Methodology v0.6

0. Verdict-Writing Principles — These Come Before Every Other Clause

Chamgap is an information provider. We do not make decisions for readers. Decisions belong to readers and, when needed, medical or legal professionals. This principle protects us legally, because factual reporting is harder to treat as defamation; it is the source of trust, because a referee without a position is trusted most; and above all, it lets each reader make the right decision in their own context.

1) Answer the reader's decision question first

A verdict first answers the reader's question: "Should I buy this, not buy this, or keep taking what I already take?" The academic evidence comes next. Conclusion first, then basis.

2) Report facts, but do not direct behavior

Stop the sentence at "what the research shows." Do not use behavioral commands or value judgments such as "buy," "do not buy," "not worth the money," "exaggerated," "effective," or "ineffective."

3) Be clear, but do not command

Do not blur the conclusion. Make it clearer than advertising. But make it a conclusion of fact, not an order for action. "No judgment" must not be used as an excuse for ambiguity; that is avoidance of responsibility, not information provision.

1. Verdict Unit — Grades Attach to Claims, Not Ingredients

Claim = ingredient x effect x dose (and duration)

"Collagen is good for skin" is not a judgeable sentence. "Low-molecular collagen peptides, 2.5-10 g/day, taken for at least 8 weeks, improve skin hydration and elasticity" is a judgeable claim. If one ingredient has multiple claims, each is judged separately: collagen x skin and collagen x joints are separate verdicts. Verdicts are written at the claim level, not for a specific product, and when advertising wording is quoted, screenshots and collection dates are stored with it.

2. Effect Grades

Grade Adjustment Factors (Stated in Every Verdict)

• Industry-funding bias — The share of manufacturer-funded studies among studies on the claim. If high, review for one-step downgrade.
• Publication bias — Signs that only positive results were published, such as a field dominated by small studies or absence of preregistration.
• Effect size — If the effect is statistically significant but too small to feel, the verdict says so plainly: "There is an effect, but only at a ruler-measured level."
• Dose gap — The gap between the dose used in studies and the actual content of products sold in Korea. Even if the evidence is A, if marketed products contain half the studied dose, that fact goes at the top of the verdict. We have not yet found a Korean-language system that applies this consistently as a verdict framework; if we find one, we will link it here.

2-1. Grade Boundary Rules v0.5 · learned through rejudgment

Boundaries that repeatedly cause disagreement are written explicitly, as revealed by the rejudgment procedure in section 6-2.

• ① Primary endpoint first (B/C boundary). The grade is based on the result of the RCT's prespecified primary endpoint. If the primary endpoint is nonsignificant and significance appears only in secondary, post-hoc subgroup, or surrogate endpoints, the maximum grade is C. (Example: lutein — AREDS2 primary main analysis HR 0.90, P=.12 nonsignificant -> C.)
• ② Independence first (C/D boundary). Even when a meta-analysis is significant, if the largest or conflict-free independent RCT is null, or if significance disappears in the high-quality subset, the grade is D. (Examples: policosanol and garcinia -> D.)
• ②-b Absence of independent replication (B/C boundary). If there is no independent RCT proven null, but all positives are concentrated in manufacturer studies and independent replication is absent, the maximum is C, not D. There is a human signal (therefore not D), but it has never been independently reproduced (therefore not B). "Null" means D; "absence" means maximum C. (Examples: hyaluronic acid and ceramides -> downgraded from B to C.)
• ③ Aggregating composite claims. If several endpoints are bundled into one claim, score each endpoint separately. If there is an ingredient with no evidence at all, review for a half-grade downgrade and state that fact separately. (Example: milk thistle — liver health is C, fatigue has 0 evidence -> stated.)
• ④ Regulatory recognition is not a bonus point. MFDS functional recognition is not an upward factor; it is recorded separately as neutral metadata (it is ingredient/label regulation, not a guarantee of effect or clinical benefit).

2-2. Score Shown Beside the Grade v0.6 · 0-100

Grades (A-F) are bands, and position within the band is shown with a 0-100 score, for example: "Grade C · 47." The grade is the broad classification; the score shows whether the evidence sits closer to the lower or upper end of that grade.

• A 80-100 · B 60-79 · C 40-59 · D 20-39 · F 0-19 · ? unjudgeable (no score)

The score is assigned within the band by combining evidence strength: existence and size of human RCTs, independence, primary-endpoint success, effect size, and consistency of results. It is not precise measurement; it is a direction marker inside the grade (47 and 49 are practically the same), and the basis is written in the verdict. After v0.6, this applies from new verdicts onward.

3. Safety Labeling (A Separate Axis from Effect Grade)

• Acceptable — Serious adverse reactions are rarely reported at studied doses.
• Caution — Caution is needed under specific conditions, such as pregnancy, drug combinations, liver disease, or kidney disease. State the condition.
• Warning — Serious adverse reactions have been reported. State the content.
• Unknown — Long-term safety data are insufficient.

