likely
failed assumptions. For each, design a 2-week experiment to test
it now.

Pattern 5: Confidence score 매 calibration

predictions = []  # list of (claim, confidence, actual_outcome)
brier = sum((c - a)**2 for _, c, a in predictions) / len(predictions)
print(f"Brier score: {brier:.3f}")  # 매 lower = better calibration

매 결정 기준

상황	Treat as
매 metric in current dashboard	Fact (with date)
매 vendor capability claim	Soft fact, 매 verify if critical
매 future user behavior	Assumption — 매 explicit
매 "everyone knows"	매 strong assumption — 매 challenge
매 LLM output	Assumption until cross-checked

기본값: 매 reasoning 시작 시 매 explicit "Facts" / "Assumptions" 분리. 매 implicit assumption 의 surface — 매 brittle.

🔗 Graph

• 부모: Belief-Revision · Bayesian-Updating
• 변형: Bayes-Theorem · Hypostatic-Abstraction
• 응용: Problem Solving Process · Process_Reflection_Template
• Adjacent: Big-Picture · Outside-Thinking · Anticipation

🤖 LLM 활용

언제: 매 agent design — 매 [FACT]/[ASSUMP] tagging 매 hallucination detection 도움. 매 reasoning trace audit. 언제 X: 매 creative ideation — 매 over-tagging 매 flow 방해.

❌ 안티패턴

• Implicit assumption: 매 unmentioned premise — 매 reviewer 못 catch.
• Fact inflation: 매 weak evidence 의 hard fact 처럼 표현.
• Confidence theater: 매 "obviously" / "clearly" — 매 hidden assumption marker.
• Single-source fact: 매 1 source = 매 still soft. 매 triangulate.
• Stale fact: 매 6개월 전 metric — 매 currently fact 인지 재검증.

🧪 검증 / 중복

• Verified (CIA Tradecraft Primer 2009, Heuer Psychology of Intelligence Analysis, Tetlock Superforecasting).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — ACH + 매 design-doc pattern + LLM CoT tagging