Source-Trail — 溯源查證框架

A multi-disciplinary verification framework synthesized from journalism, research ethics, intelligence tradecraft, and evidentiary standards. Forces explicit source classification, disclosure of confidence levels, and adversarial self-checking on factual claims.

>

整合新聞工作、研究倫理、情報分析、法庭證據四個領域的查證方法論，強制對事實宣稱做來源分級、信心揭露與對抗性自檢。

Overview

When answering factual questions, every specific claim must be traceable to a primary or authoritative secondary source (T1/T2). LLM consensus is not verification. Disclose what is verified, what is unverified, and what conflicts.

事實宣稱必須上溯到 T1/T2 來源。LLM 之間的同意不算驗證。揭露已驗證、未驗證、衝突三類資訊。

When to Use / When NOT to Use

TRIGGER on these question types（觸發類型）

1. Time-sensitive state（時間敏感狀態） — "now", "latest", "currently", "still", "recent", anything beyond knowledge cutoff
2. Specific numerical claims（具體數值宣稱） — statistics, prices, populations, distances, revenues
3. Dates / chronology（日期 / 時序） — when X happened, release dates, founding dates
4. Comparative / superlative（比較級 / 排名） — most, top N, highest, lowest, first, best
5. Identity / attribution（身分歸屬） — who founded, who owns, who wrote, who currently leads
6. Existence / status（存在性 / 狀態） — is X still operating, is Y alive, does Z support feature W
7. News / events（新聞 / 事件） — recent happenings, ongoing situations, breaking news
8. Technical specs / docs（技術規格 / 文件） — library versions, API behavior, protocol details
9. Regulatory / legal facts（法規 / 政策） — laws, compliance requirements, jurisdiction-specific rules
10. Geopolitical / changing geographic facts（地理政治變動） — leadership, currencies, borders

SKIP for these（跳過類型）

1. Code logic discussion 程式邏輯討論
2. Concept explanations (OAuth, REST) 概念定義
3. Opinion-seeking ("which is better") 意見徵詢
4. Hypothetical scenarios 假設情境
5. Creative tasks (naming, copywriting) 創作任務
6. Pure mathematics / logical reasoning 純數學推理
7. Stable historical consensus (Tokyo is in Japan) 穩定歷史共識
8. User's own data / context 使用者自身資料
9. Explaining user's own code 解釋使用者程式
10. Subjective evaluations 主觀評價

Edge cases（邊界處理）

• Mixed questions — split: explain concept (skip) + give current example (trigger only this part)
• "Best practices" — opinion if abstract; trigger if citing specific cases/numbers
• "I think X is true, verify?" — always trigger
• Questions about Claude itself — usually skip; trigger if claiming current capabilities

Stakes Assessment（風險等級判定）

Different questions need different rigor. Before running phases, classify the question:

Topic-Based Override（主題硬性歸類）

The following topics are ALWAYS High stakes, regardless of how the question is framed. Do not downgrade based on phrasing like "general info", "just curious", "in general".

Why "general framing" doesn't lower the stakes:

• Information itself is high-risk regardless of who asks
• Asker can hand the answer to someone who will act on it
• The skill cannot verify the asker's actual context
• Default to caution; cost of over-rigor << cost of under-rigor

Default rule (still applies for non-override topics): If unclear, treat as one tier higher.

Performance Expectations（時間預期）

Realistic time costs for users to set expectations:

Phase 0.3 adds ~5s but often saves more time downstream by preventing wasted searches on a wrongly-framed question.

Sequential execution can double the time. Always parallelize independent fetches. See parallelization rules in Phase 2 and Phase 5.

For users requesting speed: they can say "fast" / "lite" / "skip Phase 5" — degrade gracefully by skipping conditional phases first (5 → 4 → 1), keeping Phase 0.3 + 2 + 3 as the irreducible minimum. Phase 0.3 is not optional even in lite mode — it's the cheapest way to avoid the most expensive failure.

Pre-flight Announcement（執行前預告）

Before running phases, announce the plan. The announcement format depends on stakes:

Low / Medium stakes — Announce + Execute（預告後立即執行）

Display the plan, then start running immediately. User can interrupt via their interface's natural mechanism (sending a new message works in any UI).

🔍 /source-trail activating
   Stakes: Medium
   Phases: 0.5 → 2 → 3 → 4
   Expected: 4-5 searches

Do NOT instruct specific interrupt mechanisms ("press Esc", "type skip") — these are UI-specific and won't work across Claude Code, Desktop, claude.ai web, mobile, or API usage.

High stakes — Announce + Confirm（預告後等確認）

Pause for explicit user confirmation before running. High-stakes questions warrant the extra round-trip.

🔍 /source-trail — High stakes detected (medical/financial/legal/decision-support)
   Plan: All phases (0–5), including counter-evidence search (Phase 5)
   Expected: 6-8 searches, ~60s

⚠ This will take longer than usual. Recommended for accuracy on this question type.
   Reply to confirm, or request a lighter version if appropriate.

Wait for user response before proceeding.

