Given researcher-supplied theme name + supporting entries, AI returns
up to n entries that frame the same construct as not-Y, drawn
from the supporting entries plus a bounded, deterministic sample of
non-theme corpus entries injected into the prompt (the model sees only
what the prompt contains – it has no retrieval access to the rest of
the corpus). Per Sarkar 2024 / patterns doc: extractive only – the
model returns entries (not arguments), and a one-sentence reason per
entry.
Usage
provoke_counter_narrative(
theme_name,
theme_entries,
data,
provider,
n = 5L,
audit_log = NULL,
response_cache = NULL,
fabrication_log = NULL
)Arguments
- theme_name
Character: the theme to challenge.
- theme_entries
Tibble: entries the researcher believes support the theme (must have std_id and std_text).
- data
Tibble: the full corpus. Used to resolve and verify cited entry_ids (T0.1) and to draw the bounded candidate sample of non-theme entries for the counter-evidence categories; the model is NOT given the full corpus to search.
- provider
AIProvider object.
- n
Integer: maximum provocations to return (default 5).
- audit_log
Optional AuditLog.
- response_cache
Optional ResponseCache.
- fabrication_log
Optional FabricationLog.