Same question.
Same burden of evidence.

Apply one standard to everyone.

Republicans, Democrats, independents, presidents, governors, members of Congress, state legislators, local officials, and political organizations receive the same source checks and opportunity for counterevidence. Evidence can support unequal conclusions. Neutrality does not require a manufactured balance.

Decompose before labeling.

Labels such as “racist,” “fascist,” “corrupt,” “extremist,” “socialist,” and “authoritarian” are not self-proving factual conclusions. Define the term, identify the speaker and date, and break it into specific testable propositions about words, actions, decisions, or policies. Separate intent claims from documented conduct.

Historical party affiliation is dated evidence. Direct institutional lineage, documented association, and ideological parallels are different propositions. Record the evidence category explicitly and do not substitute one for another.

Identify the kind of statement.

Source hierarchy

Start with the record closest to the proposition. This order guides retrieval; it does not guarantee truth. A court filing may contain an allegation rather than a finding. An official statement may establish what was said without establishing that it was true.

1. 01Official legislation and statutory textPRIMARY / OFFICIAL RECORD
2. 02Roll-call votesPRIMARY / OFFICIAL RECORD
3. 03Executive ordersPRIMARY / OFFICIAL RECORD
4. 04RegulationsPRIMARY / OFFICIAL RECORD
5. 05Court recordsPRIMARY / OFFICIAL RECORD
6. 06Government datasetsPRIMARY / OFFICIAL RECORD
7. 07Full transcripts and speechesPRIMARY / OFFICIAL RECORD
8. 08Peer-reviewed researchRESEARCH / SECONDARY
9. 09Nonpartisan research organizationsRESEARCH / SECONDARY
10. 10Professional journalismRESEARCH / SECONDARY
11. 11Advocacy / partisan sources as contextual evidenceCONTEXTUAL

Preserve the publisher, title, document identifier, publication date, retrieval date, source version, URL, and exact passage or data location. Clearly distinguish original records from commentary. A repeated report of the same source is not independent corroboration.

Preserve quote context.

Find the full transcript or recording, identify the speaker and date, and include surrounding language, exclusions, and disclaimers that affect meaning. Short media clips are discovery leads, not the complete record. Note uncertainty in transcription or editing.

Trace policy through results.

Separate a statement from a bill; a bill from an amendment; a procedural vote from final passage; authorization from appropriation; and allocated funding from actual spending. Track jurisdiction, effective dates, implementation, affected populations, and changes in the program.

For outcome analysis, record the baseline, measurement period, denominator, comparison group, data revisions, and uncertainty. Evaluate other policies, economic changes, selection effects, and measurement changes. A before-and-after comparison alone cannot establish causation.

Assessments and confidence

Supported

The reviewed evidence supports the scoped factual proposition.

Mostly Supported

The central proposition is supported, with limited inaccuracies or omissions.

Mixed / Context Required

Material parts differ in support, or omitted context changes the meaning.

Insufficient Evidence

The available record cannot adequately substantiate or reject the proposition.

Mostly Unsupported

The central proposition conflicts with the reviewed evidence, with limited supported elements.

Unsupported

The reviewed evidence contradicts the scoped factual proposition.

Unresolved

Research or adjudication is incomplete, or conflicting records remain unresolved.

Confidence concerns the assessment of a particular claim, not the honesty, motives, or worth of a person. It depends on source quality, coverage, agreement, context, and unresolved contradictions. It is not a partisan score.

Prototype limitation: the sample is unaudited demonstration content and is not part of the human-audited American Racial Record. Neither sample nor custom inputs receive a numerical percentage. A production numerical score requires a documented rubric, benchmark validation, and ongoing calibration before use.

Release tiers and audited records

Tier A: every material finding is within the scope of human-audited record entries, with no risk or freshness flags. Label: Automated — from audited record. The report itself has not received human review.

Tier B: eligible unaudited primary government records pass the required checks. Automated assessments cannot be Supported or Unsupported. Label: Automated — includes unaudited sources, not human reviewed.

Tier C: high-risk, conflicting, stale, uncertain, or otherwise ineligible packets remain DRAFT — NOT REVIEWED until an authorized editor approves the exact version. Timeout never releases a draft.

Record promotion is a separate human-only action. Approving a report does not promote its sources. Verified record means human-audited under an identified methodology, not infallible or permanently current. All releases remain private until a separately authorized publication or sharing action.

Human safeguards and correction

AI outputs are research drafts. Deterministic release rules may complete low-risk Tier A and B packets automatically with explicit review-status labels. High-risk Tier C packets require an editor to inspect original sources, contrary evidence, serious allegations, and responses. Publication remains separately authorized. The tool does not automatically publish, issue endorsements, or determine guilt.

Production requirements include review history, correction notices, source versioning, access controls, and protected handling of unpublished reporting. The draft backend contains policy, review, and storage controls; live retrieval, a separate verifier service, and production activation remain unconfigured. Report a concern through DeepQAI’s contact channel.