4. Verdict Procedure — 7 Steps, All Public

• ① Claim collection — Quote actual advertising and product-page wording verbatim (screenshots and collection dates stored). Do not create straw men.
• ② Literature search — Record the search source (PubMed, etc.), search formula, and search date in the verdict. Anyone must be able to reproduce it.
• ③ Screening — Prioritize human studies, RCTs, and meta-analyses. Leave reasons for excluded studies.
• ④ Draft — AI (Claude) writes the draft. We do not hide this fact.
• ⑤ Approval and signature — The Chamgap Editorial Team directly checks the evidence list, approves it, and signs the verdict under the editorial team's name. No verdict is published before approval.
• ⑥ Publication and correction — Open an error-reporting channel. When correcting, leave the original text with strikethrough and attach the correction note and date.
• ⑦ Objection — A business with an interest in the claim under verdict may file an objection with supporting literature. We reply with review results within 14 days after receipt; valid objections are handled through correction, and rejected objections are recorded at the bottom of the verdict with reasons.

5. Conflict-of-Interest Policy

We enter into no financial relationship related to a verdict target, including sales, affiliate arrangements, advertising, manuscript fees, and sample receipt. Whether a company is a customer of the business monitoring service has no effect on public verdict topic selection or grades. If a verdict target overlaps with a monitoring-service customer, that fact is disclosed at the top of the verdict.

6. AI Use Disclosure

We use AI for draft writing and literature organization. Because AI can fabricate sources, all cited literature is checked by a human for real existence and content match before approval. Literature not yet checked is marked [verification needed], and no verdict with that mark remaining can be published.

6-1. Failure-Pattern Atlas (Learning Loop)

We accumulate the "ways information goes wrong" filtered out during verification. This is not a list of answers but a list of questions and traps. It sharpens verification for each verdict but does not replace verification. The verification stage for a new verdict reads this atlas as a checklist, and newly filtered patterns are added to the atlas. Examples: "disguised no-funding," "expanding surrogate markers into felt effects," "repurposing absorption proof as efficacy proof," and "formulation mixing (oral, buccal mucosa, patch)." Rebuttals in the correction ledger correct the atlas in reverse: over time, verification becomes faster and more accurate.

6-2. Verification and Rejudgment Procedure v0.3

Every verdict goes through the process below. This process creates trust in each grade.

• ① Draft — AI searches and synthesizes literature and proposes a draft grade.
• ② Two blind cross-checks — Two independent auditors who do not know our grade each open the citations in the original text, check them, and assign their own grade.
• ③ Classification — Concordant (both the same) / review (only one differs) / split (both differ).
• ④ Split-grade rejudgment (dialectic · alternating heterogeneous models) — Make the higher-grade and lower-grade positions compete by having different agents each newly find the real evidence. Codex (a different model family) and Chamgap's AI alternate verification, each taking over and checking the other's judgment, and the editorial team makes the final decision.
• ⑤ If ambiguous, run one more round; if it does not narrow, do not force it — If the decision remains ambiguous, run cross-checking once more. If the grade still does not narrow, do not squeeze out a single grade by force. Publish it as a "split-evidence verdict" with both sides and their reasons side by side.
• ⑥ Boundary learning — Ambiguities in grade definitions revealed by rejudgment are accumulated as methodology clarification in section 2-1.

The "Cross-check" item in each verdict discloses which path that verdict followed in this procedure: concordant, rejudgment downgrade, or adjustment.

7. Limitations

Every verdict is based on literature published up to the approval date, and grades can change when new research appears. Nothing here is medical advice. Supplement decisions should be discussed with a physician or pharmacist.

Methodology Change History

• v0.2 (2026-07-05) — Added Chapter 0, "Verdict-Writing Principles." Verdicts answer the reader's decision question first (conclusion -> basis), report facts without directing behavior, and remain clear without commanding.
• v0.3 (2026-07-06) — Added section 6-1 Failure-Pattern Atlas and section 6-2 Verification and Rejudgment Procedure. Draft -> two blind cross-checks -> dialectical rejudgment for splits. Forced conclusions prohibited.
• v0.4 (2026-07-06) — Added section 2-1 Grade Boundary Rules: primary endpoint first, independence first, composite-claim aggregation, and regulatory recognition as neutral.
• v0.5 (2026-07-06) — Added section 2-1 rule ②-b, "Absence of independent replication." Learned while filtering verdicts 013-032 twice more with a heterogeneous model (top-tier Codex): if an independent RCT is "null," D; if it is "absent," maximum C.
• v0.6 (2026-07-06) — Added score display beside A-F (0-100) in section 2-2, plus alternating cross-checking and side-by-side publication of unresolved splits in section 6-2. Verification alternates between Codex and Chamgap AI; if the decision does not narrow, it is not forced and the split is published as-is. First applied in verdict 051 (aquaporin).
• v0.1 (2026-07-04) — Initial version. (Objection procedure added on 2026-07-05.)