Skip to content

Content

The content package defines semantic types for content findings — problems with the accuracy and attribution of published content — and projects them into the domain-neutral Finding. A claim checker produces them: a static scan of claims and their sources, then models and people confirming or dismissing each finding.

Issue

Issue is a problem with one claim, or with one of the sources it cites.

Field Type Notes
ID string Stable identifier
Rule string Rule that raised it, such as content.quote.notFound
Title string Short human-readable summary
Description string Detailed explanation
ClaimID string The claim the issue is about
Statement string The claim's text
SourceRef string The cited source the issue is about, if any
Document string The file that holds the claim
Severity findingspec.Severity Impact; defaults to Medium when unset
Confidence findingspec.Confidence How sure the producer is
Evidence []findingspec.Evidence Supporting artifacts
Remediation *findingspec.Remediation How to fix, incl. an AgentPrompt
Status findingspec.Status Lifecycle state
DetectedAt *time.Time When found

Rules

Common rule IDs; scanners may add their own.

Rule Problem
content.evidence.none The claim cites no source
content.source.unknown The claim cites a source not in the source list
content.source.unreliable The source is a forum, fan site, social media post, AI answer or similar
content.source.unrated The source's reliability could not be decided
content.source.unreachable The source's text could not be retrieved, so its quotations were not checked
content.quote.missing A supporting citation gives no quotation
content.quote.notFound The quotation does not appear in the source's text
content.quote.translationMissing A quotation in another language has no translation
content.facts.missing A date, figure or name in the claim is in no supporting quotation
content.claim.singleSource A claim stating dates, figures or quotations rests on one source
content.claim.notAtomic The claim states several facts
content.claim.verdict A judge found the claim not supported as written
content.claim.lessPrecise The claim is accurate but leaves out a full name, date or figure its sources give
content.claim.subjectMissing The claim, or its supporting quotation, does not name who or what it is about
content.evidence.duplicate A quotation repeats, or is contained in, another the claim cites from the same source
content.evidence.offClaim A supporting quotation bears on no part of the claim
content.evidence.stance A citation's stance, such as disputes, is not what its passage shows
content.review.note A person's note on the claim that no other rule describes

Quotations from Recordings

A quotation from a podcast, interview or talk is checked by cutting a clip of the recording around it and comparing what several speech-to-text engines heard. content.quote.notFound applies when the quotation can't be located in the recording; three more rules cover what the comparison can't settle.

Rule Problem
content.quote.audioDisputed The engines disagree on a substantive word: a number, a negation, a name or a term
content.quote.audioMismatch The engines agree with each other but not with the quotation; the finding gives what they heard
content.quote.audioUnchecked Too few engines produced a transcript, or the clip could not be cut

Put what each engine heard, and the clip's start and end, in the finding's Evidence, so a reviewer can listen to the same seconds.

Finding IDs

content.FindingID(prefix, subject, rule) builds a stable ID: the producer's prefix, the subject (a claim ID, or a segment ID for narration) and the rule's last segment.

content.FindingID("CF", "claim-founding", content.RuleLessPrecise) // "CF-claim-founding-lessPrecise"

The same finding gets the same ID each time it is raised, so a decision about it stays attached from one review round to the next. Give each producer its own prefix, so findings from several tools can be merged without collisions (see Combining findings).

Projection to Finding

ToFinding() projects the issue into a findingspec.Finding in the content domain:

  • sets Domain to findingspec.DomainContent, Type to claim and RuleID to Rule;
  • uses the claim as the location: Location.Component is the claim ID, Location.File the document and Location.Snippet the statement;
  • defaults Severity to findingspec.SeverityMedium;
  • adds claim:<id> and source:<ref> tags.

content.ClaimID(f) returns the claim a content finding is about. It returns "" for findings of another type, such as narration; content findings with no type are read as claim findings.

Narration

NarrationIssue is a recording, or one segment of it, that doesn't say its script, such as a course lesson's voice-over.

Field Type Notes
ID string Stable identifier
Rule string content.narration.mismatch
Title string Short human-readable summary
Description string Detailed explanation
Document string The file that holds the script, such as a slide deck
Segment string The part of the script, such as a slide's ID
Expected string The script's wording
Heard string What the engines heard, when they agree
Recording string The audio file checked
Severity, Confidence, Evidence, Remediation, Status, DetectedAt As for Issue
Rule Problem
content.narration.mismatch Words are dropped, added or misheard by every engine

ToFinding() sets Type to narration and uses the segment as the location: Location.Component is the segment ID, Location.File the document and Location.Snippet the expected wording. Heard, when set, becomes a transcript evidence entry after the issue's own evidence. It adds segment:<id> and recording:<file> tags. content.Segment(f) returns the segment a narration finding is about.

Status Through a Review Funnel

The shared status lifecycle follows a finding through the stages of a review: a static scan opens it (open), a model or person confirms it (confirmed) or dismisses it (false_positive), the claim or its citation is fixed (resolved), or a person accepts it as it is (accepted), for example a single source judged sufficient.