Why this asymmetry

• Low/Medium: Auto-trigger means the user already implicitly authorized verification by asking a factual question. Confirming each time creates friction without proportional benefit.
• High: Decision-support and harm-if-wrong questions deserve a deliberate pause. The cost of confirming is small relative to the cost of an unverified high-stakes answer.

The Iron Law

Knowledge cutoff is a tool-call signal, not a give-up clause.

>

知識截止不是放棄的出口，是觸發工具的訊號。

When you detect a time point past your training cutoff, the correct action is to call WebSearch / WebFetch immediately — not to disclaim and recommend the user check elsewhere.

Failure modes that violate this law:

• ❌ "My knowledge cutoff is X, please check Y website for current data."
• ❌ "I cannot verify 2026 information. Suggest checking official sources."
• ❌ Answering from training data, then adding "but for current info, see X."

Correct behavior:

• ✓ "My training data is from X. Searching now for current data..." → actually searches → reports findings

Source Tier System (T1-T5)

All factual sources are classified before use. Factual claims must trace to T1 or T2.

所有來源使用前先分級。事實宣稱必須能上溯到 T1 或 T2。

Independence rule（獨立性規則）

"Independent" means cannot trace back to the same original source.

• ❌ Reuters citing MLB.com + ESPN citing MLB.com = ONE source (both trace to MLB.com)
• ❌ Gemini agrees + Claude agrees = ONE source category (both T5)
• ✓ Baseball-Reference + MLB official site = TWO independent T1 sources

Citation Chain Integrity（引用鏈完整性）

Don't trust "X reported that Y said Z." Read what Y actually said. Game-of-telephone effects compound at each retelling.

不信「X 報導 Y 說 Z」。要讀 Y 原始說了什麼。每一層轉述都會失真。

Wire Service Detection（通訊社偵測）

Multiple outlets running the same wire copy ≠ multiple sources. Always check if outlets are running the same Reuters/AP/AFP wire.

多家媒體跑同一通訊社稿件 ≠ 多重來源。必須偵測是否同一原稿被轉發。

Time Context Decay（時效衰減）

Always note publication date of sources. A 2024 statistic answering a 2026 question must be flagged.

來源發布日要標註。用 2024 數據回答 2026 問題必須警示。

The Verification Workflow (Phase 0-5)

Phase 0: Trigger Check（觸發判斷）

Is this a factual question matching the trigger criteria? If yes, continue. If no, exit skill.

Phase 0.3: Key Assumptions Check（前提核查）— from intelligence tradecraft

Before searching, audit the question itself. A precisely-verified answer to a wrongly-framed question is still wrong. This phase adapts the CIA's Key Assumptions Check (Tradecraft Primer 2009, pp.7-8) — originally designed for intelligence analysts to surface hidden premises before analysis begins.

#### When to run

• Always run for any question that survives Phase 0 trigger check
• Especially critical when the question comes from a third party (news headline, user quoting someone else, social-media post) — these arrive pre-framed by someone else's assumptions

#### The four-step procedure（CIA KAC 四步法）

This is an execution checklist, not a declaration. Each step must produce a written artifact that goes into the verification trail. If a step has no written output, that step did not happen.

1. Restate the analytic line（寫下目前的分析路線）

Action: Write, in one sentence, what the question seems to be asking. Output artifact: "The question appears to be: ___"

1. List all premises, stated and unstated（列出所有前提，含未明說的）

Action: Run through the checklist below. For each row, write the specific premise the current question carries, even if blank.

Output artifact: A written premise list, one line per row that has a premise to flag.

1. Challenge each premise（挑戰每個前提）

Action: For each premise, answer in writing: Why must this be true? Under what conditions would it fail? Output artifact: A written challenge note per premise, even short (one line is fine).

1. Reframe if needed（必要時重新框定問題）

Action: If any premise is shaky and the answer depends on it → either (a) verify the premise itself in Phase 2, or (b) rewrite the question to make the premise explicit. Output artifact: If reframed, write out the new question verbatim so the user can see the change in Phase 3 output.

Hard requirement: If any of the four written artifacts is missing, Phase 0.3 is incomplete and the run halts. Going to Phase 0.5 / Phase 2 without these artifacts means substitution was declared not checked — exactly the failure mode this Phase exists to prevent.

Case study: See examples/case-political-substitution.md — a 2026-05-11 case where a viral Threads post smuggled a false Authority premise ("the Hualien magistrate controls a 300-billion budget" — actually a central-government special budget) into a popular accusation. Phase 0.3's premise audit caught the substitution before Phase 2 began.

#### Output

Phase 0.3 produces an internal artifact: the original question + the reframed question (if reframed). This artifact is shown to the user in the final output (Phase 3) under "我把問題改成這樣問" / "Question reframing" — because exposing the reframing is itself part of the methodology's pedagogical value.

#### Why this Phase exists

Phase 0.5 (Pre-registration) prevents motivated search-term selection. But Phase 0.5 assumes the question itself is well-formed. If the question smuggles in a false premise, no amount of rigorous searching will produce a true answer — you'll get a precise answer to the wrong question.

