Corroboration Standard

To be eligible as a Corroboration Reviewer, an individual must satisfy all of the following:

• Active professional licensure or academic position: Currently licensed in the relevant professional domain in at least one jurisdiction (for licensed professions), OR holds a current academic appointment at an accredited research institution in the relevant field (for scientific and technical domains).
• No active disciplinary proceedings: No pending disciplinary action, suspension, or investigation by any licensing authority or professional body, confirmed at initial enrollment and annually thereafter.
• Minimum qualifying experience: At least three years of post-licensure professional practice OR two years of post-doctoral academic work in the relevant domain. Less experienced practitioners may participate as Provisional Reviewers under supervision (see CS-RE-003).
• Completion of Corroboration Standard training: Completion of the Foundation's mandatory reviewer orientation covering the information quality vs professional services distinction, blind review protocol, scoring rubric, and conflict-of-interest obligations. Training must be repeated annually.
• Signed Reviewer Agreement: Execution of the Foundation's standard Reviewer Agreement acknowledging the scope and limitations of the Corroboration Review role.

Practitioners with 1–3 years post-licensure experience may participate as Provisional Reviewers subject to: supervision by a paired Senior Reviewer (5+ years experience); Provisional reviews count toward but do not independently complete corroboration thresholds; Provisional reviews are labeled as such in the system; Provisional reviewers must complete additional training modules covering areas of particular risk for early-career practitioners. This pathway ensures the reviewer pool is not limited to senior practitioners while maintaining quality standards.

Credential verification must use the highest-confidence verification method available for the credential type. Self-attestation alone is never sufficient.

All reviewer credentials are re-verified annually. Between annual cycles, the Foundation monitors: state disciplinary board RSS feeds or equivalent for attorney and medical reviewer jurisdictions; FINRA BrokerCheck for financial reviewers; institutional affiliation for academic reviewers. Any adverse finding triggers immediate suspension of review authority pending investigation. Reviewers whose credentials lapse or who face disciplinary action have their Reviewer Integrity Score frozen and their historical reviews flagged for independent audit.

The Foundation maintains the following credential fraud detection measures: (a) Cross-referencing submitted credentials against public disciplinary databases at enrollment; (b) Monitoring for multiple account creation from the same individual using different credential claims; (c) Pattern analysis for review behavior inconsistent with claimed expertise (a claimed general surgeon who reviews oncology questions as if they are a specialist); (d) Periodic random spot-checks of 5% of reviewer credentials against primary sources regardless of scheduled re-verification dates. Confirmed credential fraud results in permanent ban, public disclosure of the ban in the Foundation's transparency log, and referral to the relevant licensing authority.

Reviewers must disclose the following at enrollment and update disclosures whenever circumstances change:

• Financial relationships with AI system developers: Any ownership, employment, advisory, or contractual relationship with an entity that develops or commercializes AI systems. Threshold: relationships with total value exceeding $1,000 per year. Disclosure is permanent — past relationships within 36 months must be disclosed even if currently terminated.
• Financial relationships with entities whose products or services may appear in reviewed answers: A physician employed by or holding stock in a pharmaceutical company may not review answers that mention that company's products. Threshold: any material financial interest.
• Institutional affiliations with positions on the subject matter: A reviewer whose institution has published official guidance on a topic is potentially conflicted in reviewing answers about that topic. Disclosure required; determination of actual conflict made by Foundation COI officer.
• Public advocacy positions: Reviewers who have publicly advocated for a specific position on a contested question (in publications, public statements, regulatory filings, or organized advocacy) must disclose this for reviews on that topic. Advocacy disclosure does not automatically disqualify but may require paired review by a reviewer without the disclosed position.
• Prior involvement with the specific question or AI answer: Any reviewer who recognizes that they have seen the specific question or answer in another context (e.g., as a legal matter they worked on, a patient they treated) must recuse immediately without disclosing identifying information about the original context.

Before a reviewer receives any review assignment, the system performs an automated conflict check against the reviewer's disclosure profile. Reviewers must also complete a self-certification attestation before each review session confirming no undisclosed conflicts exist for the assignments in that session. The self-certification reads: "I confirm that I have no financial, professional, advocacy, or personal relationship with any party, entity, or interest that would affect my ability to assess this answer's information quality objectively. If I recognize any element of this review that creates or suggests a conflict, I will immediately recuse and notify the Foundation."

Reviewers receive exactly the following — no more, no less:

• The question as submitted, fully stripped of all personally identifying information (name, location, institution, relationship, financial details, medical history) by the moderation system before any human sees it. Question stripping is performed by automated PII detection plus moderator confirmation for high-risk domains.
• The AI-generated answer, labeled only as "AI-generated answer" with no system name, model version, or provider identification. Reviewer does not know which AI system generated the answer.
• The domain, sub-domain, jurisdiction, and risk level tags.
• The date the answer was generated (for currency assessment).
• The scoring rubric and completion instructions.

