Skip to contents

Summarize quote provenance for the report's Tier-0 dashboard

Source: R/quote_provenance.R
quote_provenance_summary.Rd

Computes counts + rates from a list of verified quotes. Used by the Tier-0 dashboard (.build_tier0_dashboard) and usable programmatically for cross-run analyses (the methodology paper's KPIs draw from the same shape).

Usage

quote_provenance_summary(quotes)

Arguments

quotes

List of QuoteProvenance objects (after verify_quote).

Value

Named list:

  • total: total quote count

  • by_status: integer vector keyed by verification_status

  • by_method: integer vector keyed by verification_method

  • by_citation_source: integer vector keyed by citation_source ("anthropic_citations_api", "model_freeform", etc.)

  • verification_rate: proportion in either verified state

  • fabrication_rate: proportion fabricated

  • drift_rate: proportion drifted

  • verification_rate_by_source: named numeric vector – per-citation_source verification rate (verified / total quotes with that source). Lets the dashboard expose the differential reliability of the prevention layer (citations API) vs the detection-only layer (model_freeform + ladder).

  • n_citations_api: integer count of quotes with citation_source == "anthropic_citations_api". Convenience accessor for the dashboard's headline KPI.

  • citations_api_rate: proportion of quotes that came through the citations API path (vs. fell back to model_freeform). This is the package's empirical answer to "did the prevention layer actually engage on this run?".

Details

T0.1 part 3b adds the by_citation_source breakdown and verification_rate_by_source so the dashboard can distinguish Anthropic Citations API quotes (server-side-grounded by Anthropic plus client-verified by the ladder) from model-freeform quotes (client- verified only). Both are valid; the citations source is strictly stronger.