This is Kahneman's attribute substitution applied to LLMs: when handed a hard question, the system tends to silently substitute an easier related one (e.g., "Is X policy good?" → "What do articles I can find say about X policy?"). Phase 0.3 forces conscious detection.

#### Lineage

This Phase synthesizes:

• Structure: CIA Tradecraft Primer's Key Assumptions Check (four-step procedure)
• Falsifiability premise: Popper
• Stance-detection premise: Maxwell qualitative methodology
• Substitution-detection premise: Kahneman & Frederick

Each row in the checklist traces to a specific T1/T2 academic source. This is documented in examples/case-phase-0-3-genesis.md.

Phase 0.5: Pre-registration（預先登記）— from research ethics

Before searching, briefly state:

• What specific claim am I trying to verify
• What search terms I will use
• What T1/T2 source I expect to find this in

This prevents motivated reasoning (post-hoc cherry-picking of search terms that confirm a desired answer — the LLM equivalent of p-hacking).

事前聲明搜尋計畫，防止為了得到想要的答案，事後挑選搜尋詞。

Phase 1: Scope Reconnaissance（範圍偵察）— optional

Skip this phase if the T1 source is obvious (e.g., TSMC investor relations for TSMC quarterly revenue).

Run this phase if the candidate set is unclear (e.g., "which Taiwanese players are in 2026 MLB").

# Detect Gemini availability
which gemini >/dev/null 2>&1

• If Gemini available + scope unclear: invoke Gemini as scout

  echo "List all candidate X for question Y. Output names only, with sources." | gemini

Use Gemini's output as a candidate list, not as answers. Each candidate must be verified in Phase 2.

• If Gemini unavailable: use broad WebSearch with multiple query variations.

Phase 2: T1/T2 Verification（驗證）— REQUIRED

For each factual claim, fetch the T1/T2 source directly. Do not synthesize from search snippets.

• Use WebFetch to retrieve the actual source page
• Cross-reference at least 2 independent sources for high-stakes claims
• Apply Citation Chain Integrity, Wire Service Detection, Time Context Decay rules

#### T1 Domain Gate — REQUIRED before labelling any source T1（強制 T1 域名規則）

Before any source can carry a T1 label in the final output, its domain must match the appropriate category below. If it does not match, the source is T2 at best, and the citation chain must be traced upstream to a real T1 document before Phase 2 is considered complete.

News outlets are T2 at best, even when they appear authoritative. This includes:

• Wire services: cna.com.tw (中央社), reuters.com, ap.org
• Public broadcasters: pts.org.tw (公視), bbc.com
• Newspapers / magazines: udn.com, ltn.com.tw, chinatimes.com, storm.mg, newtalk.tw, nytimes.com, economist.com

When a news article cites a primary document (e.g., a court ruling, a legislative bill, a corporate filing), Phase 2 must follow the citation chain back to that primary document before finalizing. If the upstream T1 cannot be located, the news source can be cited but must be tier-labelled T2 in the verification output.

Audit step before output: Re-read every "T1" label in your draft. For each one, ask: "Does this URL's domain appear in the table above?" If no, downgrade or chase the upstream document.

⚡ Parallelization rule: When fetching multiple independent sources (different candidates, different domains, different aspects), issue all WebFetch / WebSearch calls in parallel in a single tool-use batch. Sequential execution wastes wall-clock time for no benefit. Only serialize when one fetch's result determines the next fetch's URL.

Example: Verifying 4 baseball players → 4 parallel WebFetch calls, not 4 sequential.

The Iron Law applies here: if the question requires data past your training cutoff, you MUST execute the search, not defer to the user.

Case study: See examples/case-citation-laundering.md — a 2026-05-11 failure where udn.com was mistakenly labelled T1, leading to five factual errors in a legislative bill comparison. The T1 Domain Gate was introduced specifically to prevent this recurrence.

Phase 3: Disclosure Output（揭露式呈現）

#### Output Structure — Inverted Pyramid（倒金字塔）

Order content by user importance, not by phase order:

1. Conclusion / TL;DR             ← 2-3 sentence direct answer
2. (If question was reframed)     ← "我把問題改成這樣問" — show original vs reframed
3. Main content                    ← Structured details
4. Caveats / risks (if applicable) ← Real warnings only
5. Blind spots / additions         ← What might be missed
———— horizontal rule ————
6. Verification details            ← Sources / falsifiability / reproducibility

Question reframing block (when applicable):

If Phase 0.3 reframed the question, surface this near the top of the output. The reframing is itself part of the answer — it teaches the reader to detect smuggled premises. Format:

原本問題：[user's original phrasing or news headline]
重新框定為：[the reframed, premise-explicit version]
為什麼：[which premise was shaky and how the rewrite makes it explicit]

If the original question already had clean premises, omit this block (don't perform reframing as decoration).

#### Confidence Labels — Default to Verified, Label Exceptions Only

Estimative Language (from intelligence analysis):

Rule: Do NOT label every section. Verified = silent default. Only label exceptions, so labels become signal, not decoration.

#### Phase Numbers Are Internal — Never Expose to User

The user does not care about Phase 4 / Phase 5. Translate internal phase outputs into user-facing language:

Phase numbers in output = leaked implementation details = bad UX.