Reviewers do not receive: the name, identity, or any demographic information about the person who asked the question; the purpose for which the question was asked; whether the question has been reviewed by other reviewers; those other reviewers' scores; the AI system's name or provider; any contextual information not present in the question text itself.

Every answer requiring Corroboration must receive a minimum of two independent reviews from reviewers who have not communicated with each other about the assignment. For High or Critical risk answers, three independent reviews are required. Reviewers are randomly assigned from the eligible pool — they cannot select their own assignments. Reviewers cannot communicate with each other during or before review completion. The system prevents reviewers from viewing each other's scores until both have submitted. After submission, aggregated scores are revealed to reviewers to support their professional development, but never in a way that attributes scores to specific reviewers by name.

Reviews submitted in less than a minimum time threshold raise a quality flag and are held for verification. Minimum times by risk level: Standard — 4 minutes; Elevated — 8 minutes; High — 15 minutes; Critical — 25 minutes. Reviews below minimum time do not automatically invalidate but trigger a Reviewer Integrity Score reduction and a request for reviewer explanation. Reviews submitted above a maximum time threshold (48 hours from assignment) expire and are reassigned. Reviewers with a pattern of near-minimum-time submissions across multiple reviews are flagged for behavioral audit.

A Disagreement is automatically triggered when two reviewers' composite scores diverge by 20 or more points, or when their CR1 (Factual Accuracy) scores diverge by 10 or more points. Disagreement status is automatically applied to the answer and remains until resolved. The answer is publicly labeled "Under Specialist Review" during disagreement resolution — the Contested chip is applied.

Stage 1 — Structured Comparison (72 hours): Both original reviewers receive a summary of the scoring criteria on which they diverged (not each other's scores or written notes). Each is asked to re-confirm or revise their score on the divergent criteria only, with written rationale. If the revised scores converge to within 15 points, the average is taken and the status assigned. If scores remain divergent, Stage 2 is triggered.

Stage 2 — Independent Adjudicator Review (7 days): A Senior Reviewer not previously involved (verified against conflict-of-interest registry) conducts an independent review. The adjudicator receives: the question and answer (blind), the two original scores by criterion, and the written rationales from Stage 1. The adjudicator produces an independent score and a written determination indicating which factual claims are in dispute and why. The adjudicator's score is the determining score if within 10 points of either original reviewer; if it falls outside both, Stage 3 is triggered.

Stage 3 — Panel Review (14 days): Three Senior Reviewers with no prior involvement in the review. Panel votes by majority on each disputed criterion. Written determination required explaining the factual basis for each criterion score. Panel determination is final subject only to the Overturn and Correction process in Section IX.

Any registered user may flag a Corroborated answer for dispute by submitting: (a) the specific factual claim they believe is inaccurate, (b) a citation to a published authoritative source supporting their position, and (c) an explanation of the discrepancy. User disputes that meet these requirements enter a triage queue. If triage determines the dispute has prima facie merit, it is escalated to a Stage 2 Adjudicator Review. If triage determines it does not, the answer receives a "User Dispute Filed — Not Escalated" notation with the reason. Triage decisions are published in the Foundation's transparency log.

A Corroborated answer may be overturned — meaning its Corroborated status is rescinded and the answer re-enters review — on any of the following grounds:

• Material factual change: A change in law, regulation, clinical guideline, scientific consensus, or other authoritative source renders the corroborated answer materially inaccurate in a way that could mislead relying users.
• Reviewer disqualification: A reviewer whose corroboration contributed to the answer is subsequently found to have had an undisclosed conflict of interest, a credential deficiency, or a violation of review protocol.
• Process irregularity: A post-hoc audit identifies that the review process did not follow the Corroboration Standard — incorrect reviewer assignment, protocol violations, or evidence of reviewer communication before independent review completion.
• Successful Stage 3 dispute: A Panel Review (CS-DD-002 Stage 3) produces a score materially inconsistent with the original Corroborated status.

Step 1: Overturn trigger is identified by the Foundation monitoring system, a reviewer, or a user dispute that reaches Stage 3. The answer is immediately relabeled "Under Review — Temporarily Uncorroborated." This label is applied within 24 hours of overturn trigger identification.

Step 2: The original AI answer is assessed: (a) Can the answer be corrected by adding a qualification or updating a specific claim? or (b) Is the answer so materially incorrect that it must be removed from Corroborated status entirely?

Step 3: If correctable — the answer is updated with a visible correction notice, the date of the correction, and the specific content changed. The correction history is permanently accessible. The answer re-enters the standard review queue for new corroboration.