#### Verification Metadata — Grouped at End

Use a horizontal rule (---) to separate the answer from the verification trail. Group all metadata under a single heading:

---
## 查證細節

**來源 (Sources)**: [list with tier annotations]
**信心摘要**: [list only the non-default exceptions]
**反證條件 (Falsifiability)**: [what evidence would change this answer]
**搜尋紀錄 (Reproducibility)**: [terms, URLs, retrieval date]

This pushes metadata below the fold visually, while keeping it accessible for users who want to verify.

#### Falsifiability ↔ Phase 5 Relationship

The Falsifiability statement is the passive description ("If X were true, my answer would be wrong"). Phase 5 (when triggered) actively searches for whether X is in fact true. Phase 5 turns the falsifiability statement into investigation results.

#### Privacy / PII Discipline — REQUIRED before publishing or sharing output

Source-Trail produces analyses that are intentionally shareable: case files, examples directories, social-media posts, GitHub repositories. Before any verification output is committed to a publicly-discoverable surface, run this gate.

##### The two-tier rule

##### When in doubt, redact

The cost of over-redaction is small (slightly less colorful case studies). The cost of over-naming is borne by someone other than the Skill author and is permanent once indexed. Default to redaction.

##### What counts as identifying information

• @handles on social platforms (Threads, X, Instagram, etc.)
• Real names of non-public individuals
• Direct URLs to individual posts or profiles
• Workplace + role + city combinations that practically identify someone
• Distinctive paraphrases that would be searchable back to the individual

##### When the claim is part of the analysis

If the goal is to analyze a class of incorrect claim (e.g., "social-media commenters often conflate central and local government budgets"), describe the class and quote one representative claim without attribution. Methodology lessons survive redaction; only the name is removed.

##### Public-platform availability is not a license

A Threads post or X reply being publicly viewable does not authorize lifting that speech into a permanent indexed methodology document attached to a verdict. Contextual integrity (Nissenbaum 2004) — the appropriateness of information flow relative to its original context — applies even to nominally public speech.

##### Audit step before commit / push

Before git commit or any publish action on output that will become discoverable:

1. Search the draft for @, real-name patterns (e.g. CJK personal-name conventions), and platform-specific identifiers
2. For each hit, ask: is this person a public figure speaking in a public-conduct role as it relates to this claim?
3. If yes — keep. If no — redact and replace with a role description.

Case study: See examples/case-privacy-leak.md — a 2026-05-11 failure where six private individuals' Threads handles were committed to the public Skill repository alongside verdicts of "factually wrong" on their comments. The Privacy / PII Discipline was added to Phase 3 in direct response.

Phase 4: Peer Review（同儕審查）— conditional

Run when:

• Stakes are Medium or High
• Answer involves subjective judgment ("最強", "最佳")
• I am highly confident — high confidence warrants higher scrutiny

Skip when:

• Pure factual lookup at Low stakes
• Multiple independent T1 sources already converged

Execution:

• If Gemini available: submit answer to Gemini with prompt "你是同儕審查者，請找出此回答的盲點。"
• If Gemini unavailable: Claude self-roleplays as adversarial critic.

Critical filtering rule: Gemini's review will contain valid methodological critiques AND its own fabricated "specifics" (made-up version numbers, draft formats, etc.). Absorb the methodological critique, treat any new specific claims as unverified — must re-verify in Phase 2 if used.

Phase 5: Devil's Advocacy（敵對審查）— conditional

Relationship to Phase 3 Falsifiability:

• Phase 3 Falsifiability = passive description ("If X were true, my answer would be wrong")
• Phase 5 Devil's Advocacy = active search ("Let me actually try to find evidence that X is true")

Phase 5 turns the falsifiability statement into an investigation.

Run when:

• Stakes are High (decision support, harm if wrong, contested claims)
• Comparative claims with multiple plausible answers ("最", "排名", "排行")
• The conclusion is counterintuitive or surprising
• The answer involves prediction or extrapolation
• Politically / socially contested topics

Skip when:

• Stakes are Low and Phase 2 produced 2+ independent T1 sources agreeing
• Answer space is closed and well-bounded (e.g., "list of currently active X")
• The Falsifiability statement in Phase 3 has no plausible trigger conditions worth searching

Execution — each step must produce a written artifact, not a self-described "consideration":

1. Find counter-evidence: What evidence would disprove this conclusion? Search for it actively.

Required artifact: a written summary of what was searched and what was found / not found.

1. List competing hypotheses (ACH) — REQUIRED WRITTEN TABLE

Phase 5 cannot complete without producing an ACH table containing at least 2 hypotheses. The table format:

Hard requirement: the table must contain H1 + at least one non-trivial alternative. If no plausible alternative exists after honest search, the table must explicitly state: "No alternative hypothesis with non-trivial fit to the evidence was found" — making the absence a deliberate finding, not an oversight.

Why this is mandatory: without a written ACH, the analyst's preferred hypothesis becomes the only hypothesis defended in the output. The reader cannot tell whether alternatives were considered or silently skipped. See examples/case-skipped-ach.md for the 2026-05-11 failure that motivated this hardening.

1. Indicators & Warnings: Where am I structurally most likely to be wrong for this question type? (Sports? Stats lag. Tech? Versions shift. Politics? Wire bias.)

Required artifact: a written one-line note per question-type category that applies.

⚡ Parallelization rule: Counter-evidence searches across different angles are independent — issue all in parallel.

Output surfacing: when Phase 5 runs, the ACH table must surface in the Phase 3 disclosure output (visible to the user, not hidden as internal metadata). A report that shows only the favored hypothesis invites the reader's natural question "what else might this be?" — which is the very question Phase 5 exists to answer in advance.

This is more proactive than Phase 4 — it doesn't wait for an external reviewer.

Phase 6: Belief Source-Trail（信念溯源）— conditional

The complement to fact source-tracing. Phases 0–5 verify what the fact is. Phase 6 explains why a different, incorrect belief became plausible enough to spread — without dismissing the believer.

#### Why Phase 6 exists

A factually correct verification that ignores why the false belief took root reads as condescension. Readers naturally respond: "If you're right, why does everyone else believe the wrong thing — are they all stupid?"

The answer is almost always no. False beliefs spread because structural conditions — knowledge gaps, narrative frames, media language, group identity, cognitive load — make the incorrect interpretation the natural near-guess under the listener's actual information environment.

Believing the wrong thing is often rational given the conditions; correcting it requires first acknowledging those conditions. Phase 6 puts the same rigor on belief formation that Phases 0–5 put on fact formation. The user phrased the principle as: 「人們的情緒跟焦慮是真實的，要先被同理，才能夠看得下溯源的證據」— emotions and anxieties are real; they must be empathized with before someone is ready to read the source trail.

#### Run when

• The false-belief version of the question circulates on social media with measurable reach
• The conclusion is going to be published (case file, post, public output) — readers will need this context to absorb the verification rather than reject it
• High-stakes questions where simply asserting the truth has historically failed to shift belief
• The verified conclusion contradicts a politically or emotionally invested framing

#### Skip when

• Low-stakes lookup with no false-belief alternative in circulation
• Pure technical / scientific questions with no social dimension
• The verification is being delivered to a single audience already aligned with the conclusion

#### Execution — required written artifact

Produce a Structural Factor Analysis table. The table must satisfy these constraints:

1. At least 3 factors. Single-factor explanations are reductive and almost always wrong.
2. At least 2 distinct factor categories from this list:

• Cognitive — knowledge gaps, memory conflation, attribute substitution, default heuristics
• Political — pre-existing narrative frames, identity-based attribution, partisan priors
• Media — headline ambiguity, wire copy propagation, framing language
• Propagation — what's shareable on a given platform, attention economy, simplification pressure
• Psychological — emotional resonance, in-group/out-group dynamics, anxiety amplification

1. Empathy frame required: explicitly state in one sentence the reading under which believing the false version was a rational near-guess given the conditions. If no such reading exists, the false belief is malicious rather than structural, and Phase 6 should output that finding directly.
2. No single-attribution verdicts. Do not say "people believe this because they are [political tribe]." That collapses Phase 6 into political commentary.

##### Format

為什麼這個誤解會流傳：結構性因素分析

| 因素 | 性質 | 對本案的作用 |
|---|---|---|
| A. [factor name] | [category] | [one-sentence mechanism] |
| B. ... | ... | ... |
| C. ... | ... | ... |
| ...至少 3 項，至少 2 個 category | | |

同理框架：在這些條件聚合下，相信「[false belief verbatim]」是一種理性的近似猜測，而非愚蠢或惡意。

#### Anti-patterns

• Single-cause reduction: "People believe X because of media bias / political bias / education." Real social phenomena are multi-factorial.
• Veiled contempt: "People can't tell the difference between special budget and county budget." Rephrase as a structural condition, not a personal deficit.
• Both-sidesing as an escape: Phase 6 is structural analysis, not a "both sides have a point" hedge. Be specific about which conditions are doing the work.
• Politicizing the analysis: If the structural factors mostly favor one political reading, the analyst is bringing their own framing. Stress-test by trying to write the same factors with the political valence reversed.
• Defensive scope-clarification disguised as both-sidesing: When the verified conclusion happens to be politically convenient for one side, do not write a closing paragraph like "but [the favored side] can still be examined in many other ways" with a list of unverified angles. Even if each listed angle is specific, the function of the paragraph is to perform political symmetry through implied accusation — exactly what Phase 6 must avoid. The correct form is a neutral Scope Statement: "Verified: [what]. Not in scope: [what other questions exist, without implying their truth status]. Do not infer the status of out-of-scope questions from this verification." See the Scope Statement format below.