Step 4: If not correctable — the answer is permanently labeled "Formerly Corroborated — Rescinded" with the date of rescission and a plain-language reason. The answer is not deleted; the rescission and its history are permanently visible. Users who previously interacted with the answer are notified if the Foundation has a mechanism to identify them.

Step 5: All overturn decisions, the grounds for them, and the correction workflow outcome are published in the Foundation's monthly Transparency Report.

A hostile reviewer who systematically assigns low scores to accurate AI answers in a specific domain or on specific topics to suppress that information is detectable through statistical pattern analysis. The Foundation monitors for: reviewers whose scores on a specific topic cluster significantly below the panel mean across multiple reviews; reviewers who consistently assign CR1 (Factual Accuracy) scores below 20 on answers that other reviewers score above 30; reviewers who show topic-correlated scoring patterns (systematic low scores on topics where they have disclosed advocacy positions). Detection triggers a behavioral audit, freezing of review assignments in the flagged domain, and mandatory review of all recent corroboration decisions contributed to by the flagged reviewer.

A coordinated attack by multiple hostile reviewers — for example, an organized campaign to disqualify accurate medical information by flooding the reviewer pool with activists who systematically mark it inaccurate — is addressed through: reviewer pool diversity requirements (no single organization, institution, or affiliated group may represent more than 20% of active reviewers in any domain); submission timing correlation detection (coordinated low scores submitted within a short time window by reviewers with shared network characteristics triggers investigation); cross-domain consistency checks (reviewers who score accurately in one domain but systematically low in another are flagged for domain-specific audit). The Foundation maintains the right to hold any corroboration decision under review if it suspects coordinated interference, without disclosing the nature of the investigation to the suspected reviewers.

No person with a material financial, employment, or advisory relationship with any AI system developer or deployer may serve as a Corroboration Reviewer for answers generated by that provider's systems. Because reviewers do not see system provider information (blind review), this is enforced through COI disclosure requirements and system-side routing. Additionally: no single AI system provider may represent more than 25% of answers in the Corroboration queue at any time — preventing a single provider from flooding the queue to build corroboration volume; the Foundation monitors for patterns where answers from a specific AI system receive systematically higher scores from certain reviewers, which may indicate undisclosed relationships.

Reviewers who systematically assign scores near the maximum (90+) across diverse answer types and quality levels are engaging in pro-AI inflation regardless of their motives. Reviewers with a mean score above 88 across 20+ reviews are automatically flagged for calibration review. Calibration review involves presenting the reviewer with a set of known-quality test answers (including intentionally flawed answers identified by the Foundation's quality team) and comparing their scores to expected scores. Reviewers who cannot calibrate accurately on known-quality test sets are suspended from the reviewer pool pending retraining.

No single employer, law firm, medical institution, academic institution, or corporate entity may represent more than 15% of active reviewers in any domain. Affiliated entities (common ownership, significant financial relationship) are aggregated for this purpose. Concentration above this threshold triggers a reviewer recruitment campaign targeted at underrepresented institutions. This limit mirrors the Foundation's anti-capture revenue policy — concentration creates undue influence regardless of whether any individual reviewer acts improperly.

For domains where commercial interests are particularly strong (pharmaceutical information, insurance claims, financial products), the Foundation employs additional monitoring: a 10% random audit sample of all Corroboration Reviews in these domains by an independent quality team; quarterly analysis of whether answers mentioning specific commercial entities receive systematically different scores than equivalent answers not mentioning those entities; mandatory disclosure of any financial relationship with any entity mentioned in a reviewed answer, not just the AI system provider. Any pattern suggesting commercial interest-correlated scoring triggers a full domain audit.

Answers in civic, electoral, regulatory, and policy domains are subject to enhanced anti-manipulation measures: a minimum of three independent reviews required (versus two for standard risk level); mandatory political balance in reviewer assignment (reviewers for politically sensitive answers must include individuals whose public advocacy positions, if disclosed, represent at least two distinct political perspectives — assessed through voluntary political perspective disclosure at enrollment, not mandatory); anonymized reviewer political perspective disclosed in the public corroboration record (e.g., "Reviewed by 3 credentialed reviewers including individuals who have disclosed conservative, progressive, and non-disclosed perspectives"); systematic monitoring for party-line scoring patterns on political content.

Current employees of any regulatory agency, legislative body, government ministry, or political party organization may not serve as Corroboration Reviewers in their area of official responsibility. A current FDA employee may not review pharmaceutical answers. A current FEC employee may not review electoral finance answers. Former government officials must observe a 24-month cooling-off period before reviewing in their former area of regulatory responsibility. This exclusion is in addition to, not instead of, the standard conflict-of-interest disclosure requirements.