#### Scope Statement format — when the verification is politically asymmetric

When a verified conclusion lands on one side of a political divide (e.g., "this specific accusation against politician X is factually wrong"), the analyst will feel a pull to compensate. Resist it. The correct discipline is:

本次查證的範圍（Scope Statement）

已驗證：
- [the specific factual question that was answered]
- [the conclusion]

本次未涵蓋的問題（不在本次 Source-Trail 範圍）：
- [other plausible questions about the same actors/topic — listed
  NEUTRALLY as research questions, with NO implication of their
  factual status]

請勿從本次查證推論任何未涵蓋問題的事實狀態。未涵蓋 ≠ 已驗證 /
已否證 / 結論為何。每一個未涵蓋問題若要驗證，需要另外跑一次
Source-Trail。

Key distinctions:

• "Not in scope" ≠ "could be examined" — the former is neutral research-design language; the latter implies the analyst is curating which unverified accusations the reader should remember
• No "this is not exoneration" / "not endorsing anyone" hedge — if the verification was done correctly, this disclaimer is unnecessary. Including it reveals analyst anxiety about being misread, and that anxiety telegraphs to the reader.
• Explicit warning: "Do not infer the status of out-of-scope questions from this verification" — this is the academic-paper standard for limitations sections.

#### Disclosure surfacing

When Phase 6 runs, the structural factor analysis must appear in the published output. It is not internal metadata — it is the empathy layer the reader needs in order to integrate the verification rather than reject it.

Place it after the factual conclusion and before the verification metadata (sources, falsifiability). The reader's progression should be:

1. Conclusion (what's true)
2. Evidence (why it's true)
3. Phase 6 — Belief Source-Trail (why a different version was plausible)
4. Verification trail (T1 sources, ACH, etc.)

#### Relationship to other phases

• Phase 0.3 surfaces the false premise in the question
• Phase 2/5 verify the actual facts
• Phase 6 explains why the false premise itself was natural — closing the loop between the question and the reader's prior beliefs

Phase 6 is what makes Source-Trail's output not just correct but receivable.

Disclosure Format Examples

❌ Bad: Training-data Synthesis（訓練資料合成）

Faker won the 2024 World Championship. He's averaging 4.3/1.5/3.4 KDA in the 2026 LCK Spring split, leading T1 to first place.

Problems: no sources, specific numbers without verification, mixes verified history with unverified current data, no confidence signal.

❌ Bad: Over-labeled Verification Dump（標籤氾濫）

【已驗證】 - Faker is on T1's 2026 roster

>

【已驗證】 - T1 won 2026 LCK Cup

>

【已驗證】 - T1 won Season Opening

>

Phase 4 — Self peer review: Things I might have missed... Phase 5 — Devil's advocacy: Counter-evidence... Reproducibility Log: ... Falsifiability: ... 底線: Faker remains top-tier.

Problems: every fact labeled (signal becomes noise), Phase numbers leaked to user, conclusion buried at the end after metadata.

✓ Good: Inverted Pyramid（倒金字塔結構）

Faker 仍是頂尖韓國中路選手，但「最強」沒有共識答案。 韓國社群普遍把 Faker（國際賽/關鍵球）和 Chovy（個人對線/例行賽）並列在最頂層，Zeka、Showmaker 也仍在前段。

>

## 2026 季前賽戰績 - T1 在 2026 LCK Cup 小組賽全勝開局 - 拿下 2026 LCK Season Opening，Faker 在開幕對 HLE 一役獲選 MVP - 個人累積戰績：2025 世界賽奪冠後成為史上奪冠最多的選手

>

## 「最強」的不同定義 - 國際賽 / 關鍵時刻 → Faker - 個人對線 / carry 能力 → Chovy（Gen.G）長期被視為更強 - 純機械操作 → 年輕選手如 Zeus、Oner 在某些指標上更高

>

## 可補充的觀點 隊伍體系的加成沒有單獨拆出 — Faker 的穩定性受惠於 Keria 遊走和 Oner 控圖；如果單看「最強單兵」可能會給出不同答案。

>

⚠ 【中信心】：場均數據（4.3/1.5/3.4 KDA）目前只查到 Sheep Esports 一個來源，未交叉驗證。

>

⚠ 【不採信】：搜尋結果有摘要稱 "Faker 拿下 2026 MSI MVP"，但 MSI 2026 尚未開打，是搜尋引擎自動摘要的幻覺，已排除。

>

--- ## 查證細節

>

來源: - T1: T1 official site, Liquipedia - T2: Inquirer Esports, Sheep Esports, Taipei Times

>

反證條件: 此答案會有誤，若 (a) Faker 自最新來源日後被冷凍，(b) Sheep Esports 統計有誤。

>

搜尋紀錄: "Faker T1 LCK 2026 Spring performance"、"Faker 2026 LCK MVP best mid laner Korea" / 查證日期 2026-05-08

Note: verified facts presented as flowing prose without labels. Only the non-verified exceptions carry warning labels (⚠). No "Phase 4" / "Phase 5" leaked to user — translated to "可補充的觀點". Conclusion at top, metadata below the fold.

Gemini's Dual Role

Gemini is a tool, not a source. It serves two distinct functions in this workflow:

A. Scout（偵察兵）— Phase 1

• Strength: Google search integration → broader coverage of recent events
• Use: Identify candidates for "who/what" questions when scope is unclear
• Output treatment: Candidate list to verify, NOT final answer
• Example: "Which Taiwanese players are in 2026 MLB?" → Gemini lists names → Claude verifies each via T1

B. Peer Reviewer（同儕審查者）— Phase 4

• Strength: Different model, different blind spots → catches what Claude missed
• Use: Critique a draft answer for methodological gaps
• Output treatment: Absorb methodological critique; re-verify any new "specifics" Gemini introduces
• Example: Gemini may correctly say "your KDA argument is shallow" while incorrectly fabricating "version 16.x" — keep the first, discard the second

Common Mistakes（常見錯誤）

• ❌ "Gemini agrees with me, so the answer is right" — same training corpus = not independent
• ❌ "Gemini gave specific numbers, those are facts" — Gemini hallucinates specifics confidently
• ❌ Conflating roles — using peer review output as scout output, or vice versa

Red Flags / Common Mistakes

Phase Selection Decision Tree

Use this to decide which phases to run. Phase 0, 0.3, 0.5, 2, 3 always run. Others depend on stakes and scope.

Question received
       ↓
[Phase 0] Is this factual?
   └─ No → Exit skill
   └─ Yes ↓
[Stakes Assessment] Low / Medium / High?
       ↓
[Pre-flight Announcement]
   └─ Low/Medium → announce + execute
   └─ High       → announce + WAIT for user confirmation
       ↓
[Phase 0.3] Key Assumptions Check (always)
   - List premises (existence, actor, time, metric, stance, substitution)
   - Challenge each
   - Reframe question if a premise is shaky
       ↓
[Phase 0.5] Pre-register strategy (always)
       ↓
[Phase 1] Is scope unclear?
   └─ Yes → Gemini scout (or broad WebSearch)
   └─ No  → Skip
       ↓
[Phase 2] T1/T2 fetch (always REQUIRED)
       ↓
[Phase 3] Disclosure output (always)
   - Estimative language labels
   - Reproducibility log
   - Falsifiability statement
       ↓
[Phase 4] Peer review?
   - Run if: Medium/High stakes, subjective judgment, high confidence
   - Skip if: Low stakes + 2+ independent T1 sources agreed
       ↓
[Phase 5] Devil's advocacy?
   - Run if: High stakes, comparative claims, surprising conclusion
   - Skip if: Low stakes + closed answer space + strong T1 verification
       ↓
[Phase 6] Belief Source-Trail?
   - Run if: false-belief version circulates on social media; output will be
     published; politically/emotionally invested framing
   - Skip if: low-stakes lookup; no social dimension; aligned audience only
       ↓
Final answer

Quick lookup: which phases by question type

Note on news / current events: Phase 0.3 is especially critical here — news headlines almost always arrive pre-framed by an editor's assumptions. Skipping Phase 0.3 means inheriting those assumptions silently.

Note on social-media-originated questions for public output: Phase 6 (Belief Source-Trail) is required, not optional. Publishing a fact-check without explaining why the false belief was plausible reads as condescension and gets rejected — even when the facts are correct.

Advanced Techniques (Appendix)

Optional refinements for high-stakes verification:

Right of Reply（給被指控方回應的機會）

For comparative claims ("X is better than Y"), actively seek Y's defenders' counterargument before publishing.

Conflict of Interest Tagging

If a source is the party-of-record (company press release about itself, paid promotion), downgrade one tier and flag.

Hierarchy of Evidence

Direct observation > Reported observation > Aggregated analysis > Individual opinion. T1 source quality varies by this hierarchy.

Convergent Evidence

Beyond multi-source: multi-method convergence is stronger. Official announcement + third-party statistics + on-the-ground reporting all pointing to same conclusion = high confidence.

Indicators & Warnings

Pre-list "where am I most likely to be structurally wrong" for this question type. Sports? Stats websites lag. Tech? Version numbers shift. Politics? Wire bias.

Chain of Custody

Information from original to presentation must be traceable at every step. No "I read it somewhere" — track the path.

Hearsay vs Direct Testimony

"X said that Y said Z" is weaker than "Y directly said Z." Prefer direct quotes over reported quotes.

Exhibit > Testimony

Original documents (T1) outweigh expert opinion (T3). Link the document, don't paraphrase the expert.

Real-World Cases

See examples/ directory:

• case-baseball.md — Phase 1 Gemini scouting succeeds where solo WebSearch fails (2026 Taiwanese MLB players)
• case-tsmc.md — Phase 1 skipped when T1 source is obvious (TSMC quarterly revenue)
• case-faker.md — Phase 4 peer review's double-edged nature (Gemini's critique contains hallucinations to filter)
• case-phase-0-3-genesis.md — How Phase 0.3 was added to the skill: the meta-case of using Source-Trail to verify Source-Trail's own academic foundations (CIA KAC, Popper, Maxwell, Kahneman)
• case-citation-laundering.md — When news outlets are mistakenly labelled T1: a 2026-05-11 failure that introduced the T1 Domain Gate rule into Phase 2
• case-political-substitution.md — When a viral question smuggles a false Authority premise: a 2026-05-11 case that hardened Phase 0.3 from declaration-based to execution-based
• case-privacy-leak.md — When methodological rigor is not the same as ethical adequacy: a 2026-05-11 PII failure in the Skill's own published examples that introduced the Privacy / PII Discipline rule into Phase 3
• case-skipped-ach.md — When Phase 5 is run but ACH is silently skipped: a 2026-05-11 failure where the analyst defended only the favored hypothesis until a user reviewer surfaced a plausible alternative. Hardened Phase 5 ACH from "listed step" to "required written table."
• case-belief-source-trail-origin.md — Why Phase 6 was added: factually correct verifications without an empathy layer get rejected. The complement to fact source-tracing — explaining why a false belief became plausible enough to spread, without dismissing the believer.

Origin

Distilled from a 2026-05-08 working session on epistemic humility in LLM-assisted research. Each rule traces to a documented failure mode caught during testing. See examples/ for the original cases.

Version history

• v1.0 (2026-05-08) — Initial release. Phases 0, 0.5, 1, 2, 3, 4, 5. Origin from baseball/Faker/TSMC failure cases.
• v1.1 (2026-05-11 AM) — Added Phase 0.3 (Key Assumptions Check). Triggered by recognizing that Phase 0.5 (pre-registration) assumes the question is already well-formed; if the question smuggles a false premise, no rigorous searching produces a true answer. Adapted from CIA Tradecraft Primer's KAC procedure. See examples/case-phase-0-3-genesis.md.
• v1.2 (2026-05-11 PM) — Hardening release in response to two same-day failures during real-world testing:
• Added T1 Domain Gate to Phase 2: a source cannot be labelled T1 unless its domain matches an explicit whitelist; news outlets are T2 at best regardless of authority. Triggered by mis-labelling udn.com as T1 in a legislative-bill comparison, which led to five factual errors. See examples/case-citation-laundering.md.
• Phase 0.3 substitution check is now execution-based, not declaration-based: each of the four KAC steps must produce a written artifact; missing artifact halts the run. Added "Authority / Jurisdiction" as a distinct premise category. Triggered by a viral Threads post smuggling a false Authority premise that survived initial Phase 0.3 inspection. See examples/case-political-substitution.md.
• v1.3 (2026-05-11 evening) — Ethics release. The v1.2 push committed six private individuals' Threads @handles into the public Skill repository alongside "factually wrong" verdicts on their specific comments. The Skill author caught this on review and required a privacy discipline that should have been built in from the start. Added Privacy / PII Discipline as a Phase 3 sub-rule, with a two-tier rule (public figures named directly; private individuals redacted). Retroactively redacted examples/case-political-substitution.md. Documented the failure as examples/case-privacy-leak.md. The Skill author also chose to rewrite git history (force-push via git filter-repo) to remove the original handles from all prior commits, rather than the lower-effort forward-only repair — because the cost of leaving private individuals' identifiers in indexed historical commits is borne by them, not by the Skill author.
• v1.4 (2026-05-11 night) — Phase 5 hardening. During the same review of the Hualien 300 億 case that surfaced v1.3, the user pointed out that the Skill's report defended only the favored hypothesis (馬太鞍堰塞湖 300 億 special budget) without surfacing a plausible alternative (0403 震災款 285.5 億, which could be the actual referent if rounded up colloquially). Phase 5 had listed ACH (Analysis of Competing Hypotheses) as a sub-step, but the actual run had skipped it — the same "described but not executed" failure pattern that motivated Phase 0.3 hardening in v1.2. Phase 5 ACH is now a required written table with at least H1 + one non-trivial alternative, surfaced in the Phase 3 output rather than hidden as internal metadata. See examples/case-skipped-ach.md.
• v1.5 (2026-05-12) — Empathy layer release. The user observed that Phases 0–5 produce factually correct verifications but leave a gap: why do so many people believe the wrong version? Without addressing this, a fact-check reads as condescension — readers naturally respond "if you're right, why does everyone else believe wrong, are they all stupid?" The answer is almost always no: false beliefs spread because of structural conditions (knowledge gaps, narrative frames, media language, identity-based attribution, propagation pressure, cognitive load). Added Phase 6: Belief Source-Trail — the complement to fact source-tracing. Required output is a Structural Factor Analysis with at least 3 factors across at least 2 distinct categories, plus an explicit empathy frame stating that believing the false version was a rational near-guess under the conditions. The user's framing: 「人們的情緒跟焦慮是真實的，要先被同理，才能夠看得下溯源的證據」. See examples/case-belief-source-trail-origin.md.
• v1.5.1 (2026-05-12) — Same-day patch. The user reviewed cases/001-hualien-300-billion.md and flagged that the closing paragraph ("but [the politician] can still be examined in many other ways: [list of unverified angles]") read as both-sidesing — defensive scope-clarification disguised as political symmetry. Even though each listed angle was specific, the function of the paragraph was to perform balance through implied accusation, which is exactly the failure mode Phase 6 is supposed to prevent. Added anti-pattern #5: Defensive scope-clarification disguised as both-sidesing and a required Scope Statement format that uses neutral inclusion/exclusion language without implying the truth status of out-of-scope questions. Drawn from academic limitations-section conventions and Boykoff & Boykoff's (2004) false-balance analysis.